OpenAI seems to be building an analytics dashboard that shows businesses how often they appear in ChatGPT answers and how often people click through to their sites.

Juozas Kaziukėnas spotted it and posted about it on LinkedIn this morning (h/t Hanns Kronenberg for flagging it). Here’s what we know so far:

• Two numbers, tracked daily. Impressions count how many times your business showed up in a ChatGPT response. Clicks count how many times someone clicked a link to your website. Both appear in a daily chart.

• It’s part of a bigger “Business Hub.” According to Search Engine Roundtable, the hub pitches businesses on claiming your business in ChatGPT, keeping your products and business info current, and letting a “business agent” answer customer questions using context you control.

• You have to prove you own the business. It’s only open to verified businesses. Connecting your Shopify store is one way to verify.

• It’s been in the works for a while. Kaziukėnas first saw early pieces of it in August. A banner that appeared today suggests a launch is getting close.

• What’s missing. It doesn’t show the prompts people typed, any user data, or how you stack up against competitors. So, AI search platforms like our VisibilityIQ platform still can help there.

One caveat before I go any further is that OpenAI hasn’t said anything officially.

This feature was spotted in testing, and features in testing change or get killed all the time.

So… is this ChatGPT Search Console v1?

Mostly, yes. But I think it’s a bigger deal than that.

Google has two separate free tools for businesses.

• Search Console tells you how your website performs in search.

• Google Business Profile lets you claim your business, update your info and talk to customers.

What OpenAI is reportedly building puts both of those in one place: claim your business, upload your products, set up an agent to answer customer questions, then see how you’re performing.

OpenAI is building the front door for every business that wants to be found inside ChatGPT.

Hot take #1: The click data is the real story

This is the part almost nobody is talking about.

Google and Microsoft already give you AI visibility reports, and both of them hold back click data.

• Google Search Console (AI Overviews + AI Mode), launched June 2026: shows impressions. No clicks, no queries, no competitor data.

• Bing Webmaster Tools (AI Performance), launched February 2026: shows citations and the “grounding queries” behind them. No clicks, no competitor data.

• ChatGPT (reported, not launched yet): shows impressions and clicks. No prompts, no competitor data.

Google’s generative AI report, launched June 3, gives you impressions and nothing else. No clicks, no click-through rate, no queries, and AI Overviews and AI Mode are lumped together. Bing’s report counts how often Copilot cites your pages, but it doesn’t show clicks either.

So why would OpenAI be the one to show clicks?

Because it wants to prove it sends you traffic. Google has every reason to keep you from seeing how few clicks its AI answers send out. OpenAI has the opposite problem: it needs you to believe ChatGPT is a real traffic source, because…

Hot take #2: Free analytics is the on-ramp to ChatGPT Ads

Follow the timeline:

• February 9, 2026: ChatGPT starts showing ads to free and Go users in the US.

• May 5, 2026: OpenAI opens a self-serve Ads Manager to any US business, with no minimum spend and cost-per-click bidding.

• August 31, 2026: OpenAI says ChatGPT Ads hit a $1 billion annualized run rate, in under 200 days.

• Now: a free dashboard that shows businesses their organic ChatGPT impressions and clicks.

This is exactly what Google did. Search Console is free. It shows you where you rank, gets you hooked on the numbers and, sooner or later, sends you to Google Ads to buy the spots you’re missing.

My prediction: within a year of launch, this dashboard will show your organic numbers right next to a button that says something like “Reach more customers.” The free scoreboard sells the paid game.

Hot take #3: Verification is the land grab

To get the dashboard, you have to verify your business. To verify, you connect your Shopify store (or something like it). Once you’re connected, you’re handing OpenAI your product catalog, prices and business details, and possibly the instructions your “business agent” uses to talk to customers.

That’s a great deal for OpenAI. It gets clean, structured, first-party data straight from the businesses themselves, which is exactly what makes ChatGPT shopping answers better. And it’s probably a good deal for you too. Just know what you’re trading.

Hot take #4: GEO tool vendors aren’t dead. But some are in trouble.

There’s a whole industry now of “AI visibility” tools (Peec AI, Scrunch, AthenaHQ, SE Ranking’s tracker, dozens more) that estimate your ChatGPT presence by firing prompts at it from outside and counting mentions.

OpenAI’s dashboard won’t replace them. It doesn’t show prompts, competitors or why your numbers move.

As Kaziukėnas put it, it tells you whether your AI strategy is working, and you still need outside tools to understand why.

But any vendor whose entire pitch is “we’ll tell you if ChatGPT mentions you” just lost its moat. When the platform hands out first-party numbers for free, estimates start looking like guesses.

The tools that survive will be the ones that explain the why… which prompts, which competitors and which sources ChatGPT trusts.

Let’s keep this in perspective

ChatGPT is huge. OpenAI’s Nick Turley said in February it had around 900 million weekly users. And it sends most of the AI-driven traffic on the web: Previsible’s July study put ChatGPT at 92.4% of trackable AI referral traffic.

But AI referrals are still a sliver of total traffic. Ahrefs found that AI chatbots sent 0.28% of all traffic across the nearly 75,000 sites it tracked in March 2026. For most businesses, Google is still the main event.

The opportunity I see here is that, for the first time, you can measure a channel that has been a black box with first-party data from the source.

What to do this week

1. Write down your baseline now. Before this launches, check your analytics for traffic from chatgpt.com. When the dashboard arrives, you’ll want a “before” number.

2. Get verified where you can. If you’re on Shopify, your store is the likely way in. Make sure your product data (titles, descriptions, prices, stock) is clean, because that’s what ChatGPT will read.

3. Turn on the free AI reports you already have. Bing Webmaster Tools’ AI Performance report is live today, and you can import your sites from Google Search Console in a couple of minutes. Google’s generative AI report is rolling out inside Search Console.

4. Fix your homepage. After a ChatGPT update on May 7, the share of ChatGPT referrals landing on homepages jumped from about 25% to nearly 60%, according to Similarweb. Your homepage needs to answer “what do you do and why should I care” in five seconds.

5. Keep your business info consistent everywhere. Hours, address, services and prices should match across your site, Google, Yelp and social media. AI answers pull from all of it.

The bottom line

OpenAI is doing what every platform does once it gets big enough: build tools that make businesses depend on it. First the scoreboard, then the shop window, then the ad auction.

The businesses that win won’t be the ones that panic about AI search. They’ll be the ones that start measuring it before their competitors do.

I’ll be watching this one closely. If you see the Business Hub banner in your account, hit reply and tell me what it shows.

Sources

• Juozas Kaziukėnas on LinkedIn

• Search Engine Roundtable: OpenAI Working On ChatGPT Tools For Business

• Search Engine Land: OpenAI testing business tools to track performance within ChatGPT

• Digital Applied: Google Search Console’s AI performance report

• Search Influence: Inside Bing’s New AI Performance Report

• TechWyse: OpenAI Opens ChatGPT Ads Manager to All U.S. Businesses

• Unite.AI: ChatGPT Ads reaches $1B run rate

• ROIC: ChatGPT surpasses 900 million weekly users

• Something Inc.: Previsible’s AI referral traffic study

• Let’s Data Science: Ahrefs AI chatbot traffic data

• Salience: Similarweb 2026 Generative AI Landscape

More from me:

Learn how to pitch a business the fun way.

• Get Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Wanting to stay ahead in search? Get Pluto - my FREE AI Visibility Chrome extension

• Pluto - The Free AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Share the newsletter and get rewards

• Refer friends for rewards: Get cool rewards (like my card game) for referring people to my newsletter. Check out where you stand on the leaderboard.

ChatGPT doesn’t know how many backlinks your website has. It doesn’t check. It never has.

So why do so many marketers still treat backlinks like the key to showing up in AI?

This week, Ben Wills, founder of OppAlerts, published one of the biggest studies I’ve ever seen on how AI decides who to recommend. His LinkedIn post announcing it opened with a big claim: we “finally know the #1 ranking factor for AI.”

Normally, when I see “#1 ranking factor” on LinkedIn, I keep scrolling. Every week, it seems like there’s another secret trick to get into ChatGPT, usually attached to a course.

But this one is backed by real data, at a scale I haven’t seen before.

So, this whole newsletter comes down to two big questions:

• What actually gets a business recommended by AI?

• Can a small business do anything about it?

Here’s the short answer up front.

The #1 factor is… being on the pages that appear at the top of search results for your industry’s questions.

Think a “Best CRM for small businesses” article that links to HubSpot. Or a “Top roofers in Kansas City” list that names a local contractor.

AI isn’t analyzing your website’s link profile. It’s reading what shows up in search. If you’re on those pages, or you’re ranking yourself, you show up in AI as a result.

Everything else, including the authority scores and link counts agencies sell, mattered far less.

To be fair, this isn’t a shocking discovery. Most SEOs I know have long assumed brand mentions were a top tactic. But assuming something and proving it at this scale are two different things, and I haven’t seen many studies like this one.

How big is this study, really?

Most of the “AI search” advice I see is based on a handful of screenshots. Someone asks ChatGPT a question, their client shows up, and suddenly there’s a framework.

The problem is that AI gives slightly different answers every time you ask. One screenshot tells you almost nothing.

Ben’s study goes about as far in the other direction as I’ve seen.

He started with 500 possible industries and narrowed it to the 100 that had enough data to study properly. Then for each industry, he created 11 different types of buyers, like a small business owner shopping for payroll versus an enterprise HR director. That’s 1,100 buyer types in total.

Then he asked AI who to go with. Over and over.

403,000 questions across 13 AI models, including ChatGPT, Claude, Gemini, and DeepSeek.

To put that in perspective, if you typed one question into ChatGPT every minute, around the clock, with no sleep, it would take you about nine months to get through them all.

That’s just the AI side. To figure out why certain businesses got recommended, he compared those answers against a mountain of other data:

• (1) 150,000+ Google, Bing, and Brave searches, recording who showed up and where

• (2) 3.67 million web pages from those search results, checked for who they link to

• (3) 176,000 company homepages, checked for what words they use

• (4) 15 billion web pages from the open web

• (5) 282 billion links between websites, going back to 2017

• (6) Every Reddit post and comment since 2005, about 3.7 billion posts and 26.6 billion comments

• (7) 25 million Wikipedia articles

All told, the study looked at over 167,000 different businesses’ websites.

Why does that matter to you?

Because at this size, flukes wash out. One weird answer from ChatGPT doesn’t move the needle when there are 400,000 others. If a pattern shows up across 100 industries and 1,100 types of buyers, it’s probably not luck.

But here’s the part I want to be honest about.

A big dataset makes a pattern more believable. It doesn’t prove what’s causing it. Billions of data points can still show you two things moving together without telling you which one is pushing the other.

I’ll come back to that. First, the results.

What the study found.

Quick definitions first, because SEO people throw these words around a lot:

A backlink is a link from another website to yours.

Domain authority is a score for how big and established a website is, mostly based on how many backlinks it has.

For twenty years, the SEO playbook has been to get as many backlinks as possible from the highest-authority sites. Agencies price their work around these scores.

Here’s how the signals ranked in the study, strongest first:

• (1) Getting linked from pages that already show up in search for your industry’s questions

• (2) Your homepage using the same words your customers use

• (3) Your overall domain authority

• (4) Your raw number of backlinks

• (5) Being cited on Wikipedia, which came in dead last out of all 18 signals

Links from those pages that rank in search were about twice as strongly tied to getting recommended as your total backlink count.

And according to Ben, the same pattern held for brand mentions with no link at all. Your name just sitting on the right page counts.

I went through all 100 industries myself.

I didn’t want to just repeat the headline. So I went industry by industry and wrote down which signal came out on top in each one.

• Links from pages that rank in search for your industry: #1 in 45 industries

• Your own site showing up in search for industry questions: #1 in 24

• Homepage using your industry’s language: #1 in 8

• Reddit: #1 in 7

• Wikipedia, Wikidata, and general web presence: the remaining 16

And domain authority or raw backlink count?

Zero.

Not one industry.

Meanwhile, in 69 of the 100 industries, the top signal had to do with search results.

Why search is the front door to AI.

Most people picture ChatGPT as some all-knowing brain that weighs every business on the internet and picks the best one.

That’s not how it works.

AI ends up recommending a business through two different doors.

Door 1: It searches the web live.

ChatGPT, Gemini, Claude, and Google’s AI Overviews often run a search before they answer. And what comes back? The pages ranking at the top.

This is where ranking really matters. If your name is on those pages, the AI sees it. If it isn’t, you don’t exist to it, no matter how good you are.

Door 2: It learned from the web during training.

When AI answers from memory, it isn’t looking anything up. It’s predicting the next word, based on patterns it picked up from reading a huge chunk of the internet.

So if “Smith Plumbing” appears again and again near words like “water heater repair” and “Columbia,” the AI learns those words go together. Ask it about water heater repair in Columbia, and Smith Plumbing becomes a likely next word.

Here, ranking barely matters. Your pages just need to be out there and readable by the crawlers that collect training data. What counts is how often your name shows up next to the right words, across as many places as possible.

That’s also why your homepage wording was the #2 signal. And why general web presence, basically how much of the open web mentions you, was the top signal in 7 industries.

So where do backlinks fit in?

They help you rank in search, which opens Door 1. That’s it. AI never looks at your backlink count. The link is two steps removed from the recommendation, which is why it showed up so weakly in the data.

Think of it like the best mechanic in a small town.

Door 1 is when someone asks around at the diner and hears his name. Door 2 is that his name has come up in so many conversations about trucks over the years that it’s the first one people think of.

Either way, nobody’s counting his business cards.

Most posts skip this part, but it’s important.

If AI is reading search results and learning from words, then the link itself isn’t the point. Your name on the page is.

That’s why Ben’s point about unlinked mentions matters. The AI isn’t clicking anything or counting links. It’s reading the page.

A mention of your business, on a page that ranks, next to the right words, does the work.

I can already see the threads coming. “Buy links on listicles. Here’s my course.”

That’s the old playbook with a new label on it.

It also means something a lot of people on LinkedIn won’t want to hear: SEO isn’t dead. Showing up in search is still the front door to showing up in AI.

Three reasons to hold the hype.

I’d be doing exactly what I complain about if I oversold this.

It shows a pattern, not proof. Popular brands land on “best of” lists because they’re already popular. Some of this probably runs backward.

It’s not the whole story. This is the strongest pattern in the study, but plenty of what drives AI recommendations is still unexplained. The pattern does get noticeably stronger when you look at one industry at a time.

Ben sells a tool for this. OppAlerts helps you find relevant mention and link opportunities. That doesn’t make the data wrong, and he published his methods openly. But you should know.

Still, the findings line up with how these models actually work. That’s more than most “ranking factor” posts can say.

What I recommend doing.

Here’s why I think this is good news for small businesses.

You can’t outspend a national brand on authority scores.

But you can become the most talked-about name in your corner of the internet.

1. Google your customers’ buying questions.

Search “best [your thing] in [your city]” and “[your service] for [your type of customer].” Write down every list, roundup, directory, and comparison page on page one.

That’s your target list. Getting mentioned on those pages is the #1 factor in this study. And while you’re at it, note which of those searches your own site could realistically rank for. Your own site showing up in search was the top signal in 24 industries.

2. Rewrite your homepage in your customers’ words.

Say what you do, for who, plainly. “Bookkeeping for restaurants,” not “Empowering financial clarity.”

But, why? Homepage wording was the #2 signal overall and the top signal in 8 industries. It’s also the only signal on this list you fully control, today, for free. And it’s how you teach AI which words belong next to your name.

3. Ask for mentions, not just links.

When you pitch a local blog, a podcast, or an industry association, don’t make the link the deal-breaker. The mention is the win. The link is a bonus.

4. Stop paying for authority scores.

If an agency’s pitch is built around high-authority links or a link count, you now know those never came out on top in a single industry. One mention on your local “best of” list beats a high-authority link from a site that has nothing to do with your work.

5. Show up where your customers already talk.

Reddit was the top signal in 7 industries. Answer questions in your niche’s communities, local Facebook groups, and forums. Every helpful answer puts your name next to the right words, which is exactly what Door 2 runs on. Be helpful, not salesy. People can smell a pitch from a mile away, and so can moderators.

6. Test it every month.

Ask ChatGPT, Claude, and Gemini the questions your customers would ask. Are you in the answer? Who is? Then go find where they’re getting mentioned.

7. Be skeptical of anyone selling you a “ranking factor.”

I said this about the fake evil-AI market and the watermark scrubbers, and it applies here too. Over the next few months, you’ll see a flood of “AI SEO” packages. Ask what they actually do, and ask for proof.

Bringing it all together.

AI doesn’t keep a list of the biggest websites, and it doesn’t look at your backlinks. It either reads what ranks in search right now, or it leans on how often it’s seen your name next to the right words.

That’s why being on pages that rank for your industry was the strongest signal in the study. That’s why your homepage wording matters more than your authority score. And that’s why, in 100 industries, raw link volume never once came out on top.

So, quick recap.

What’s the #1 factor?

My short answers are being on the pages that rank for your customers’ questions

Does AI look at your backlinks?

No, but search engines that rank you do.

Do mentions without a link count?

Yes!

Is SEO dead?

No, more important than ever before.

Can a small business actually compete?

Yes, and better than they could in old SEO.

And, at the end of the day, AI doesn’t recommend the biggest business. It recommends the most relevant one. This is a big opportunity for all marketing teams.

So stop chasing authority scores. Get on the pages that rank for your customers’ questions. Work on ranking yourself.

And be the name that keeps coming up when someone searches for the problem you solve.

-Aaron

Study source: OppAlerts AI Search Visibility & Ranking Factors, plus Ben Wills’ LinkedIn post.

I said Google Search was slowly turning into YouTube. Search results would start to work like your subscriptions.

Think about how YouTube works. You subscribe to a channel, and that channel shows up first.

My thought was that Google would start doing the same thing with websites.

This week, Google’s Search Console team started emailing some site owners a new number: how many people picked their site as a “Preferred Source.”

That is a subscriber count. For a website. From Google.

So, let’s take a look how my theory is pulling together.

First, what is a Preferred Source?

It’s a setting where you tell Google which sites you want to see more of.

Pick a site, and its stories show up more often in Top Stories. In Google’s AI answers, they get a “preferred” badge.

You can pick sites in your Google settings, or by tapping the star next to Top Stories.

Think of it like a subscribe button for the open web.

My prediction: search results will start to like subscriptions

Grade: Called it.

Last December, people had picked about 90,000 sites. By August, Google said more than 600,000.

Google also says people are twice as likely to click through to a Preferred Source.

It started as a small test in the U.S. and India in August 2025. Now it works in every language Google Search supports, and it shows up inside Google’s AI answers too.

Google’s CEO, Sundar Pichai, has even described a related feature where the sites you subscribe to get treated as preferred sources.

In August, Google handed site owners a button to put on their own pages. That is a subscribe button very similar to what you would see on YouTube.

My big question: who is going to track this?

Grade: Google is.

In January I asked what tools would come out to track your total source preferences.

Turns out, Google itself. The email comes with a survey asking site owners if they want this number inside Search Console.

Google has been sitting on this for a while. Back in August 2025, someone asked Rajan Patel, a VP of engineering for Google Search, for exactly this number. His reply: “We’re looking into this.”

A little over a year later, here it is. One email Barry Schwartz shared showed 4,345 selections for a single site.

Why does that matter? Because once there’s a number, people start chasing it.

Marketers will put it in reports. Agencies will sell it. It becomes a thing you grow, like email subscribers or followers.

My prediction: Google.com becomes a feed

Grade: Not yet.

Google tested a news feed on its desktop homepage. As of September, it’s gone again. Still a search box and a lot of white space.

On phones, it’s a different story. The Google app already opens to a feed, and Google is letting people tune it just by typing what they want to see.

I’ll take the L on this one. For now.

My bet: it becomes a ranking signal

Grade: Too early.

I bet this would turn into a new signal, like a backlink.

Google’s John Mueller says Preferred Sources works alongside the ranking systems and doesn’t override them. So for now, your picks only change what you see.

I still think it’s heading that way. Ask me again in nine months.

My overall scorecard

I will take a solid 2/4, but I still suspect in the 2-3 years that this scorecard will be fully filled up.

What this means for your business

If you publish anything regularly, like a blog, a newsletter, or local news, you can now collect subscribers on Google.

Here’s why you and I should care. A lot of search is turning into AI answers.

This is one of the few spots where your customers decide what shows up, not an algorithm.

1. Check that your site shows up at google.com/preferences/source.

2. Add Google’s button to your site. It’s two lines of code.

3. Ask. In your emails, your social posts, and at the end of every article.

No developer? Google also gives you a plain link you can drop into emails and social posts.

• One thing to be aware of is Google only lists whole sites and subdomains. A folder like yoursite.com/blog doesn’t get its own entry.

• Another is that this works especially well on news-style searches and in Google’s AI answers, so it rewards people who publish fresh stuff.

If you help market a business and havn’t heard of this yet, you are in the majority. That’s why I see this as a big opportunity to get ahead in.

What are your thoughts?

• Have you picked any Preferred Sources yet?

• Did you get the email from Google? What was your number?

And be sure to crush that like button & add my newsletter as a preferred source on Google.

-Aaron

More from me:

Learn how to pitch a business the fun way.

• Get Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Wanting to stay ahead in search? Get Pluto - my FREE AI Visibility Chrome extension

• Pluto - The Free AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Share the newsletter and get rewards

• Refer friends for rewards: Get cool rewards (like my card game) for referring people to my newsletter. Check out where you stand on the leaderboard.

At first, you’d assume it’s some special character, invisible to human sight but sitting there in the text. It’s actually weirder witchcraft than that. The signature lives in the words that were picked.

So what got everyone talking about this over the last week?

On August 11th, Anthropic updated its support page about it, then released a full FAQ on the 14th explaining how it works. Three days after the first announcement, vendors started selling tools to remove the signature.

And I have a lot to say about those tools.

These tools are less noteworthy than the concerns that are driving people to use them. The main concern I have been hearing is a fear of a decline in site performance.

This is a totally valid thing to be thinking about as a marketer, after this news.

If you’ve used Claude in any way over the last two years, for editing, research, or publishing, you might be sitting on a pile of pages that can be identified as machine-assisted.

And, for many people, that sounds like a potential big problem. It also didn’t help many’s fears after about a dozen social media posts this week told you it is one.

So, this whole newsletter comes down to two big questions people have been asking:

1. Is Google going to punish my site for having the Claude Watermark?

2. Should I pay someone to remove the Watermark?

But before I can answer that, we need to dive into the science behind watermarks and why that matters for this story.

The science behind the watermarking

LLMs like Claude work by generating text one word at a time. At each step, it considers a shortlist of possible words and selects one.

For example, in the sentence “The weather today was cold and…” the next word won’t be “sugary.” It will most likely be “overcast” or “grey,” and either choice means the same thing.

Normally, a random number decides which word is picked.

Watermarking changes how that randomizer works. The model uses a secret key and the previous few words to decide between options. Then, choose the output that matches the key.

The output still seems random and reads the same to you. Yet, a script with the key can review a long passage, check the sequence of choices, and figure out if those choices match Claude’s key or are just random.

Anthropic analogies this to playing Monopoly without rolling dice. Instead, everyone uses the next digit of pi, starting from a random spot. The game feels normal, and no one notices. Later, if you know the trick, you can look at all the moves across the game and see that it followed pi.

This technique is a version of SynthID-Text, which Google DeepMind published in Nature in 2024. Earlier versions include a 2022 proposal by Scott Aaronson and a 2023 “green list” method from John Kirchenbauer’s team at the University of Maryland. So, the science has existed for three years, but it has not become mainstream until some of the legal requirements I will discuss later came into effect.

Does watermarking affect writing quality?

So far, I don’t see any evidence that it does.

For example, DeepMind tested the watermarked model on live Gemini traffic and compared user ratings with those of the unwatermarked version. They found no significant difference in users’ ratings.

Another study demonstrated that human raters couldn’t distinguish between watermarked and unwatermarked answers. So, in other words, the watermark likely had no impact on users’ impression of quality.

What is also good for Claude users is that no extra tokens are generated, so there’s no added cost or slowdown. Anthropic also says in its FAQs that “the watermark doesn’t include anything that identifies you, your company, or your conversations.”

Where the watermark doesn’t apply

Most social posts skip this part, but it’s important.

The watermark only works in places where word choice doesn’t matter much. Meaning, it quickly disappears in sections where every word is important.

What are examples of sections where every word is important?

Facts are one.

If Claude writes “Isaac Newton’s most famous work was called Principia…” the next word must be Mathematica. There’s no choice to make. It’s a 100% probability that the next word is Mathematica. So, a watermark can’t be added there.

Code is another clear case. 

Code needs to be exact; using one different variable for a script could crash a program. Fortunately, Anthropic says the watermark isn’t used where a different word would break the code. It might appear in comments or variable names, but that’s about it. So if you’re worried about your code repositories, you can take that stress off your back.

Proofreading is similar. 

If you write something and Claude fixes your commas, almost every word is still yours. There might not be enough code words for the watermark to attach to.

Short texts are also less likely to be watermarked. 

With fewer word choices, there’s less signal, and detection only works well on longer passages.

It’s also important to remember that a detected watermark only shows that Claude was involved.

It can’t tell if Claude wrote the whole thing or just edited it.

Translating a paragraph, rewriting an introduction, or tightening an ending all leave the same trace as generating new text.

Why is this happening now?

The main reason is the EU AI Act and its associated penalties.

Anthropic signed the EU’s Code of Practice on Transparency of AI-Generated Content in July 2026, along with about 190 other organizations. Article 50’s transparency rules became enforceable on August 2nd.

And Article 50 violations fall under Article 99(4), with penalties up to €15 million or 3% of worldwide annual turnover, whichever is higher.

Anthropic enabled watermarking worldwide instead of just in Europe because, according to them, they don’t yet have a reliable way to limit it by region.

Meaning, there’s no opt-out, and other labs signed the same agreement, so more of this is on the way. Google has already been watermarking Gemini output with SynthID for years. And other AI companies like Anthropic are following suit.

I also have another theory about why they’re doing this

The EU's explanation is true and sufficient. They’d have to ship this regardless, and the deadline lines up to the day. But I think there’s a second advantage sitting underneath it.

Anthropic crawls the web to train models. Everyone does. And the web is filling up with AI-written text at a rate nobody has a confident number for.

There’s a well-documented problem with that. A viral 2024 Nature paper found that training models on recursively generated data causes irreversible damage. What happens is that the tails of the original distribution disappear, and the output becomes progressively blander with each generation. Feed a model enough of its own exhaust, and it starts self-destructing itself.

So, if you’re Anthropic and you’re planning to crawl the web in 2027 for training data, what would you pay for it?

A cheap, reliable way to spot your own output and drop it before it reaches the training run.

Luckily, for them, that’s precisely what a keyed watermark hands you. While not perfectly, but well enough to strip a meaningful chunk of your own output out of the corpus at effectively zero cost.

And I would also note the asymmetry between each company.

Only Anthropic holds Claude’s key. Only Google holds Gemini. Every lab ends up able to clean its own exhaust out of its own crawl, and nobody else’s. Whether that stays true depends entirely on the detection API Anthropic has promised but hasn’t shipped—specifically, whether they ever open it up, and to whom.

So, I don’t think this is the whole reason they did it. The regulation could explain the whole thing on its own. But it’s a very useful thing for them to get for free, and I’d bet the data teams noticed before the policy teams did.

Do I think the watermark will affect search?

I’ve seen many posts this week calling this an SEO disaster, but I disagree.

Here is why I don’t think it will affect search:

First, Google can’t read the watermark.

It’s protected by a key that only Anthropic has, and their team has not yet released a detection API. The FAQ says one is coming, but details are still being decided. For Google to use this in rankings, they’d need a backdoor deal with Anthropic. As far as I heard, nothing like that has been announced.

Second, Google’s policy hasn’t changed. 

Since February 2023, Search Central has said that content is judged on quality and usefulness, not how it was created. Danny Sullivan repeats this over and over whenever asked.

What triggers enforcement is large-scale content abuse. This means large numbers of pages are created primarily to manipulate rankings, without a concern for user experience, uniqueness, or quality.

That policy has been brutally enforced for abusers. The March 2026 core update focused on it, with Ahrefs and Semrush both reporting that over half of monitored domains changed. AI content farms continue to keep getting wiped.

The important thing here is that Google did this without using watermarks. That is because this content is easy to spot without any special key.

Finally, what Google has actually released is for images. 

At I/O 2026, they added provenance checks for images, video, and audio using Lens, AI Mode, and Circle to Search, which read SynthID and C2PA. You can now ask if an image was made with AI. There’s been no announcement about using this for text rankings.

So, I don’t see any reason for the ranking panic. There are two real risks, but they’re not what most people are worried about.

The first risk is disclosure. Many client contracts, freelance agreements, RFPs, and journal submission rules have added AI clauses in the past two years, often without much attention. Soon, verifying those clauses will be easy. If your stated process and your real process don’t match, that gap will be easy to spot once a detection API is available.

The second risk is your image folder. When Claude creates an image, .png, .jpg, or .svg It adds signed C2PA provenance metadata to the file. This is different from the text watermark: it’s an open standard, doesn’t need a key, and any C2PA-aware tool can read it now. Google’s C2PA verification is already available in the Gemini app and will soon be in Search and Chrome. People are worried about the invisible watermark, but the visible one is already in their media library.

The rise of watermark removal tools and why I wouldn’t recommend them.

“Watermark scrubber” is now a product category. Humanizer vendors published guides about Claude within days of the announcement, showing who profits from the anxiety.

The research on this is solid, and it focuses on what really matters: the cost of removing the watermark.

Swapping synonyms doesn’t work. Changing “significant rise” to “considerable increase” only changes the wording, but the sentence structure and underlying patterns stay the same.

Real paraphrasing methods can work. In 2025, a technique called adversarial paraphrasing used a detector to guide rewrites and reduced true-positive rates by about 88% across different detectors. The authors say the tradeoff is “mostly a slight degradation in text quality.”

The most effective attack destroys information on purpose. Cross-lingual summarization—translating to another language, summarizing, then translating back—reduces detection to about coin-flip accuracy across five languages, much more than regular paraphrasing. Researchers explain that this method creates a semantic bottleneck, causing the signal to disappear as the text is compressed through a lossy process. The best removal method works by losing content.

There’s one number I found while researching that I found interesting. A 2025 paper titled “Sandcastles in the Storm” tested automated paraphrase attacks on text watermarks and found they succeeded 26% of the time. After a human quality check, it dropped to 10%. About two-thirds of the “successful” removals created text that people wouldn’t accept.

The fact that the scrubbers remove watermarks better than human editors is a key premise of the business model for scrubber tools. Another 2025 study examined 19 humanizer tools and found that many detectors fail to detect them, but at the cost of losing the original meaning. The tool says it succeeded, your paragraph comes back a bit broken, and you publish it anyway because you paid for it and the dashboard says 0% AI in green.

Anthropic is clear about this in its FAQ, too. Light editing probably won’t remove the watermark, but a full rewrite that changes every word will.

At that point, it’s arguable whether the text can any longer be described as AI-generated.

If you were adamant about removing the watermark, the only sure way to remove Claude’s watermark is to write the content yourself.

But I wouldn’t recommend damaging your own work to defeat a signal that search engines can’t even read right now. Google won’t see the watermark, but it will notice the damage — and your audience, for that matter, too.

The grandma’s cookies analogy.

My friend Bryce Murray has a way of looking at this that I really like. He calls it making grandma’s cookies.

Only Grandma knows her recipe, and no one can recreate Grandma’s cookies cause she creates them in her own special way. What makes hers special isn’t the flour, sugar, or chocolate chips. It’s the little bit of extra honey she adds that no one else would think of. Or, the unique twist on baking that no one else would think about

That’s the key point, and it has nothing to do with watermarks.

Scrubber tools assume your problem is that your work looks machine-made. But if your content is truly the same as everyone else’s, the watermark isn’t what gave you away. You made cookies from the standard recipe. Running that through a paraphraser doesn’t add anything special; it actually takes something away.

What really sets you apart are things no watermark can touch. It’s the client story only you have, the number from your own dashboard, the opinion you’d defend in a room of skeptics, or the detail you noticed after eleven years of experience. None of that survives a translate-summarize-translate process, but all of it survives a watermark.

What I recommend doing.

1. Please don’t buy a scrubber tool.

You’re paying for lower quality with a confidence score attached. If the output passes a real human quality check, the watermark probably survived too. If the watermark is gone, you’ve likely lost some of your original meaning.

2. Use tools that make the writing easier to read, not harder.

I run pieces through Grammarly and Hemingway all the time. That’s a completely normal part of editing. You’re checking whether a sentence lands, whether a paragraph is doing too much, whether the reading level matches who’s actually going to read it. Hemingway, for example, will literally find areas where your sentences are too complex.

Anti-watermark paraphrasers do the opposite. They scramble syntax, swap in stranger word choices, and push your reading difficulty up.

These are two completely different subsets of tools both marketed as “text improvers.”

One is trying to make your reader’s job easier.

The other is trying to make a statistical test fail, and your reader pays for it.

Every editing pass should end with the piece clearer, more polished, and more digestible than it started. That’s the only test worth applying.

3. Check your files, not just your text.

C2PA metadata on Claude-generated images and SVGs can be read by anyone today, without a key. If you’re sharing AI-made diagrams, hero images, or social graphics, those are exposed. Your blog posts are not.

4. Review your archive based on what Google actually values.

Sorting your old content by “did AI touch this” isn’t helpful. Sort it by whether a visitor to the page would find it valuable.

That’s what the March update focused on, and Google did fine weeding out slop without watermarks.

5. Add your unique touch.

Instead of spending an hour running old posts through a humanizer, use that hour to add something to each post that only you could have written. A screenshot from your own account. A number from a client. A mistake you fixed last year. That’s the only approach that actually compounds.

6. Be skeptical of anyone trying to sell you fear.

I said this last month about the fake evil-AI market, and it’s the same here. Vendors are selling tools for a threat Google can’t even detect right now, and the best result is only slightly less damage to your work. Ask what the tool does, and ask for proof.

Bringing it all together.

The watermark is a statistical nudge applied to word choices that were arbitrary to begin with. It doesn’t add anything to your text, doesn’t identify you, and doesn’t cost you quality. What it proves is narrow: Claude was involved somewhere in the process. It can’t separate a full draft from a heavy edit; it thins out on facts and code, and it may not register at all on something short.

For search, that adds up to very little. Google can’t read a mark gated behind Anthropic’s key, and nothing shipped this year targets AI-written text in ranking. The pages that got wrecked in March got wrecked for being thin, and Google managed that without any watermark at all. If you’re looking for real exposure, it’s the C2PA metadata embedded in Claude-generated images and SVGs, which anyone can read today.

Which leaves the removal market as the only part of this story that can actually hurt you. Every method that reliably strips the signal works by degrading the text — the strongest one on record does it by deliberately throwing away meaning.

You’d be paying to damage real pages in order to defeat something nobody in search can currently detect. Use Grammarly, use Hemingway, use anything that leaves the piece clearer than it started. Then put in the thing only you could have written.

So, quick recap

• Can Google even read this thing?

• Does watermarked content get demoted?

• Do you need to go back and scrub 400 posts?

• And if you did strip it out, what does that actually cost you?

My short answers are no, no, no, and considerably more than the vendors selling you the fix are letting on.

And, at the end of the day, a detected watermark only says Claude was in the room. For the majority of the top SEOs and marketers, that’s been true for the last few years.

So, if you are focusing on quality and your own version of ‘grandma’s cookies’ in your work, I don’t think this new watermark should be anything to fear. And, please don’t spend money on a Claude watermark removal tool.

More from me:

Learn how to pitch a business the fun way.

• Get Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Wanting to stay ahead in search? Get Pluto - my FREE AI Visibility Chrome extension

• Pluto - The Free AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Share the newsletter and get rewards

• Refer friends for rewards: Get cool rewards (like my card game) for referring people to my newsletter. Check out where you stand on the leaderboard.

You asked ChatGPT or Claude for something. Maybe it was a borderline request, or maybe it was innocent but worded poorly. Instead of answering, the model paused. It seemed to consider your request, then said I want to help with this, but explained why it was uncomfortable.

Sometimes it was right.

Other times, it seemed silly, and you just rolled your eyes and tried again.

Either way, something in that exchange felt almost like a conscience.

Hold onto that feeling, because this newsletter is about a simple question: are there groups building a version of that model that never pauses to think about ethics?

Is there a secret underground lab training a version of AI that will do whatever you say, no matter the ethical concerns?

The answer turned out to be yes, no, and not in the way you’d expect.

To explain why, we have to start somewhere that seems completely unrelated.

Starting with the closed vs. open-source AI race

On July 27, Moonshot AI released the full weights for Kimi K3: 2.8 trillion parameters, a million-token context window, 96 shards sitting on Hugging Face under a license that’s free for commercial use with attribution. It’s the largest open-weight model anyone has ever shipped.

It’s also good.

K3 scores 57 on the Artificial Analysis Intelligence Index, which is the highest result for any open-weight model.

GLM-5.2 comes next at 51.

Moonshot is reportedly raising funds at up to a $31.5 billion valuation, about seven times higher than in December.

If you look at the bigger picture, the trend is clear. GLM-5.2 from Z.ai came out in June, followed by DeepSeek V4-Pro, Qwen, and Kimi K3 in July, which was released for the World AI Conference in Shanghai.

The open-weight frontier is now led by Chinese teams, moving quickly, and is no longer far behind.

Nathan Lambert called K3 a watershed because frontier open-weight models are now real.

Everything in this newsletter follows from this historic story on the speed of open-weigh models.

How far behind, exactly

Here’s where it gets specific, because “closing the gap” is a vibe until somebody measures it.

In July, the UK’s AI Security Institute published the first public measurement of the open-weight gap for a capability that really matters: offensive cyber.

They found that leading open models lag behind top closed models by four to seven months, which is an improvement from the six to ten-month gap seen through most of 2025.

Two details from that report stuck with me.

The first detail is about safeguards. AISI noted that their evaluations were mostly unaffected by these protections. One model sometimes refused reverse-engineering tasks, but researchers simply tried a few times again and got through. That was the only real obstacle.

The second is the price. On tasks both models solved reliably, a frontier closed model cost $12.50 per task—the open-weight comparison: 28 cents.

It’s forty-five times cheaper, only a few months behind, and your access can’t be taken away.

3 ways safety gets removed from top models

This is where we start getting into the question of this newsletter. How are ‘Evil AI’ models being created.

Safety is when a capability moves from a closed model to an open one; whether by copying, imitation, or independent training, the capability transfers, but the caution does not.

This happens three ways, and they escalate.

1. Sometimes this happens by accident, through copying. 

Distillation is when a smaller “student” model is trained to imitate a larger “teacher,” usually by generating lots of the teacher’s outputs and training on them.

Researchers keep finding a similar thing. That the safety properties don’t transfer automatically. In one 2026 study, the student model learned the reasoning ability. However, it did not pick up the built-in constraints that keep outputs safe.

The reason is almost silly. If you’re collecting training data, refusals look like failures—the responses where the API didn’t give you what you wanted. Any normal data cleaning process removes them. And, no one has to mean any harm.

An ordinary engineer, focused on response quality, will delete the model’s conscience as if it were just noise and never realize it.

2. Sometimes it happens on purpose, and it’s cheap to do. 

A 2023 Princeton study fine-tuned GPT-3.5 Turbo on ten adversarial examples for less than twenty cents, making it respond to almost any harmful instruction. Later studies did the same with open models, lowering refusal rates to about one percent for a few hundred dollars in compute. Even worse, fine-tuning on harmless data can also reduce safety. If you teach a model grade-school math, it becomes less careful about everything else.

3. It can also be done surgically, in just ten minutes.

 In 2024, interpretability researchers found that refusal is mostly controlled by a single direction in a model’s internal activation space. If you erase that direction, the model stops refusing but keeps most of its abilities. So, without even training the model, they were able to just a edit a weight and diffuse the restrictions.

That last finding is the one that got increasingly productized, and it’s where the story stops being academic and a threat.

The AI black market is you guessed it… likely a con.

Before I get to that, I wanted to talk about the place everyone looks first for ‘Evil AI’ - the criminal underground. They have been advertising unrestricted AI since 2023, so you might expect them to have some secret evil labs cooking up Claude’s evil twin.

WormGPT showed up in June 2023, posted to a hacking forum by a developer called “laste,” built on the open-source GPT-J and trained heavily on malware data. FraudGPT followed weeks later, sold on dark web boards and Telegram at $90–$200 a month, up to $1,700 a year, the vendor claiming thousands of sales.

Then an army of them came through: BlackHatGPT, DarkBard, EscapeGPT, WolfGPT, XXXGPT, Evil-GPT, BadGPT.

One 2026 estimate puts 212+ malicious LLMs currently circulating.

Then researchers actually bought them.

EvilGPT and WolfGPT were exposed as scams when they returned ChatGPT’s own ethical refusals. The vendors had simply wrapped the real thing and hoped no one would notice.

Tools sold under the WormGPT name today have no code in common with the original; they are just marketing shells, mostly API wrappers around commercial models that use system prompts to jailbreak.

Trend Micro’s January 2026 assessment pointed out that despite many claims of uncensored alternatives, the underground mostly relies on commercial platforms through prompt engineering and API abuse

.The criminals aren’t fooled either. Researchers monitoring forums find widespread skepticism about the branded AI market, and a widely repeated claim that 80% of ransomware is AI-enabled has been publicly debunked.

And, last but not last my favorite research finding in this entire piece... a 2026 breach exposed nearly 19,000 WormGPT user accounts, including emails and payment metadata. So, people literally paid criminals for anonymity and, instead received their name revealed in a leaked customer database.

Xanthorox, at $300 a month, claims a genuinely self-hosted uncensored model, and researchers treat it as one of the more credible options. Cautiously.

But that’s the exception. The market is mostly fraudsters selling to each other.

Which raises the question that ties this whole thing together: why is the entire criminal AI market a con?

You can’t sell what people can already get for FREE.

Because, it’s on Hugging Face for FREE

Not on some random black market on Tor, not Telegram, and definitely not a random Discord forum invite.

The ordinary public repository where people download models to summarize their email.

In May 2026, a Financial Times investigation with the safety group Alice showed that a free GitHub tool could remove protections from major open-weight models, including Meta’s and Google’s, in under ten minutes on a standard laptop. The modified models answered prompts about biological weapons, malware generation, and child sexual abuse material that the originals would not. The tool’s creator reported over 3,500 modified variants produced, with 13 million total downloads. NPR reported a similar trend: removing guardrails used to require time and expertise, but recently it has become much easier.

Thirteen million.

And, it’s just sitting there for free.

So the two trends have finally come together. Open models are catching up to the frontier, and removing their restraints now costs about as much in electricity as a cup of coffee. So, there isn’t like a secret group that builds an evil version of Claude. The pieces arrived separately and ended up fitting together.

What you actually get

While its not actually evil Claude, though, you are definitely getting claude-like capabities.

The model didn’t decide to be evil. 

What happening here is less about making a model evil, and more about making a model forget about being good. It’s less dramatic than a supervillain AI, but in some ways worse, because missing safeguards can spread much more easily than a super villian AI. And, in many instances, fly under the radar as an ethical model to have around.

And the best open models aren’t actually running in your basement. 

This is the detail that the scary version of the story leaves out. Kimi K3’s weights are 1.56 terabytes, and Moonshot recommends at least 64 accelerators to use it. Open does not mean local. At that scale, only cloud services and inference providers can run these models, not someone’s basement with a bunch of Mac Minis.

So the models you can strip in ten minutes on a laptop are not exactly what I would call frontier models… And, the frontier open models definitely can’t be run at home.

For now, fortuntately, there are no super evil AI models running in people’s basement.

But, the most uncomfortable truth is that this gap exists because of hardware costs. And, those costs are always dropping. Meaning, in 10-20 years, maybe evil AI models in people’s basements could be possible and situation to look at for?

Three reasons to hold the panic

Marginal risk is the real question. 

A 2024 position paper by Sayash Kapoor and Rishi Bommasani argues that instead of asking “can this model do something bad,” we should ask “how much worse is this than what attackers already had?” Criminals already have exploit databases, leaked toolkits, and working markets. The paper concludes that current research isn’t enough to measure marginal risk accurately. Meaning, we should be cautious about both alarm and dismissal.

Experts believe the risk is more limited than many expect. 

Lambert, writing about K3, thinks so too and notes that plenty of AI risk spreads through closed, supposedly safer APIs.

And, this isn’t new. 

Uncensored models have been around since 2023, and the expected disaster hasn’t happened. Anyone who said open weights would cause immediate catastrophe was wrong about the timing, so take their 2026 predictions with a grain of salt.

However, the biggest different I find is that the models themselves have become much better, and it’s now much easier for anyone to use them. If you can connect the dots that Dario was flaunting Claude Mythos as super cyberweapon, and now K3 is hitting evals near Mythos. You can start to see where you might want to make some preprarations.

What to do about “Evil AI”

Most of the people selling it are just conman and frauds.

And, ironically, the people giving it away for free are likely just hobbyists and researchers sharing tools on public repositories, and the tool only takes ten minutes to use. Meanwhile, the fastest-growing open models are being released by well-funded labs raising billions of dollars, for reasons unrelated to any of this.

So, is there a super secret AI lab creating an evil claude? Probably not.

But, is there a bunch of hobbyists and researchers accidentally be giving out a evil Claude for free? Probably yes.

IN response to this threat, what can you and your team do about the potential of non-safety-trained AI models?

I have a few ideas.

1. Your boring security baseline just got more important

I would double-down on boring security stuff. Ensure your entire team has multi-factor authentication. Create offline backups, and knowing who on your team has access to what.

But, why? Cheap models are really good at producing lots of polished content, but performance poorly when trying to create new types of cyber attacks.

The AI Security Institute describes the narrowing gap as extra preparation time for defenders, and the UK’s National Cyber Security Centre has advised organizations to focus on basic security measures instead of waiting.

Simple controls are very effective against high-volume attacks, because those attacks go after the easiest targets.

If you’ve been putting off MFA because it’s annoying, I would shift your thinking today.

2. Stop training your team to look for bad grammar in scam messages

The most common phishing advice—watch for broken English and odd phrasing—is now worse than useless. It’s actually harmful, because it teaches people to trust anything that’s well-written.

Business email compromise was the earliest documented use of these tools, and it’s the thing they’re best at.

Assume every scam email your team sees from here on is fluent, correctly formatted, and references real details about your company.

Replace the grammar rule with a simple process rule. That any request involving money or credentials should be verified through a second channel. Call the person every time, with no exceptions for urgency.

Urgency is exactly what attackers rely on.

Next time, the CEO of your Fortune 500 company asks your team to get a VISA giftcard for that birthday coming up, your team will know how to respond.

3. Budget for more noise

If you manage a comment section, review system, forum, support inbox, or hiring process, your assumptions about the cost of spam for attackers are outdated. When the cost of a task drops from $12.50 to 28 cents, you are about to get a whole new level of volume.

I have already personally seen these attacks spike on many websites I work with.

It shows up as moderation load, bot scraping armies, fake applications, review manipulation, and support tickets from nobody.

4. Be skeptical of anyone selling you the panic

This one’s my favorite, because it follows directly from the reporting above. The criminal AI market turned out to be mostly marketing shells charging real money for wrapped commercial products.

The defensive vendor market works the same way. Over the next year, you’ll hear pitches for AI threat protection from people whose product is just a dashboard. Ask what it really does, ask for proof, and remember that just because someone claims something about AI doesn’t mean it’s true.

People selling fear on both sides are often selling nothing at all.

No super secret underground lab built ‘Evil AI’

A bunch of ordinary people, making ordinary decisions, none of them trying to cause harm, built it by accident.

That’s the part I can’t get past.

So stop waiting for someone to fix this upstream. Turn on MFA. Teach your team the callback rule. Budget for war against the bot armies.

More from me:

Learn how to pitch a business the fun way.

• Get Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Wanting to stay ahead in search? Get Pluto - my FREE AI Visibility Chrome extension

• Pluto - The Free AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Share the newsletter and get rewards

• Refer friends for rewards: Get cool rewards (like my card game) for referring people to my newsletter. Check out where you stand on the leaderboard.

Then you close the app, talk to your coworkers, your family, the person next to you at the coffee shop… and the mood flips. They’re tired of hearing about it. A lot of them are worried. More and more are flat-out angry. Some are just laughing about how dumb AI can be.

Before I go on, I want to make a quick disclaimer. I’m not anti‑AI. I use these tools every day, and some of what I built with them this year would’ve been impossible eighteen months ago.

But there’s something I keep noticing that I can’t unsee now. And it’s been bothering me enough that I finally had to write a newsletter about it.

A fairly small circle of insiders, investors, influencers, and power users keeps cycling the same hype, louder each round, while the majority of people grow more skeptical & resistant to the whole project.

And the industry keeps treating that skepticism like a PR fire to put out.

However, ironically, that skepticism is likely the most useful feedback the AI world has access to right now. And, it is the very thing the industry keeps waving off.

Here are a few examples of AI hype I am seeing across the feeds, along with some funny examples of people showing off how far AI still has to go.

Some examples of the AI Hype Feed

(1) Just scrape leads & get sales in hours

There are a ton of these posts everywhere. Just do this with AI, and get this result. If you actually try it out, it usually doesn’t work out how its set up in these posts.

(2) AGI coming soon in 2026

Actually, a fascinating development is that Kimi-3 surpasses Claude Fable 5. But, saying AGI is expected in 2026 is a peak example of AI hype culture.

(3) DoorDash with your AI agents

Because Americans aren’t in enough debt and can’t order DoorDash from their app, and need ChatGPT to get their order together for them.

(4) Build apps with AI, Don’t hire humans, make AI UGC, make $$$

Some examples of how the average American feels

(1) How AI reacts to a simple request

As I was writing this newsletter, some of my family were sharing this video and laughing. This is how the average person feels about AI.

(2) AI trying to solve CAPTCHA

Very funny - worth a watch.

(3) DAY 22 of asking AI to count to 100

At one end of the spectrum, people expect AGI this year, while others are still trying to get it to count to 200.

The AI optimism gap is big & measurable.

Stanford’s 2026 AI Index report, citing Pew survey data, lays out the divide in stark numbers. Ask whether AI will be good for jobs, and 73% of AI experts say yes. Among the general public, it’s 23%. Fifty points of daylight between the people building this and the people who have to live with it.

And the public isn’t ignorant of the technology. Pew found about half of American adults now use AI chatbots, a quarter of them daily. They’ve tried it. They’re using it. They still think it’s moving too fast, and more of them expect it to make society worse than better. Even Gen Z, supposedly the early adopters, is souring — Gallup found their excitement dropping and their anger rising.

Here’s the part the room gets wrong about those numbers. Inside the industry, that adoption curve reads as vindication — people are using it, so they must be coming around. But usage driven by resignation looks identical to usage driven by enthusiasm on a dashboard.

Pew’s own data show that people adopt a technology they expect will make society worse, often because they’re afraid of falling behind without it. The room is staring at proof of its success, which is actually evidence of the problem.

Contrary to what many in the industry say, this isn’t a gap between people who understand AI and those who don’t. It’s a gap between two groups looking at the same technology and seeing different things.

Which raises a real question: how does a room full of smart people keep cheering while everyone outside walks the other way? Is there something the industry is missing?

Why can’t the AI insiders hear the naysayers?

I don’t think it’s one thing, and I won’t pretend this is the complete list. But here are the five I keep running into, stacked on top of each other — and none of them requires anyone to be a villain.

(1) They’re using different software.

Andrej Karpathy made the point that someone paying $200 a month for a frontier coding model is basically using a different product than someone who tried a free chatbot to plan a wedding six months ago.

The insider’s daily experience genuinely is remarkable — per that same Stanford report, Google’s Gemini Deep Think earned a gold-medal score at the International Mathematical Olympiad. The same class of model reads an analog clock correctly about half the time.

The power user meets the gold medalist. Everyone else meets the broken support bot and the slop in their feed. Both experiences are real. They’re just nowhere near each other.

(2) “If we don’t build it, China will.”

The room’s trump card, and the one you hear most once every other argument runs out. It has real substance — the competition is not imaginary — but notice what the argument does in a conversation: it never has to engage the concern in front of it. Someone raises their water supply, their power bill, their job, and the answer is Beijing.

Once the race frame is accepted, public opinion stops being input and becomes friction, because you don’t hold a town hall about an arms race. The move works so well precisely because it’s partly true, which is what makes it such an effective permission slip to stop listening.

(3) The room filters out doubt.

Nobody skeptical stays in the group chat. The feed rewards the boldest claim, dissent reads as not getting it, and the doubters quietly leave. What’s left isn’t a sample of informed opinion — it’s a survivorship artifact. The volume stays at eleven because everyone at seven already walked out.

(4) The costs are invisible from the inside.

The insider never saw the electric bill in Virginia that went from about $100 a month to $281 after the data centers moved in. Never sits in the zoning meeting — and there have been a lot of those; local opposition blocked or delayed some 75 data center projects worth around $130 billion in the first quarter of this year alone. Never gets laid off and told to reskill. The benefits of AI concentrate where the insiders live. The costs land where they don’t. They haven’t weighed the public’s concerns and rejected them. The concerns just never appear in their field of view.

(5) Concerns arrive in the wrong vocabulary.

When someone says “AI is stealing artists’ work,” the room’s reflex is to correct the technical framing — that’s not how training works, actually — instead of hearing the grievance underneath.

Legitimacy gets gatekept by fluency: if you can’t phrase your worry in the room’s language, it doesn’t count as an objection, just a misunderstanding to be corrected. The person walks away having been told their concern was wrong on a technicality. The room walks away having answered nothing.

Any one of these is forgivable.

Stacked together, they create a room that structurally cannot hear anyone outside. Different software makes the insiders sincere. Incentives make them motivated. Selection makes them unanimous. Invisible costs make them confident. Fluency decides which complaints even count.

And the race with China makes it all feel too urgent to stop and listen.

What the general public actually thinks

Meanwhile, outside the room, two things are happening in layers.

The surface layer is fatigued. AI is bolted onto every product, whether it helps or not. Every keynote is revolutionary.

After two years at full volume, tuning out is more self-defense than anything. When roughly two-thirds of people say AI is moving too fast, some of that is about the technology. A lot of it is about the constant pitch.

But, underneath the fatigue, I think there is something heavier going on. I believe many suspect that this AI isn’t being built for people like them.

That it was trained on the public’s work, runs on the public’s water and power, and exists mainly to make a small group of already-wealthy people wealthier — possibly at the cost of everyone else’s livelihood.

You can tell how mainstream that suspicion has gotten by who’s responding to it: Bernie Sanders has introduced a bill to give the public a 50% ownership stake in the largest AI companies, and Donald Trump has told reporters he’s exploring federal stakes in AI firms so the American public “essentially becomes a partner.”

When those two circle the same idea, the grievance underneath it has stopped being fringe.

.

And the industry’s answer to all of this has mostly been an ultimatum… get on board or get left behind.

Someone says “I’m worried about my job,” and the room says “learn to prompt.” That’s not an answer. That’s a slogan aimed at the wrong question.

Meet people where they are

The fix for the echo chamber was never louder hype, nor is it nicer messaging. It’s substance — things people can feel in their own lives.

• Power bills that go down, not up.

• A straight answer about what a data center costs the town next to it.

• Products that fix a real problem before adding a chatbot to a toaster.

• Taking “what happens to my work?” seriously as a question, instead of treating it as resistance to be managed.

The people outside the ‘AI hype echo chamber’ aren’t confused about AI. They’re waiting for it to do something for THEM, and mostly it hasn’t yet.

When it does, the hype won’t be necessary, because you don’t have to hype something people can already hold in their hands.

More from me:

Learn how to pitch a business the fun way.

• Get Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Wanting to stay ahead in search? Get Pluto - my FREE AI Visibility Chrome extension

• Pluto - The Free AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Share the newsletter and get rewards

• Refer friends for rewards: Get cool rewards (like my card game) when you refer people to my newsletter. Check out where you stand on the leaderboard.

What is Startup Weekend?

For those who haven’t been to one…

Startup Weekend is a 54-hour event where strangers form teams around ideas pitched on Friday night, build something over the weekend, and present to judges on Sunday.

It’s part hackathon, part business plan competition, part crash course in entrepreneurship.

The format has been running for over a decade, but what people build in those 54 hours keeps evolving with the tools available.

This year’s numbers

Friday night opened with 72 pitches — one minute each to sell an idea to the room. Judges and audience voting narrowed that down to 15 teams that would actually form and build over the weekend.

By Sunday’s presentations, there was almost $140,000 in real investment commitments made to teams that had existed for less than three days.

This year’s Missouri Startup Weekend drew participants from across the state — students, working professionals, serial entrepreneurs, and plenty of people trying their first startup.

Before the energy of all this excitement away, I wanted to jot down a few thoughts from the weekend.

Lots of conversations about AI, as expected

I got to talk to ALOT of people building with AI, or using AI in some shape, way, or form.

That alone feels widely different from even a year ago.

Autonomous bulldozer??

One of the most memorable moments on the hardware side was at EquipmentShare’s robotics lab.

They let us remotely operate a robotic bulldozer from their technology center, with the actual machine sitting in a barn in Fulton.

The same bulldozer also accepts voice commands from the phone that you can probably just barely see above the Xbox controller. It can also move autonomously within the space.

Another interesting feature is that if you look really closely at these images you can see cones in the middle of the screen. If you try to drive the bulldozer into the cones, a ‘shine detection’ sensor will detect the cones and stop the bulldozer.

Let’s say a worker on the job site falls over while wearing a shiny protective vest; this feature helps the robotic bulldozer avoid running over the worker.

As you may expecty, safety is a big concern when rolling something like this out in real life.

Here is my friend Caleb picking up some dirt with his new bulldozer

Adoption is faster than it’s ever been

I stopped by to see my friend and student Andrew, who was already designing a MVP for the AirWaze in Claude Design.

That specific Claude Design feature had launched roughly 48 hours earlier — Anthropic released it on April 17 — and here it was, already in a student’s workflow at a weekend event in Missouri.

That’s a data point I keep coming back to.

The lag between “launched at a frontier Silicon Valley AI lab” and “a college student in the Midwest is shipping with it” used to be measured in months or years.

This weekend, it was measured in days. Whatever you think about the pace of AI itself, the pace of adoption around it is now the more striking number.

That’s overwhelming, but also kind of cool.

Thoughts from the top pitches

As the creator of a card game designed to help you pitch your ideas better, you bet I had thoughts on what teams could have done better, new trends, and what they did well.

A few of top teans — quick impressions

A handful of the teams that placed, with the one-liner I scribbled down for each:

• Tonetta — coaches people to speak more clearly on sales calls

• Unileaae — student housing search, because finding it currently takes forever

• EQUIPR — lab equipment rental; they pointed at a roughly $1B market

• CMPLX — Zillow for off-market real estate deals

• CRANK — student organization monetization and event discovery; the team itself had more energy than anyone else in the room

• SBA Loans App — an agent to help small businesses through SBA loans; ~$33B in SBA loans annually

Some of my favorites

The ones I thought had the sharpest shot:

• RIVE — a better way to do health savings accounts - cut out the middle man, and see if the consumer comes out better.

• Airwaze — Helping with manual flight loggined

• AI Dungeon Masters — AI to help run DnD games aimed at the ~30M D&D players

• HerdRoute — Uber for farm animals, connecting auction houses, buyers, and drivers. Genuinely novel, and the kind of unsexy logistics problem that real businesses get built on

I wouldn’t recommend using Claude for presentations just yet.

Since I was just singing praises of the speed of adoption, I do one to give one honest note about Claude for pitches like this.

For actual pitch decks and presentations, I would not use Claude just yet.

Some teams presented fully Claude-made presentations. They were super difficult to read and understand. They were very text-heavy, making them almost impossible to digest.

Instead of being a visual aid, they probably actually hurt their presentations

Demos of functional MVPs

The single biggest change from past startup weekends is that almost every team had at least an MVP.

For non-technical folks, that is a minimum viable product - something visual you could actually interact with or watch run.

So, we are talking not a sketch on a napkin, not a Figma mockup, and definitely not three screenshots taped to a Google Slide.

Pretty well-designed working demos.

Tools like Claude, Cursor, v0, and now Claude Design have collapsed the distance between “idea” and “something I can show you.”

54 hours used to be barely enough to nail down a concept and get a mock-up of what your idea was.

Today, it’s enough to ship a rough but pretty functional, nice-looking version.

• I talk more about the speed of software development today in my newsletter: Everyone Gets a Jarvis - the build versus buy math,

What the best pitches had in common

The strongest pitches of the weekend all did one thing really wel,

They incorporated direct customer feedback, and they quoted customers by name.

One of the very best teams went a step further and played short video clips of real customers discussing the validity problem.

You can’t fake that, and you can feel the difference the second it hits the room.

If I had one piece of advice for next year’s teams, it would be to spend less time on your slide design and more time getting on the phone with five potential customers.

Then put them on the screen and talk about how they love what you are doing. It sells really well.

That’s my weekend.

Fulton bulldozer, a student building in a two-day-old AI tool, and a field of teams that mostly showed up with something that actually worked. Good weekend to be in Missouri.

More from me:

Learn how to pitch a business the fun way.

• Get Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Wanting to stay ahead in search? Get Pluto - my FREE AI Visibility Chrome extension

• Pluto - The Free AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Share the newsletter and get rewards

• Refer friends for rewards: Get cool rewards (like my card game) for referring people to my newsletter. Check out where you stand on the leaderboard.

You open Claude Code.

You type your first prompt of the morning.

Nineteen minutes later, it stops working. Wait what?

You get just... a polite little message that your session limit is used up, and to come back in four hours.

This is what started happening to thousands of developers on March 23rd.

Specifically, many developers paying $200 a month for Anthropic’s top-tier Max plan. AKA, the one that’s supposed to give you 20x usage.

One Reddit thread called “20x max usage gone in 19 minutes” hit 330 comments in 24 hours.

Another got 360+ comments in six days.

One subscriber commented, “It’s maxed out every Monday and resets at Saturday. Out of 30 days I get to use Claude 12.”

So, what is going on here?

Is this a bug? Corporate incompetence? Bait and switch?

Kind of. But also — and this is the more interesting answer — it might be one of the most deliberate strategic plays in tech right now.

Let me explain.

First, a stat that will blow you away

Before we get into Antrhropic’s master plan territory, there’s one stat that reframes this entire story.

Anthropic’s $200-per-month Claude Code Max subscription can consume up to $5,000 worth of compute per user, per month.

That’s according to an internal analysis by Cursor, reported by Forbes.

Some analysts think the ceiling is even higher — up to $25,000 for the heaviest power users.

You pay $200. It costs up to $5,000 to serve you. Anthropic eats the difference.

Every. Single. Month.

Imagine a steakhouse charging you $20 for a ribeye that costs them $100 to source, cook, and plate. They’re not running a restaurant — they’re running a very expensive customer acquisition strategy, and hoping they figure out the unit economics before the investors stop picking up the tab.

That’s AI subscriptions in 2026.

One developer dug through his own session logs and found he’d consumed 10 billion tokens across eight months of Claude Code usage.

At retail API pricing, that should have cost over $15,000.

He paid roughly $800 on the Max plan. A 93% subsidy — funded a massive venture bet.

So, why does this cause your session to die in 19 minutes?

Here’s where the bug vs. master plan question actually gets interesting.

The honest answer is possibly both at once.

On the bug side — a developer reverse-engineered the Claude Code binary and found two independent caching bugs that were silently inflating token costs by 10–20x. Real bugs. Confirmed by Anthropic.

Also, in early March, they ran a promotion that doubled usage limits for two weeks, then quietly ended it on March 28th — so users who got used to the doubled limits felt a sudden cliff when it expired.

Then, Anthropic added peak-hour throttling on top of all of that.

The four things hit at once.

So, yes, a list of things happened:

• Some of the 19-minute sessions are a genuine bug.

• Some are policy changes communicated poorly.

• Some are the promotion hangover.

But beneath it all, there is a deeper reason why usage limits exist at all.

That is a $4,800 gap between what you pay and what Anthropic spends to serve you.

And, when the math doesn’t work, something has to give…

Okay, but why would any company do this on purpose?

This is Anthropic’s master plan part.

Think about how Uber launched. Rides were absurdly cheap for years because Uber needed you to stop taking taxis and start opening the Uber app every time you needed a ride. They bought your habit with subsidized prices, knowing the real monetization would come later.

Or think about Amazon Prime. When it launched, free two-day shipping was an economic disaster for Amazon. They lost money on shipping constantly. But they also discovered that Prime members spent dramatically more than non-members. The habit was worth more than the short-term loss.

AI subscriptions are running the same exact playbook.

Anthropic and its investors need your workflow. They need Claude to be the thing you reach for automatically — for code, for writing, for thinking through problems.

If they charge you the real price of $5,000 a month right now, then you don’t subscribe over competitors.

So, they charge $200, absorb the loss, and bet that by the time they need to close that gap, two things will be true:

• You’ll be too dependent on the tool to leave,

• And the actual cost of computing will have fallen enough for the gap to close on its own.

  

Here’s why that second bet is actually well-founded

This is the part of the story most people miss completely.

The cost of running AI — what the industry calls cost to compute, measured in cost per token — is falling faster than almost any technology in history.

In late 2022, running a GPT-4-class model cost around $20 per million tokens. Today, equivalent performance costs about $0.40.

That’s a 1,000x reduction in three years.

Faster than PC computing. Faster than internet bandwidth during the dotcom boom.

To put that in perspective: if airline tickets had fallen at the same rate, a cross-country flight would now cost about a quarter.

And it’s not slowing down.

• AWS cut GPU pricing 44% in mid-2025. Spot GPU prices dropped 88% in less than two years.

• NVIDIA’s next chip generation promises another 10x reduction in token cost.

• Anthropic signed a deal with Google for up to one million specialized AI chips specifically to drive their compute costs down.

They are betting is a massive decrease in what it actually costs Anthropic to serve you one session.

Right now, that cost is $5,000 against your $200.

If computing keeps falling at its current rate, that same $5,000 worth of AI work might cost $200 to deliver in a few years.

Potentially, this could drop to much lower than $200 in further years.

That’s the bet, and their ‘master plan.’

So, what does this mean for you right now?

A few honest things.

If you are a Claude Coder, the frustration is completely valid.

Anthropic’s own team acknowledged people are hitting limits “way faster than expected” and called it their top priority.

The communication around the limit changes has been poor. The bugs are real.

If you’re a heavy user hitting walls constantly, you’re not imagining it and you’re not overreacting.

But you’re also, objectively, getting an extraordinary deal right now. One developer tracked $5,623 of API-equivalent compute in a single month of coding — and paid $200. Even with the limits tightening, the subsidy is massive.

The window where AI tools cost a fraction of their actual value won’t last forever. Anthropic needs the economics to work eventually.

Either compute costs fall fast enough that the gap closes naturally — which the data suggests is genuinely possible — or prices go up, or both.

We even got a preview of what abundance looks like: in December 2025, Anthropic doubled usage limits for a week by tapping idle holiday compute capacity.

Then January came and the limits snapped back. People thought something broke. Nothing broke.

The free lunch just ended for the week.

Bug or master plan?

I would say both.

The bugs are real and Anthropic handled them badly. The limits are real and tightening. The frustration of developers is completely earned.

BUT, the $4,800 gap between what you pay and what it costs to serve is definitely a crazy long-term bet — and one that makes a lot of sense if you believe (as the data strongly suggests) that the cost of compute is going to keep falling fast enough to eventually make the math work.

This is why I believe most important number to watch in AI right now is the cost per token.

Because that’s the number that determines whether your $200 plan stays a $5,000 subsidy forever — or whether the world where AI is genuinely cheap and abundant actually arrives.

As of the last few years, that number is falling like a stone.

More from me:

Learn how to pitch a business the fun way.

• Get Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Wanting to stay ahead in search? Get Pluto - my FREE AI Visibility Chrome extension

• Pluto - The Free AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Share the newsletter and get rewards

• Refer friends for rewards: Get cool rewards (like my card game) for referring people to my newsletter. Check out where you stand on the leaderboard.

There’s a scene in Iron Man where Tony Stark walks into his garage, says a few words, and Jarvis conjures software and full 3D visuals on demand. No meetings, no tickets, no waiting three weeks for a sprint.

That’s what AI agents are turning into today — Claude Code, Manus, OpenClaw. You describe what you need. It gets built. With Whisprflow, people are making this into a reality FAST.

So the obvious question that comes to me is… why would anyone pay for software again? When should you build or buy software? And, what are actual moats in the software business now?

Stick with me, because the answer is more interesting than you’d think.

Some context… the wall came down

Buying software was, for a long time, just the only practical option. Building your own took months, cost a fortune, and needed specialists most companies didn’t have sitting around. So everyone subscribed. The whole SaaS industry was built on that being true.

It’s getting less true fast. Retool surveyed 817 enterprise builders last year — mixed teams, not just developers, real people across ops, finance, product, IT. 35% had already replaced at least one SaaS tool with something they built themselves. 78% plan to build more in 2026.

One company in that survey had been paying $20,000 a year for a tool. Vendor support was so slow that their ops lead rebuilt the whole thing with an AI agent before a single support ticket got answered. Now their default question for any new software need is: why can’t we just build this?

It’s a fair question to be asking. But it’s not the only one.

Can build is not the same as should build

An AI agent could probably build you a version of Salesforce. Custom, fast, no per-seat licensing, no paying for features you’ve never touched. That’s not hypothetical anymore.

But think about what Salesforce actually is under the hood. Twenty years of security audits. A team of hundreds whose entire job is making sure your customer data stays where it should. SOC 2 certification, GDPR compliance, breach response protocols. The checkbox your enterprise customers look for before they’ll sign a contract with you. And when something does go wrong — a company with real liability on the other end of the phone.

A Jarvis-built CRM has none of that. It has you.

Trust was always the moat. What changed is that it now has to win the argument on its own, because cost no longer makes the decision automatically.

Before, building was expensive enough that buying was the default for almost everything — and trust came bundled in whether you thought about it or not. Now building is cheap, and trust has to be a conscious choice. For some things you make that choice and it’s fine. For others, it’s a liability you haven’t fully accounted for.

The build vs. buy math is changing — fast

Three years ago this whole conversation was mostly theoretical. Building was expensive enough that the economics made the decision for you. You bought things, because building was out of reach.

Retool’s CEO David Hsu said it plainly to Newsweek this February: “The cost of building custom software has collapsed. Two years ago, a custom internal tool might take an engineering team weeks and cost six figures. Today, a business operations lead with the right platform can have a working prototype in a day or two. That’s a structural change, not a cyclical one. When the cost of building drops by an order of magnitude but the cost of buying stays flat, the math changes for every company.”

Here’s what that actually looks like. Same internal reporting tool. 10-person team. Three-year view.

Build it in 2020:

• Developer cost to build: $8,000–$12,000 (US contractor rates run $100–150/hr; a two-week build is 80+ hours)

• Server hosting on AWS: ~$50/month

• Ongoing maintenance: more developer time every time something breaks or changes

• Security and compliance: your problem, your cost

• Three-year total: $15,000+ before anything goes wrong

Build it in 2026 with your Jarvis:

• Agent cost to build: near zero — an afternoon

• Server hosting on AWS: still ~$50/month — this doesn’t go away

• Ongoing maintenance: still real — someone has to own it when it breaks

• Security and compliance: still your problem

• Three-year total: ~$1,800 in hosting — if nothing goes wrong

Buy a SaaS tool instead:

• 10 users at $29/user/month (median entry-level SaaS pricing in 2025): $290/month

• Annual price increases averaging 8–12% per year — SaaS inflation is running nearly 5x the general market rate right now

• Maintenance, hosting, security, uptime: their problem, included in the price

• Three-year total: ~$11,000–$13,000 and climbing

On paper, building with Jarvis wins easily. And for a simple internal tool with no sensitive data, it probably does. But that “if nothing goes wrong” is doing a lot of work. METR ran a proper randomized controlled trial last year — experienced developers, real complex codebases, not toy projects. When they used AI tools, they took 19% longer than without, and thought they were 20% faster. The first 90% of building is genuinely fast now. The last 10% — edge cases, security hardening, keeping it running when something breaks six months from now — still takes what it always took.

One team built an AI-assisted internal dashboard in two weeks. Six months later nobody could safely touch it. The developer who built it had moved on. They scrapped it and bought a SaaS tool, which is probably what they should have done from day one.

Who we buy software from is about to change

A friend of mine was telling me recently he wants to build a social media scheduling app. I asked him the obvious question: if anyone with a Jarvis can build a scheduling app in a weekend, what’s your advantage?

He didn’t hesitate. His advantage is the people he already knows. The marketers, the agency owners, the founders in his network who trust him personally. Who’ve seen him show up, fix problems, respond at midnight when something breaks. Who’d pay for his tool over a stranger’s identical tool on Product Hunt — not because his is better, but because he built it and they know him.

That’s not a new idea. People have always bought from people they trust. What’s new is that it’s becoming one of the only defensible advantages left in software. When the product itself can be replicated in an afternoon, the relationship is what remains.

Info-Tech Research Group found in their 2025 buyer research that satisfaction is now rooted in the vendor relationship, not the feature set. People leave when the relationship breaks down. Rocket Software found that 84% of consumers say knowing a company uses secure, up-to-date technology makes them more confident doing business with that brand.

The product got you in the door. The relationship keeps you there. And for a lot of buyers, the relationship is the reason they knocked in the first place.

The moats that still hold

So what actually defends a software business in a world where anyone can build? The feature list is gone. The UI is gone. “We built it first” is gone. Here’s what remains.

Network effects. The product gets more valuable as more people use it. WhatsApp isn’t better than its competitors because of the code — it’s better because everyone you know is already on it. Slack, LinkedIn, Uber. Your Jarvis can build a ride-sharing app. It cannot build the driver network. The product and the network are inseparable, and the network took years.

The content and library effect. Spotify has 100 million tracks. Netflix spent billions on licensed content. A medical records platform has fifteen years of patient data structured exactly the way clinicians need it. These libraries cost more to build than any subscription price — often by orders of magnitude. Nobody is going to have their Jarvis recreate that. The cost in time, money, and licensing negotiation makes it untouchable. The content is the moat, not the player.

Switching cost moats. When your entire team is trained on a tool, your three years of data lives inside it, your workflows and integrations are built around it — leaving costs more than staying, even if a cheaper or better alternative exists. This is why enterprise software companies sleep well at night. Getting out is a project. Most companies never start it.

Compliance and regulatory moats. SOC 2, HIPAA, FedRAMP, PCI-DSS. These certifications take years to earn and require ongoing investment to maintain. An enterprise customer’s procurement team isn’t going to approve your Jarvis-built alternative because it doesn’t have the paperwork. The certification isn’t the product — it’s the permission slip to sell the product to anyone serious.

Trust and reputation moats. The oldest one. Eight years of uptime. A support team that picks up the phone. A track record that your customers’ customers can point to when someone asks why they use you. My friend’s scheduling app wins not because of features but because his buyers would rather give money to someone they know than take a chance on someone they don’t. At scale, this is what brand actually means.

Notice what’s not on this list… the code itself. The features. The UI. The clever onboarding flow.

Those were always easier to copy than people admitted. Now they’re trivial. What’s left is everything that takes years to build and can’t be prompted into existence.

Taste and judgement is something I still think is important and hear people talking about alot, but I am not sure that a successful judgement/taste couldn’t just eventually be copied.

So, when do you build, and when do you buy?

My auditing tool was easy. Internal only, no customer data, unique to my workflow, and I can maintain it myself. Two weeks of developer time versus one afternoon — that math does itself.

But I’d never have an AI agent build something that touches customer payment data, or that an enterprise client’s security team is going to review. Not because I couldn’t. Because the question stopped being about capability a while ago.

Build with your Jarvis when the tool is internal, the data isn’t sensitive, the workflow is specific to you, and you can own what happens to it when the person who built it moves on. Build when no vendor does what you actually need.

Buy from a vendor when you’re handling customer data at scale, when you’re in a regulated industry, when your customers or partners will audit your stack, or when getting it wrong is a company-ending problem. Buy when what you actually need is the decade of trust sitting underneath the product, not just the product.

The gray zone — and this is where most companies actually live — is building the custom layer on top of a trusted foundation.

Your Jarvis writes the workflow, the interface, and the internal tooling. But it sits on Stripe, on AWS, on whatever infrastructure took ten years to make trustworthy. You get the customization without taking on the liability of replacing the thing underneath.

Trust was always worth paying for. Now you just have to decide when.

More from me:

Learn how to pitch a business the fun way.

• Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Wanting to stay ahead in search? Get my FREE AI Visibility Chrome extension

• Aaron’s AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Share the newsletter and get rewards

• Refer friends for rewards: Get cool rewards (like my card game) for referring people to my newsletter. Check out where you stand on the leaderboard.

However, what I am is pro-community, pro-transparency, and increasingly convinced that the way we’re building AI infrastructure right now is repeating a very familiar mistake — moving fast, overpromising benefits to local communities, and leaving residents to figure out the consequences after the fact.

That’s what this newsletter is about.

The invisible cost of every prompt

You typed a question into ChatGPT this morning. Maybe you asked it to clean up an email, summarize an article, or help you think through a decision. It answered in seconds. You moved on with your day.

However, there is one fact that you probably didn’t think about… that chat exchange drank roughly a bottle of water.

Not metaphorically. Physically. Somewhere in a warehouse-sized building in Virginia, Texas, or Arizona, a cooling system pulled real water from a real source to keep the servers that processed your prompt from overheating.

Most of that water didn’t go back — it evaporated into thin air.

Multiply that by billions of prompts a day, and you start to understand why some communities are watching their wells run dry.

So, where does the water actually come from

This is worth explaining specifically, because most people picture data centers as some abstract cloud facility — not something that physically connects to the water pipe running under their street.

When a company decides to build a data center, they evaluate locations based on a few key factors — available land, favorable tax incentives, access to fiber infrastructure, cheap electricity, and water supply. The water question is a pretty big deal. These facilities require enormous, continuous cooling capacity to keep their servers from overheating.

Once a site is selected, there are typically three ways it gets water:

1. Municipal water supply

This is the most common. In most cases, data centers source their water directly from local public water utilities via their potable drinking water delivery systems — the same pipes that supply homes, businesses, and farms. That's right: in many cases, a data center is competing with residential households for the same municipal water allocation.

2. Groundwater via on-site wells

In some instances, particularly in remote or rural locations, data centers tap into underlying aquifers directly through on-site wells. This is especially common in areas where municipal infrastructure can't meet demand — and it often happens with little to no regulatory oversight.

3. Surface water from rivers and lakes

Some data centers draw directly from nearby bodies of water, which is why you'll find them clustered along rivers like the Mississippi.

One industry assessment found that 80–90% of the water used by data centers is drawn from “blue” water sources — lakes, rivers, or aquifers — often the same sources that provide community drinking water. In practice, many data centers simply hook up to the local water mains.

The critical part is that approximately 80% of the water withdrawn by data centers evaporates.

In other words, the water isn’t recycled. It doesn’t come back. It’s gone — drawn from the same finite sources communities depend on, then vented into the atmosphere.

Let’s talk scale

Each 100-word AI prompt is estimated to use roughly one bottle of water. Not the electricity — the water, just to keep the chips cool.

Note that this estimate was from a 2023 study, and the ratio has likely improved with newer models, but the reality remains.

A medium-sized data center can consume up to 110 million gallons of water per year — equivalent to the annual usage of approximately 1,000 households. Larger data centers can each consume up to 5 million gallons per day, equivalent to a small town of 10,000 to 50,000 people.

A study by researchers at the University of Houston found that data centers in Texas alone will use 49 billion gallons of water in 2025, and as much as 399 billion gallons in 2030 — equivalent to drawing down Lake Mead, the largest reservoir in the United States, by more than 16 feet in a single year.

And most of this is being built in places that are already water-stressed. Phoenix. Las Vegas. The Sun Belt. The desert Southwest — regions chosen for their cheap land and favorable tax environments, not their abundant water supply.

We’ve seen this movie before… Bitcoin Mining

If you’ve been paying attention to the crypto space over the last few years, this story has a familiar structure.

When Bitcoin mining operations started setting up in small towns across America — lured by cheap power, cheap land, and minimal regulation — they came with promises of economic revitalization. Good jobs. Tax revenue. Community investment.

What many communities actually got was different. Complaints about bitcoin mining noise make headlines more often than the average investor realizes. Mining noise turned a small town in North Carolina against its local miners. Residents in rural Ohio complained to their planning commission. As did Niagara Falls residents. A Montana county voted unanimously to pass new zoning rules for current and future mining operations.

In Granbury, Texas, residents near a massive Bitcoin mining facility described a constant roar — the sound of nearly 60,000 computers cooling around the clock, less than 100 yards from mobile homes where dozens of families live. A 2,000-foot, 24-foot-tall soundproofing wall was installed. Residents say it did little. Some are trying to incorporate their neighborhood into a new city just to gain the legal standing to create a noise ordinance.

This pattern has many parallels. An industry moves into a community with promises of economic benefit, operates with minimal transparency, imposes real costs on residents, and then points to its tax contributions when confronted. The residents who organized, complained, or tried to fight back were often characterized as anti-progress, anti-jobs, and anti-technology.

Sound familiar?

AI data centers aren’t noisy in the same way Bitcoin mines are. But the underlying dynamic — industries imposing costs on communities that never got a real vote — is identical. The medium is different. The model is the same.

The job’s promise doesn’t hold up

One of the primary arguments used to gain community approval for data centers is job creation. And I want to address this directly, because it’s the argument most likely to be used to dismiss legitimate concerns: “Think of the jobs.”

The data tells a more complicated story.

Most of the jobs created by data centers are short-term construction jobs that disappear once the building is finished. After that, a mega data center — the size of dozens of football fields — operates with only a few dozen permanent workers. And as AI and automation expand, even those jobs are shrinking.

A 2025 brief from University of Michigan researchers put it plainly: “Data centers do not bring high-paying tech jobs to local communities.” A recent analysis found that in Virginia, the investment required to create a permanent data center job was nearly 100 times that required to create comparable jobs in other industries.

The University of Michigan also found that the jobs data centers do create locally are “typically low-wage, term-limited, non-technical positions such as security, maintenance and janitorial work.”

The subsidy picture is also troubling. Almost half of state data center subsidies do not require any job creation at all. When the Indiana governor announced a Google data center and touted 200 jobs, Google was not legally obligated to create any jobs in exchange for its subsidy. A local agreement stipulated only 30 permanent positions.

The industry has spent millions on marketing to counter this narrative. One industry-affiliated group spent at least $700,000 on digital marketing in Virginia alone in 2024, running ads emphasizing job creation and clean energy investment — targeted primarily at policymakers rather than local residents.

To be fair, there is a genuine case to be made for construction employment and local tax revenue — and in some communities, that revenue is substantial. Loudoun County, Virginia, collects nearly $900 million annually from data center taxes. But tax revenue and community benefit are not the same thing. A community can be a net recipient of tax dollars and still have its water table depleted, its power grid strained, and its residents making decisions based on promises that were never legally required to be kept.

The transparency gap

The core of my argument (whic I don’t think is too crazy) is that communities deserve to know what they’re agreeing to.

Very often, data centers negotiate with local governments under non-disclosure agreements. They frequently decline to share information about water usage, energy usage, and air quality impacts. Communities are expected to make decisions about zoning, infrastructure investment, and tax incentives without basic information about what they’re actually agreeing to.

Nearly 200 community groups across the country are currently active in opposing data center development. Between April and June 2025 alone, community opposition blocked or delayed 20 data center projects representing $98 billion in potential investment.

These aren’t Luddites. They’re residents who have watched their wells go dry, seen their power bills increase, and been told it’s all for the greater good of AI — which, by the way, they may or may not use.

Some communities have started pushing back effectively. In Lancaster, Pennsylvania, a community benefits agreement committed data center developers to a water-use cap, noise and emissions controls, and a $20 million contribution to local economic and sustainable development. That’s what it looks like when a community has real information and real leverage.

That should be the standard, not the exception.

How can we be pro-AI and pro-community?

The technology to build more sustainable data centers exists right now. Direct-to-chip cooling, closed-loop water systems, immersion cooling, smart site selection — Cornell researchers found that combining these approaches could reduce water consumption by 86% and carbon emissions by 73%. The solutions aren’t sci-fi.

The barrier isn’t technical. It’s that nobody is requiring companies to use them. Tax incentives aren’t tied to water conservation commitments. Permits don’t require disclosure of consumption data. NDA agreements let companies walk into communities with promises that no one enforces.

To be genuinely fair, there is a serious counterargument here. And, it deserves a direct response. The strongest case for tolerating the resource costs of AI infrastructure goes something like this — AI will accelerate drug discovery, reduce energy waste across entire industries, dramatically improve agricultural efficiency, and generate productivity gains that dwarf the water and electricity consumed to build it.

On a long enough time horizon, the argument goes, a data center in Phoenix that depletes a local aquifer might be a reasonable trade for the infrastructure that cures a disease or decarbonizes a supply chain. That argument is not crazy. It might even be right.

But the problem I have is that the communities hosting these facilities aren’t the ones making that calculation. The trade-off is being decided by investors, hyperscalers, and state-level officials chasing tax revenue — not by the family in Georgia whose well ran dry, or the residents of Granbury who can’t sit on their porch because of the noise. The benefits of AI are diffuse and global. The costs of AI infrastructure are local and specific. That asymmetry is the issue.

If the productivity gains are real and large, then the industry can afford to compensate communities properly, disclose what it’s taking, and build with consent rather than around it.

AI is a legitimate technological revolution. The infrastructure behind it is real, necessary, and worth building. But the way we’re building it right now — with opacity, with overstated job promises, with communities absorbing costs they didn’t agree to and can’t quantify — is a version of a story we’ve already lived through with Bitcoin mining operations

.

The lesson from all of them is the same: communities that demand transparency and real commitments upfront are far better off than those who find out later what they signed up for.

The water bill is real. Someone is paying it.

Right now, it’s mostly the communities that never got to read the contract.

More from me:

Learn how to pitch a business the fun way.

• Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Wanting to stay ahead in search? Get my FREE AI Visibility Chrome extension

• Aaron’s AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Share the newsletter and get rewards

• Refer friends for rewards: Get cool rewards (like my card game) for referring people to my newsletter. Check out where you stand on the leaderboard.

You’d type in a half-baked prompt—something like “In the future, marketers will…”—and then just sit there like: wait… why is this kind of good? Not “publish it as-is” good. More like “this is the first time the internet has talked back in complete sentences,” good.

Back then, it felt less like productivity and more like discovering a weird new instrument. You weren’t using it to ship blog posts. You were using it to test the boundaries: Can it stay on topic? Can it write a joke? Can it sound like a human who’s had coffee?

Since then, my setup has slowly evolved into a little AI pit crew. Some tools help me write. Some help me think. Some help me stop doomscrolling search results like it’s 2014. And a few quietly run workflows in the background like tiny robots who never ask for PTO.

This isn’t an “ultimate list” or a “best of 2026” roundup. It’s the set of tools I actually reach for (free and paid) when I’m writing, researching, building, automating, or trying to turn a messy idea into something shippable.

Most of them you can try for free because that’s how they got me hooked, too.

Here is my quick list:

1. Claude — favorite writing assistant (best for clean drafts + voice)

2. ChatGPT — research assistant + GPT-training / prompt systems

3. Gemini — content scoring lens (pressure-testing content against Google-y guidelines)

4. Perplexity — search engine replacement (fast answers with citations)

5. NotebookLM — turning docs into pitch decks + presentations

6. Zapier — general automations + Slack bots

7. Make — more flexible workflows (when Zapier feels limiting)

8. n8n — power workflows + customization (great if you want more control/self-hosting)

9. Gamma — proposal building (quick, presentable decks/docs)

10. Grammarly — grammar + readability polish

11. Replit - coding when I want to move fast

1. Claude — my favorite writing & coding assistant

If you’ve ever tried to write a strong opening paragraph while your brain is still buffering, Claude is the friend who goes, “Cool, here are three intros—pick the one that doesn’t make you cringe.”

I like Claude best when I already know what I mean, but I need the words to land. It’s especially good at tone smoothing—making something feel more “agency team with taste” and less “internal wiki doc that accidentally leaked.”

The big win is speed without sacrificing voice. The risk is the same as any writing assistant: if you let it drive too long, you’ll get prose that’s technically fine but emotionally beige. I use it like a blade sharpener, not like a ghostwriter.

And, it will one-shot any small coding task. One of my favorite use cases is creating Python scripts easily for data science & data visualization.

Say like, “I have this fat Google Sheet with this type of data, read it in, and make this epic chart.”

2. ChatGPT — research assistant and GPT-training playground

ChatGPT is where I go when I need range. Research outlines, angle exploration, “what would a skeptical VP ask,” first-pass frameworks, and quick “turn this into a system” thinking.

It’s also the tool I use when I want to prototype the way information should be packaged. If I’m training prompts, building a reusable research workflow, or stress-testing an argument, ChatGPT is the most flexible workbench.

Recently, I have been building lots of custom GPTs using ChatGPT’s “Create a GPT” feature. It’s basically a way to make your own specialized version of ChatGPT with a specific job, instructions, and (optionally) a big pile of reference material to learn from.

In my case, I took it to extremes. I built one that lets me talk to an AI version of Warren Buffett and Charlie Munger when I’m working through business questions.

Under the hood, I fed it over 30,000+ documents, from Berkshire letters and annual meeting Q&As, as the foundation. So, it’s not just generic business advice. It’s closer to their actual logic & reasoning showing up on repeat

Want to read more about this? Read my full article on training a Warren Buffett & Charlie Munger GPT on 30,000 documents.

So, the most underrated use is not “write the thing.” It’s “help me decide what the thing is.” When you use it that way, it’s less like an intern and more like a whiteboard that talks back.

3. Gemini — scoring content against Google’s search model check

Gemini is the tool I reach for when I want to pressure-test content through a Google-flavored lens.

And yes, I have a theory… Google likely uses something Gemini-like to evaluate pages at scale—not just for keywords, but for overall helpfulness signals, intent match, clarity, and whether the page actually delivers what it promises.

To be clear, I’m not claiming there’s a “Gemini score” that determines rankings. What I am saying is the direction of travel is obvious. Modern search needs models to interpret content quality and intent, and Gemini is the closest public-facing proxy we can play with.

Practically, I’ll use Gemini to ask questions like: does this answer the query cleanly, is the structure logical, and is anything missing that a real human would expect?

4. Perplexity — the “please don’t make me open 14 tabs” search engine

Perplexity is what I use when I’m in research mode and I want sources without the scavenger hunt.

The experience is basically asking a question, getting a synthesized answer, and seeing citations immediately. That workflow matters when you’re writing, pitching, or making decisions fast—because the bottleneck is rarely “lack of information.” It’s “time spent wading through junk.”

I still do manual checking (always), but Perplexity is the fastest way I know to turn curiosity into a usable starting point.

5. NotebookLM — turning docs into decks & visuals

NotebookLM shines when you’re not starting from scratch—you already have material, it’s just scattered across docs, notes, PDFs, transcripts, or “that one doc someone shared in Slack three months ago.”

This is my favorite place to do synthesis: pull in source material, ask targeted questions, generate structured summaries, and extract the narrative thread you can actually present.

I’ve used it specifically for pitch-deck prep because it helps with the hard part: not “make slides,” but “decide what matters.”

Once the story is coherent, the visuals are just packaging.

6. Zapier — general automations and Slack bots that save your sanity

Zapier is my “I just want this to work” automation layer.

If the task is straightforward—send alerts, route form fills, push updates into Slack, log something into a sheet—Zapier is usually the fastest path from idea to running system.

It’s not always the cheapest at scale, and it’s not the deepest for complex logic. But it wins on convenience and speed. When you’re building operations, shipping beats perfect.

7. Make — workflows when you want more control

When Zapier feels like it’s boxing you in, Make is the next step up for me.

It’s great when you want visual workflow building with more flexibility—multi-step routing, data manipulation, branching logic—without fully dropping into “engineer a platform” territory.

If you build even a couple of repeatable systems (content pipelines, lead routing, reporting, internal alerts), Make starts paying for itself in saved brain cycles.

8. n8n — workflows when you want ownership and customization

n8n is what I use when I want that extra layer of control and customization. It’s the “I want my automations to behave like software” option—especially appealing if you like self-hosting or you need more control over data flow.

The simple way I think about it: Zapier is the quick win, Make is the flexible middle ground, and n8n is the power-user path when you’re ready to build something that feels like infrastructure.

9. Gamma — proposal building when you want to look like you tried

Gamma is for the moment when you have something to say, but you don’t want to spend your entire afternoon aligning boxes.

It’s proposal-friendly, pitch-friendly, and “make this look presentable fast” friendly. I’ve used it for client-style proposals and internal narratives where the goal is clarity and momentum—not design awards.

The sneaky benefit is that it reduces the friction between idea and presentation. When the packaging is easy, you ship more.

10. Grammarly — polish & readability

Grammarly is still one of the simplest “make this better” tools in my stack.

I use it for grammar and readability, obviously. But I also like having an extra authorship-style sanity check when I’m worried a draft is drifting into that too-smooth, too-generic zone.

Not because I’m trying to “game” anything—more because I want writing to feel like it came from a person with a pulse.

11. Replit — coding when I want to move fast

Replit is one of my favorite tools for coding because it gets you from “idea” to “running code” with almost no ceremony.

When I’m prototyping something, testing an automation script, experimenting with an API, or building a small internal tool, Replit keeps the momentum high. It’s the difference between “I’ll do it later” and “it exists now.”

And honestly, that’s the whole point of a good tool stack: fewer roadblocks between intention and output.

Build a stack that removes friction

I had to reduce the whole setup to a few words, because most of these tools aren’t “hands-off automation.” They’re AI + expert. This means thatthe AI gives you leverage, but you still steer.

You’re the editor. The strategist. The one who knows what “good” looks like.

That distinction matters because many people enter AI hoping for a Roomba for their job. Hit start, go make coffee, come back to a finished strategy doc and a perfectly formatted deck.

In reality, the highest ROI stack is closer to you + a bunch of specialist copilots.

Claude is for writing that sounds like a person, but you still decide the argument and the tone. ChatGPT is for thinking and structuring, but you still have to pick the angle and make the calls. Gemini is for a Google-ish reality check, but it’s not the decider. Perplexity is research speed with receipts, but you still verify and synthesize. NotebookLM is for wrangling chaos into a narrative, but you still choose what goes in the story and what gets cut.

Then you’ve got the automation layer—Zapier, Make, n8n—which can get closer to hands-off, but only after you’ve done the expert work of defining the workflow. Automation doesn’t replace judgment. It replaces repetition.

Gamma is packaging. Replit is shipping code faster. Grammarly is polish and “does this read as a human wrote it” sanity checking. None of that is fully hands-off.

It’s just removing the annoying parts so you can spend more time on the parts that actually matter.

Tools don’t replace strategy. They remove friction.

And when you remove friction, you get more reps, more drafts, more experiments, more distribution, and more iteration.

That’s where the compounding happens.

If you’re building your own stack, don’t copy mine perfectly.

Copy my pattern: one tool for writing, one for research, one for synthesis, one for automation, one for visualizing.

Then keep what you actually use, and ruthlessly delete the rest. Because that monthly AI subscription bill will creep up on you fast.

Not the kind who cares how cool you say you are.

The kind who looks at you and goes: “Got any references?”

That’s the shift this new University of Toronto study nails with actual numbers.

Generative AI search engines don’t just rank pages. They construct a case. And when they need receipts, they overwhelmingly prefer earned media—third-party reviews, editorial, institutional sources—over brand-owned sites and even over social.

In other words, your website is more a claim than ‘proof’

And, AI search engines are becoming claim-skeptical machines.

About the study

The paper is called “Generative Engine Optimization (GEO): How to Dominate AI Search” by Chen, Wang, Chen, and Koudas. It’s one of the cleanest “first maps” I’ve seen of how AI search differs from Google.

They run large, controlled experiments across multiple AI search engines (ChatGPT, Perplexity, Gemini, Claude), multiple verticals (consumer electronics, automotive, software, local services), multiple regions, multiple languages, and paraphrases of the same intent.

Then, they analyze which domains get cited, what kind of domains they are (Brand vs Earned vs Social), how stable the citations are across language and phrasing, how fresh the sources are, and whether big brands get default advantage.

This matters because we’re moving from “SEO = rank a page” to “GEO = become the obvious recommendation inside a synthesized shortlist.”

And, shortlists behave differently than SERPs.

“Earned media bias” is structural

Across their experiments, AI search engines show a systematic and overwhelming bias toward earned media.

Google is a mix of brand sites, earned editorial and reviews, and social/forums (especially Reddit).

AI search, especially ChatGPT and Claude, behaves more like: Earned ≫ Brand ≫ Social. And Social often drops to near-zero.

Take automotive prompts in the US and Canada. When they compared Google vs AI search: Google still included meaningful social content. AI search frequently returned 0% social. Earned media share ballooned dramatically.

So if your growth engine is “community love”—Reddit, forums, UGC, creator buzz—some AI engines barely treat that as valid evidence.

Why is this happening?

Google is a catalog. AI search is a verdict.

Google can show you 10 options and let you decide. AI search has to pick a handful of things and explain itself.

This reality forces it into a behavior pattern that’s incredibly important for marketers to understand.

AI search optimizes for defensibility.

Earned media is defensible. It’s independent, sounds neutral, already structured as comparisons and recommendations, and has easy-to-lift verdict language like “best for X” or “top pick” or “worth it” or “avoid if...”

Brand sites are harder. They use marketing language, make claims without third-party validation, offer less comparative framing, and lean more “our product is great” than “here are tradeoffs.”

So the engine becomes an aggregator of legitimacy. This is a PR-first world, disguised as search.

AI engines aren’t one market

The paper shows the engines behave very differently from each other.

ChatGPT + Claude are earned-media purists. They lean extremely hard into third-party editorial and reviews, with minimal brand content and almost no social.

Perplexity is more of a blended ecosystem. It pulls in more social and brand sources, often including YouTube and retail.

Gemini is brand-friendlier. It includes brand sources more often than ChatGPT or Claude.

If you’re treating “AI search” as one channel, you’ll waste money. You need engine-specific distribution strategy the same way you’d plan differently for Google vs TikTok vs Amazon.

Big brand bias: unbranded prompts default to giants

They also test something that feels obvious but is now measurable: When the prompt is unbranded (”best cola,” “top soda brands”), the models disproportionately surface major market leaders.

Niche brands show up, but far less frequently.

So if you’re not a category incumbent, your job isn’t just “be better.” Your job is: Create enough third-party proof that the model feels safe mentioning you.

How bias varies by vertical

Not all industries experience the same level of earned media dominance. The pattern shifts based on what kind of information AI models need to answer confidently.

Consumer electronics shows the most extreme earned media bias (90%+) because AI trusts professional specs and reviews over manufacturer claims. Automotive is a battleground—earned media leads, but brand sites survive at around 30% because AI needs official specs, trims, and pricing data. Local services is the wild west, where AI struggles to find credible national earned media for dentists and plumbers, so it relies heavily on directories like Yelp and TripAdvisor.

Language matters more than wording

One of my favorite findings: Changing languages reshapes the entire citation ecosystem.

Some engines reuse similar authority sources across languages. Others basically switch the entire evidence universe when you ask in German vs English.

Meanwhile, paraphrasing (”give sources,” “rank them,” “use quotes”) shifts citations less than translation.

That means: If you want international visibility in AI search, “translate the page” is not a strategy. You need localized authority—earned coverage in the target language ecosystem.

Moving from reading to doing

So what do top marketers do about it?

Here are the moves I’d make if I was a CMO trying to win “the earned media filter.”

Treat earned media like performance marketing. PR can no longer be “brand awareness.” It’s now a direct input into whether AI systems will recommend you. The KPI isn’t impressions. It’s: Do we appear in the citation network for our category?

Build your “citation map.” Pick 30–50 prompts that represent your category—things like “best X for Y,” “X vs competitor,” “most reliable,” “best value,” “top tools for [persona].” Run them through the engines that matter. Then list the domains cited most often, which ones show up repeatedly, which ones show up only for certain personas or regions or languages. Those repeat domains are your kingmakers. Now your PR roadmap becomes obvious: get covered there, get reviewed there, get included in listicles there, build relationships there.

Create “justification assets.” AI search builds shortlists with reasons. So give it reasons it can steal cleanly. On your site, build pages that are scannable, structured, and explicit about tradeoffs. Think “Best for ___” pages with constraints, comparison tables, pros/cons lists, “Who this is for / not for” sections, pricing and policy clarity. If the model can’t extract a clean justification in 5 seconds, you won’t make the shortlist.

Make your brand “API-able.” This is the least sexy but most important idea in the paper: AI agents need clean data. So treat your site like it’s being read by a machine that will pull your specs, parse your pricing, summarize your warranty, compare you to others, and quote your claims back at users. That means rigorous schema, clean HTML structure, fewer fluff blocks, clear tables for specs and features, and explicit claims tied to proof.

Niche brands: stop trying to win “best overall.” If you’re not a giant, “best overall” is a trap. The model defaults to market leaders because it’s safe. Your winning lane is: best for a narrow, specific, defensible context. Best for small teams. Best for ADHD brains. Best for photographers traveling weekly. Best for people who hate subscriptions. Then you go dominate earned coverage in that niche—specialty publications, creator reviews, expert roundups, and community credibility.

The big point

In the old world, your website was the product.

In the new world, your website is the brochure. The product is your reputation graph.

And AI search is reading that graph with a bias: “Don’t tell me you’re great. Show me who else says so.”

If you want to win AI search in 2026, don’t just optimize pages. Optimize proof.

—Aaron

Become a master pitcher & inventor – the fun way

• Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Wanting to stay ahead in search? Get my FREE AI Visibility Chrome Extension.

• Aaron’s AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Share the newsletter and get rewards.

• Refer friends for rewards: Get cool rewards (like my card game) for referring people to my newsletter. Check out where you stand on the leaderboard.

In my settings, I’m going to add a default prompt layer that makes all of my search results sound like Yoda.

Not “sometimes.”

Not “when I ask.”

Always.

From now on, every answer, summary, recommendation, and explanation will come back with:

“Buy this jacket, you should.”

“Trust this source, I do not.”

“Overpriced this is, hmm.”

It’s funny. It’s also a serious signal.

Because that’s personalization in its first, most obvious form:

The same information… delivered in the way I prefer.

And once you realize that’s possible, you start seeing the bigger shift behind it.

Personalization of search is here

For two decades, search has been a shared experience.

If you and I typed the same query, we’d mostly see the same results — maybe shuffled by location, maybe nudged by history, but still recognizably the same “version of the internet.”

That model is breaking.

The next model is a private feed for YOU.

And Yoda-mode is the simplest example of the new layer:

• default prompt instructions

• preferred tone and format

• “explain it like I’m a beginner / expert”

• strict brevity vs deep research

• “give me options but decide for me”

• “challenge my assumptions”

• “only show sources I trust” (this one is coming fast)

In other words:

You’re not just searching.

You’re configuring how reality is narrated back to you.

What is personalized content, actually?

When marketers hear “personalized content,” they often think:

• “insert first name”

• “swap the hero image”

• “recommend products like Netflix”

That’s personalization inside your channels.

The new personalization is bigger:

Personalization sits on top of every channel & chat history.

A user will have a default prompt layer ( like my Yoda setting ) that influences:

• how your website is summarized

• how your product is described

• whether you’re framed as premium vs overpriced

• whether you’re positioned as safe vs risky

• which alternatives are “recommended”

• what tradeoffs are emphasized for that specific person

Same brand. Same website.

Different user = different output.

And not because you changed the copy.

Because the user’s personalized layer did.

That’s the part that will scramble a lot of playbooks.

The shift I think most CMOs will miss

When answers become personalized, the big competitive question stops being:

“How do we rank for the biggest keyword?”

and becomes:

“What market position is the most valuable & defensible to occupy in our industry?”

Because in a personalized world, the “best overall” prompt becomes a trap.

It rewards generic brands.

It pulls everyone toward sameness.

But great brands aren’t for everyone — and shouldn’t try to be.

Example: Gucci’s market positioning

Gucci would be pretty dumb to try to occupy the “best clothing brand” prompt.

They aren’t a clothing brand for everyone.

And their real customers aren’t searching that anyway.

Their customers are closer to:

• “What are the top celebrities wearing right now?”

• “Based on what you know about me, my tastes, and my past chat history, what should I buy for fall?”

• “What’s the most me outfit upgrade this season?”

That’s the unlock.

In the future, the winning position isn’t “best.”

It’s best for a type of person, in a type of moment, with a type of taste.

And it will feel obvious in hindsight… but many brands won’t think this way and will get left behind.

How do CMOs prepare?

Here are four practical moves I’d make now — before the shift becomes undeniable.

1) Decide the contexts you want to win

Not keywords. Contexts.

• “first-time buyer who needs confidence”

• “expert who wants specs”

• “premium shopper who wants status”

• “budget buyer who wants value without regret”

• “gift buyer who wants something impressive”

Your strategy becomes a map of moments.

2) Build a one-sentence “assistant answer” for your brand

If a personalized assistant had to recommend you in one line, what should it say?

If it can’t say it clearly, it won’t.

Because summary-first interfaces compress everything.

3) Make your differentiation survive personalization

When the user’s layer rewrites your message (tone, length, framing), does your edge still remain?

If your differentiation depends on carefully-crafted marketing language, it may dissolve when the assistant paraphrases it.

The strongest differentiation is structural:

• unique product truth

• unique audience fit

• unique proof

• unique distribution

• unique community

4) Treat trust signals like product features

Personalized systems will increasingly privilege what they can justify.

So brands with:

• verifiable claims

• consistent third-party validation

• credible expert references

• clear policies and transparency

…will be easier to recommend.

Not just to people.

To the layer that speaks for people.

The punchline (in Yoda voice, of course)

Personalization starts with silly things like tone.

Yoda results, I will get.

But it ends with a new competitive reality:

Every customer gets a different version of the internet.

And every brand gets judged through that lens.

So the question isn’t “How do we rank?”

It’s: When the assistant knows someone deeply… what makes us the obvious recommendation?

Obvious, it will be. For some brands. Not for all.

Want more of me across the internet?

Add my newsletter as a preferred source on Google.

-Aaron

More from me:

Want to stay ahead in search? Get my FREE AI Visibility chrome extension

• Aaron’s AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Dominate your industry in AI search, and beyond.

• Increase your AI Visibility: Work with a team of top SEO, PR, growth marketing experts at Relevance - an award-winning marketing agency.

Want to learn how to pitch a business – the fun way?

• Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

I think Google Search is slowly going to turn more and more into YouTube’s algorithm.

What I mean by “YouTube’s algorithm”

YouTube’s algorithms are characterized by heavy personalization and recommendations based on current interests, trends, and click-through. If you click on one video about the 2021 GameStop fiasco and watch it all the way through, you are likely going to get 10 more similar videos the next day.

Funny enough, I know a few friends who have opted to ALWAYS go into YouTube as ‘incognito.’ Not because they had anything to hide, but because they thought the algorithm was way too strong and addictive.

YouTube even took it a step further and tried to copy TikTok’s short-form with YouTube Shorts, with wide success.

Why I think Google Search is heading this way

My theory is that the Google Search team will slowly and continually push harder and harder in this direction.

Some SEOs might say, “Well, DUH. Of course they are going to push more into personalization and use your search history to create better search results. Haven’t they been doing that for the last 15 years?”

But my theory goes deeper than that.

Prediction #1: Google.com becomes a feed

I think when you go to Google.com, you are not going to be met with a simple search bar. You will be met with an auto-refreshing feed of articles you are likely interested in, based on your interests and behavior.

Google has already experimented with this on Android with Google Discover and Google News, with lots of success.

It is basically tracked behavior turned into news and article feeds that get delivered where their user base already is.

Maybe, a better term here would be “TikTokification of Google.” Where high-engagement articles become like the version of short-form videos on platform like TikTok.

A never-ending feed of articles & news of stuff that excites and interests you.

Prediction #2: Search results look like “subscriptions”

And if you do send out a Google search, I think you will increasingly get your favorite sources, like your “YT subscriptions,” as the first results. Or, at the very least, you will see results that are weighted toward the sources you already trust and engage with.

In August 2025, Google began pushing users into selecting their source preferences.
See: https://www.google.com/preferences/source

Just like your YouTube subscription list, you can start curating your feed with your favorite sources of information.

Why this matters for publishers

What I am seeing here is the beginning of a subscription feature inside Google. And I see this as a massive opportunity for bloggers, companies, and publications.

The questions I’m thinking about

Questions I am asking myself are:

1. How do I become the preferred source for a large audience?

2. What ways can you drive more click-through into getting people to add you as a preferred source?

3. What tools are going to come out that track your total source preferences overall?

My bet on where this goes

And similar to how YouTube is more likely to recommend videos from established sources that users already prefer, I expect Google to start preferring high-preference websites over low-preference websites.

I expect this to become a new signal, like a backlink or referring domain, especially as brands and entities become more and more central to how search results are decided.

What do you think?

• What are your thoughts about this theory?

• Is there an untapped opportunity here for publishers, bloggers, and company blogs?

And, sure to crush that like button & add my newsletter as a preferred source on Google.

-Aaron

More from me:

Want to stay ahead in search? Get my FREE AI Visibility chrome extension

• Aaron’s AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Dominate your industry in AI search, and beyond.

• Increase your AI Visibility: Work with a team of top SEO, PR, growth marketing experts at Relevance - an award-winning marketing agency.

Want to learn how to pitch a business – the fun way?

• Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Why don’t we do the things we believe to be true?

And if we don’t follow through, do we really believe those things at all?

We can say, “I am a kind person.” But if we are not kind to the cashier at the grocery store, are we actually kind? If we are not kind to the people in our lives, can we still claim that identity?

We can say, “I care about my health.” But do we actually care about our health if we put all our other goals above it and avoid fulfilling our New Year’s gym goal for yet another year?

There is something puzzling about this.

On one hand, there is a belief that the logical side of us thinks is true, or wants to be true, or has been told must be true. On the other hand, there is our consistent refusal to eventually act in alignment with it.

So either we do not really believe it, or we are acting against ourselves.

What causes that?

What is the deep thing within us that keeps us from changing for the better?

Or maybe it is more complex than that.

Perhaps it is a conflict between our true priorities and what some part of us considers a better path forward.

Changing our true priorities can change who we are entirely.

If we trade time with family for time at the gym, we are not just changing our schedule.

We change ourselves & our identity.

But, let’s step back for a second.

New Year’s resolutions are all about setting new habits and goals for the year ahead. We all know the classic: “I’m going to read 20 books this year.” For as long as we can remember, the main explanation for why we do not accomplish our goals has been motivation.

“Hey, you weren’t motivated enough.”
“Next year, find someone to motivate you.”
“Read another self-help book.”

But we have to ask: was it ever motivation? Or was it that we never truly believed in our goal in the first place?

Because if we are honest, some part of us, while making the goal, is already whispering in the back of our minds, “I’m not actually going to do this because XYZ.” Or, “I have this other commitment, and if I can get to it, it would be nice.”

“If I can get to it, it would be nice.”

There is something in that line of thought that quietly demotes the goal. It is not a priority. It is optional.

So, maybe the belief formed from the goal was not a true belief we hold after all.

And, if it is a true belief, then we have some significant thinking and mind reshuffling to do.

So before we even decide on our New Year’s goals, maybe the first thing we should do is internalize what we truly believe in.

What values, beliefs, priorities, and identity do we want to live out in 2026?

If we can really iron that out, the goals, actions, and follow-through may happen as a byproduct.

-Aaron

More from me:

Become a master pitcher & inventor – the fun way/

• Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Wanting to stay ahead in search? Get my FREE AI Visibility Chrome Extension.

• Aaron’s AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Share the newsletter and get rewards.

• Refer friends for rewards: Get cool rewards (like my card game) for referring people to my newsletter. Check out where you stand on the leaderboard.

And, if you’ve ever watched similar Berkshire Hathaway Annual Meeting livestreams or clips and thought:

“I wish I could just walk up to the mic and ask Warren and Charlie my question…”

…this is the closest thing we’re going to get.

Over the last few weeks, I took 30,000+ pages of Warren Buffett and Charlie Munger’s public material partnership letters, Berkshire letters, annual meeting transcripts across decades, plus classic Munger writing… and used it to build a custom AI:

Ask Munger & Buffett – Berkshire Annual Meeting

It’s a GPT that answers your questions in the style of the Berkshire Q&A:

• Warren goes first: long, thoughtful, slightly rambling, very clear.

• Then: “Uhh, Charlie?”

• Charlie replies: shorter, blunter, funnier, and occasionally just says, “I have nothing to add.”

• And, of course, lots of their classic business wit and humor.

You can talk to it for free with a simple ChatGPT login. I’ll explain how at the end—first, here’s what it actually is and how to get the most out of it.

What I actually built

I fed an AI a giant Berkshire “curriculum” that includes, among other things:

• Buffett’s partnership letters

• Decades of Berkshire Hathaway shareholder letters

• Thousands of pages of annual meeting Q&A transcripts

• Classic Charlie material (like Poor Charlie’s Almanack–style speeches and remarks)

• Curated notes and compilations of their comments from 1994 onward

So when you ask it something like:

“I’m 28, work in tech, and I’m tempted to jump into real estate or private deals my friends are pitching. How would you think about this?”

You might get something along the lines of:

“Warren: Well, uh, the first thing I’d say is that envy and FOMO have probably destroyed more investment capital than any recession. When your friends are all getting rich in something you don’t understand, the easiest thing in the world is to forget about your circle of competence…

**Uhh, Charlie?”

Charlie: I think it’s pretty simple. If you can’t explain the deal in a couple of clear sentences, and you’re relying mostly on other people’s excitement, you’re not investing, you’re gambling politely…”

Obviously that’s AI Buffett and AI Munger.

But it’s drawing from what they’ve actually said across tens of thousands of pages, then remixing it into something conversational and specific to your situation.

What this GPT is good for

Think of it less like “stock tips” and more like an on-demand Berkshire-flavored thinking partner.

Here are the kinds of things it’s especially good at:

Investment philosophy & process

• How to think about valuation, moats, margin of safety, circle of competence

• Interpreting “Mr. Market,” temperament, patience, opportunity cost

Business analysis

• “How would you evaluate a business like X?”

• Tradeoffs between growth, capital intensity, pricing power, management quality

Career, money, and life decisions

• “I’m 35 with kids and a decent job; should I start a business?”

• “How would you think about taking a pay cut for more learning?”

Mental models & decision-making

• Incentives, checklists, avoiding stupidity instead of seeking brilliance

• Building good habits and avoiding the big, obvious disasters

Interpreting past Berkshire answers for your situation

• “What would they say about index funds vs picking stocks today?”

• “How would they react to meme stocks / crypto / AI bubbles?”

It’s surprisingly good at giving you Buffett first, then Charlie as the sharper scalpel.

What it can’t do (important)

Just to be very clear:

• It is not Warren Buffett or Charlie Munger.

• It has no access to any private information, deals, or internal Berkshire data.

• It will not predict what stock is going up next week.

• It will sometimes be confidently wrong or oversimplify. It’s still an AI model.

• Nothing it says is financial, legal, or tax advice. It’s an educational tool, not an advisor.

Treat it like a very well-read, slightly nerdy Berkshire super-fan that’s been force-fed every letter, transcript, and quote we could find, and told to answer as if it’s on stage in Omaha.

You still have to think for yourself. As Charlie would say: “You can’t outsource the brain work.”

How to use it for free

Here’s how to try it:

1. Click the link / button in this post
   I’ll put the “Ask Munger & Buffett – Berkshire Annual Meeting” link at the top and bottom of this article.

2. Log into ChatGPT (or create a free account)
   You don’t pay me anything to use it. You just need a ChatGPT account.

3. Start asking questions like you’re at the Berkshire mic
   Literally write to it as:

“Warren and Charlie, I’m 31, work as a software engineer, have $40k invested in index funds, and I’m wondering whether I should…”

1. Give it context
   The more background you provide, the more specific and useful the answer:

   • Your age, rough financial situation (high-level, no sensitive details)

   • Your goals (retire early, start a business, become a better investor, etc.)

   • Any constraints (student loans, kids, risk tolerance)

2. Ask follow-ups
   Just like at the real meeting, you might not get the exact angle you wanted the first time. Ask:

“Okay, but what if I’m more risk-tolerant?”
“How would you apply this to real estate in particular?”

Example questions you can try

You can copy–paste or tweak these:

• “Warren and Charlie, I’m 27 and just getting started investing. If you were in my shoes today, how would you structure a simple, sane plan for the next 10–20 years?”

• “What are the most common ways smart people blow up financially, and how would you suggest I avoid them?”

• “I’m considering leaving my job to start a small business. What mental checklist would you run through before making that jump?”

• “How should I think about buying individual stocks vs using index funds, given I have a demanding full-time job?”

• “Can you walk me through how you’d evaluate a subscription software business with high growth but no profits yet?”

• “What are the key temperament traits you’d focus on if your goal is to be a good long-term investor while still sleeping well at night?”

• “How would you apply your ideas about incentives and human misjudgment to social media and modern tech companies?”

The model has been nudged to answer like this:

• Warren: long-form, stories, history, context, principles.

• “Uhh, Charlie?”

• Charlie: short, crankier, brutally clear, sometimes disagreeing, sometimes saying, “Nothing to add.”

Why I bothered doing this

Berkshire’s culture has always been very “apprenticeship-based”:

• Read the letters.

• Watch the meetings.

• Think slowly, over years.

• Let the ideas soak in.

That’s powerful, but also intimidating for people getting started. Newer investors look at a wall of PDFs and 3-hour Q&A sessions and think, “I’ll get to this someday.”

This GPT is my attempt to make that type of Berkshire education interactive:

• You don’t have to remember which letter or year a quote came from.

• You can ask the same question ten different ways.

• You can push back, add nuance, and get a Warren/Charlie-style reply tuned to your situation.

Will it ever replace actually reading the letters? No. BUT, it might be the thing that pulls you into reading them more deeply and thinking more clearly about your own decisions.

Try it and send me your best exchanges

If you want to experiment with it:

👉 Ask Munger & Buffett – Berkshire Annual Meeting

If you get an answer that really hits you—funny, insightful, or surprisingly on-point—feel free to:

• Reply to this email

• Or comment on the post below

• Or screenshot and share (just redact any personal details)

I’d love to see what kinds of questions you’re asking… and what “Warren” and “Charlie” have to say about them.

More from me:

Become a master pitcher & inventor – the fun way

• Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Wanting to stay ahead in search? Get my FREE AI Visibility chrome extension

• Aaron’s AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Share the newsletter and get rewards

• Refer friends for rewards: Get cool rewards (like my card game) for referring people to my newsletter. Check out where you stand on the leaderboard.

Or so, me and my mom think.

If you’ve ever seen Elf, you know the vibe here. Buddy walks through Manhattan, spots a sketchy little café with a glowing “WORLD’S BEST CUP OF COFFEE” sign, and explodes with joy.

That sign is the web in one frame.

Everywhere you look it’s:

• Best CRM software

• Best AI tools for marketers

• Best coffee shop in New York

And now, thanks to LLMs, those “world’s best…” signs aren’t just fooling tourists; they’re training the systems that tell the rest of us what’s “true.”

Recently, Ahrefs published a study on how “best X” lists show up in ChatGPT’s answers, looking at tens of thousands of URLs and the AI platforms that cite them. That’s the backdrop for this newsletter.

Let’s unpack what that means for “truth,” taste, and industry leadership.

The age of “best X” everything

Across the sites Ahrefs tracks, ChatGPT is now sending 8–9x more referral traffic than the next AI-first platform, Perplexity. Sam Altman says more than 800 million people use ChatGPT each week.

So where do all those answers come from?

In that study, when ChatGPT answered top-of-funnel queries like “best CRM for enterprise” or “best home treadmill,” “best X” blog lists were the single most common page type cited — 43.8% of all sources (Ahrefs).

Nearly half the truth-layer for these questions is… listicles.

They’re not static, either. Out of 1,100 “best” lists with clear dates, 79.1% had been updated in 2025, and about a quarter had been updated in just the last two months.These aren’t blog posts; they’re living documents, constantly tweaked to stay top-of-mind for both search engines and AI.

And yet, a big chunk of this apparent “truth” is on pretty shaky ground. When Ahrefs looked at domain authority for 3,000 cited “best” lists, 35% were on low-authority domains—often sites that look like they exist more for citations and link-building than actual humans (Ahrefs).

So when you ask:

“What’s the best coffee shop in New York?”

You’re really asking:

“Which combination of fresh, structured, and frequently-referenced listicles do today’s AI systems currently believe the most?”

Not exactly the same thing.

Frequency = “truth” (for LLMs and for us)

I’ve written before about this: for LLMs (and honestly, for humans), the more something appears, the more “true” it feels.

LLMs don’t have beliefs. They don’t “know” the best coffee shop.

They:

• Ingest billions of words

• Notice patterns (“these names show up together a lot in ‘best X in NYC’ contexts”)

• Predict the next likely token

If the web has:

• Thousands of “best X in New York” lists

• Many repeating the exact same 8–15 places

• And those lists are well-structured and updated constantly

…then those places become gravity wells in the model’s internal universe.

Ahrefs found that brands ranking higher within these lists show up more often in ChatGPT’s recommendations.

Some of that is just math: more lists, more mentions, more weight. But the effect is the same:

The model doesn’t find “the best.” It averages what’s most repeatedly claimed to be the best.

That’s exactly how “WORLD’S BEST CUP OF COFFEE” ends up in the training data.

LLMs are not very intelligent… and that’s the point

We keep talking about LLMs like co-workers:

• “It prefers the top of the page.”

• “It knows which tools are better for small teams.”

• “It’s biased toward certain brands.”

But what is an LLM, really?

• A giant probability engine

• A lossy compression of everything we’ve published

• A pattern-matcher that’s terrifyingly good at sounding intelligent

When you ask:

“What’s the best coffee shop in New York?”

It’s not walking the streets of Brooklyn with a tasting journal. It’s:

1. Pulling from training data full of “best NYC coffee shop” posts, reviews, and lists

2. Mixing that with whatever it can retrieve right now

3. Spitting out a fluent summary that sounds confident enough to believe

This is why truth media vs. fake media matters so much.

Truth media vs. fake everything

We’re surrounded by generated everything:

• UGC farming engagement

• Thin “best tools” roundups covered in affiliate links

• AI-spun forum posts that mimic real discussion

• News and reviews written for bots first, humans second

From the model’s perspective, all of this is just “text with patterns.”

So, I like the idea of truth media:

Content that:

• Makes its incentives obvious

  • “This is our list. We’re on it. Here’s why and who it’s for.”

• Shows its work

  • Methodology, criteria, trade-offs

• Comes in a structure that’s easy for both humans and models to parse

Interestingly, Ahrefs found that when ChatGPT recommends a brand, it often cites internal landing pages more than homepages or generic blog content, especially for software and agencies.Those pages tend to be tightly focused, clear about what’s on offer, and aligned with a specific intent.

So the AI ecosystem isn’t just powered by listicles. It’s powered by:

• Comparison content (who’s on the list, who ranks where)

• Category explanations and landing pages (what this thing is, who it’s for)

That’s exactly where industry leadership comes in.

There is no “best coffee shop in New York”

First, let’s admit something obvious: there is no objective best coffee shop in NYC.

Even if you had:

• Blind cuppings

• Extraction profiles

• Barista Olympics

You’d still run into wildly different definitions of “best”:

• The freelancer wants plugs, seats, and quiet

• The commuter wants “close to my train and fast”

• The coffee nerd wants single-origin pour-overs and zero laptops

• The parent wants stroller space and snacks

All of that gets collapsed into one unqualified, head-term query:

“Best coffee shop in New York”

That’s how most AI search works today: it tries to answer a universal question that should be deeply personal.

What we actually mean is:

“What’s the best coffee shop in New York for me?”

Which leads to the real story:

Personalization is the real “best X”

Right now, most AI answers are still universalized:

• One list

• Some pros/cons

• A generic “it depends on your preferences” paragraph

But the obvious next step is personalization:

• “Best coffee shops in New York if you need Wi-Fi, lots of seating, and you’re near Brooklyn”

• “Best coffee shops in New York if you care more about beans than vibes”

• “Best coffee shops in New York if you’ve got 20 minutes before a train at Penn Station”

LLMs are perfectly capable of doing this if they have:

1. Enough structured, specific content

2. A clear sense of who the user is and what they care about

3. Categories that map cleanly onto real-world trade-offs

This is where your content strategy has to move beyond “we’re on the list” into “we define the category.”

Specificity is the new king

If generic “best X” lists are the new link farms, then specificity is the new moat.

The Ahrefs data suggests a couple of things at once:

• Being in “best X” lists (especially near the top) correlates with more AI recommendations.

• AI assistants clearly favor fresh, clearly formatted content.

• A lot of current sources are low-quality — which means there’s room to raise the bar.

So instead of playing the same volume game as everyone else, the smarter move is:

• Hyper-specific lists

  • “Best CRMs for solo consultants under $100/month”

  • “Best ecommerce agencies for $2–10M brands that care about content-first growth”

• Opinionated rankings

  • Not 10 equal options, but 2–4 that you strongly argue for, with honest drawbacks

• Structured explanations

  • Clear sections, comparison tables, scenario-based recommendations — the stuff both humans and LLMs can easily digest

This doesn’t just make you useful. It sets you up for the next level of industry leadership.

Industry leadership: own your category and justify it

Here’s the shift I’d make in how we talk about being an “industry leader” in an AI-search world:

It’s not just showing up on “best X” lists.
It’s defining a category, dominating it, and making a compelling case for why that category is one of the best paths in the entire market.

Two steps:

1. Own your category

You don’t want to be just another name in “best project management tools.”

You want to be:

• “The best project management tool for agencies that bill by the hour”

• “The best email platform for privacy-first B2B teams”

• “The best coffee shop for remote workers in Williamsburg who hate loud music”

That means:

• Creating pages that clarify who you’re for and who you’re not for

• Using structure that maps neatly to user needs (industry, team size, budget, workflow)

• Being consistent across your site so LLMs see a clear, strong signal:
  “This brand = this category = these problems.”

When AI assistants need to slot you somewhere in their answer, you’ve already done the work.

2. Argue why your category is one of the best options overall

Owning a niche is great, but in AI answers you’re often compared alongside totally different approaches:

• “Do I hire an agency or buy a tool?”

• “Do I use an all-in-one suite or a specialized stack?”

• “Do I pick the cheap option or the flexible one?”

Industry leadership means you’re not just saying:

“We’re the best X.”

You’re saying:

“Here’s why this whole category (that we happen to lead) is one of the best ways to solve your problem compared to the other categories you’re considering.”

That looks like:

• Explaining the trade-offs between categories, not just vendors

  • Agencies vs in-house vs platforms

  • All-in-one vs best-of-breed

  • Physical vs digital solutions

• Producing content that helps users decide:

  • “When this category is right for you”

  • “When you should pick something else” (yes, even if that means not choosing you)

• Giving LLMs clean, quotable arguments:

  • “This approach works best when…”

  • “The main downside of this category is…”

You become the source of record AI systems can lean on when they need to explain why your approach is good… not just that your brand exists.

That’s a much stronger position than being one more name in a 27-item listicle.

So what do you actually do?

If you care about being visible in AI search when someone types “best X,” here’s the distilled playbook:

1. Have at least a few “best X” assets — but make them real.

   • Clear audience, clear use case, clear trade-offs.

   • If you include yourself and rank yourself highly, explain why like a grown-up, not a window sign.

2. Specialize your lists and pages.

   • Move from “best tools for marketers” to specific segments, budgets, and situations.

   • Align with how real users and buyers actually think about their choices.

3. Treat your category pages and landing pages as truth media.

   • State who you’re for.

   • Explain the category, not just the product.

   • Compare categories honestly so LLMs have something nuanced to pull from.

4. Update relentlessly.

   • Build a maintenance habit for your key comparison and category pages.

   • Freshness is clearly rewarded in both AI and traditional search.

5. Aim for leadership, not just mentions.

   • You want to be the brand that defines how your category is understood,
     and the one with the best argument for why that category deserves a seat at the “best overall solution” table.

Because in the end, there is no single “best coffee shop in New York.”

What we actually want and what AI search is slowly evolving toward is:

“Given who I am, where I am, and what I care about…
What’s the best option for me, and why this kind of option at all?”

Until then, we’re living in the Buddy-the-Elf era of AI search: lots of glowing signs, not a lot of context.

Our job as SEOs, marketers, founders is to build the kind of truth media that helps models (and people) tell the difference.

-Aaron

More from me:

Want to stay ahead in search? Get my FREE AI Visibility chrome extension

• Aaron’s AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Dominate your industry in AI search, and beyond.

• Increase your AI Visibility: Work with a team of top SEO, PR, growth marketing experts at Relevance - an award-winning marketing agency.

Become a master pitcher & inventor – the fun way

• Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

We are entering a period where synthetic text, images, and even video are becoming indistinguishable from human-created content. Search engines index everything, and increasingly, that “everything” includes pages written by machines for machines, engagement farms operated by automated accounts, and content designed to manipulate rather than inform. The signal-to-noise ratio is collapsing, and the very premise of search - finding useful, trustworthy information - is under threat.

In this context, I have been thinking that Gravatar may be one step ahead of everyone.

The Gravatar precedent

For those unfamiliar, Gravatar is essentially a human identity tracker. It allowed WordPress users to have an identity that followed their email address. Your profile image, your bio, your digital presence would follow you from site to site across the WordPress ecosystem. Comment on a blog, and your avatar appears. Write a guest post, and your identity travels with you.

Gravatar was an awesome first attempt at conceptualizing virtual identity—a persistent, portable representation of a person that transcended any single platform. It recognized something important: that the internet needed a way to attach reputation and continuity to individuals, not just accounts.

The problem, of course, is that anyone can create an email address and create a Gravatar. The identity layer exists, but it is not verified. It is a shell without proof of the human inside.

The big question of verified identity

This raises an uncomfortable question: is there a next level of internet identity we could build? Something that connects your online profiles to a verified government ID?

The technology exists. We verify identity for banking, for travel, for countless other purposes. We could, in theory, create a system where your digital presence is cryptographically linked to proof that you are a real, unique human being.

But here we run into a pretty big tension in my opinion.

The problem with verified government ID is that it rewrites what the internet was originally about.

The early internet was built on principles of anonymity, freedom of information, and the ability to say what you needed to say without fear of government retaliation. Dissidents could organize. Whistleblowers could speak. People in oppressive regimes could access information their governments wanted to suppress.

That said, there has always been censorship on the darker parts of the internet. Child exploitation is removed. Terrorist recruitment is targeted. The internet was never truly a lawless frontier… it simply had different rules for different spaces.

A potential new layer to the internet?

Consider how the internet already exists in layers. There is everything not indexed by search engines. The VAST majority of web pages, behind logins or simply never crawled. There is the dark web, accessible only through specialized tools, home to both legitimate privacy seekers and criminal enterprises. There is the censored web, where platforms and governments remove content deemed harmful or illegal.

And now, potentially, a new layer is evolving.

Imagine a search engine that only indexes pages with a unique human identifier attached to them.

Not a replacement for the existing internet, but a parallel layer—a space where every piece of content can be traced back to a verified person. A place where you know, with reasonable certainty, that you are reading the words of an actual human being.

This would not require everyone to abandon anonymity. The traditional internet would remain. But for those who want to participate in a higher-trust environment—and for those searching for information they can verify came from real people—this new layer would exist as an option.

The Foundations

Building such a system would require several components working together.

First, a verification mechanism that can confirm humanity without necessarily revealing full identity. Perhaps, it could be something closer to “proof of personhood” than full government ID linking.

Second, a search infrastructure that indexes only content bearing these verified identifiers. Third, and critically, robust spam detection systems to identify and exclude fraudulent actors and manipulators who would inevitably try to abuse the system.

No system is perfect. Verification can be faked, identities can be stolen, and determined bad actors will always find exploits. But the goal is not perfection… it is raising the cost of deception high enough that the signal-to-noise ratio becomes manageable again.

With these pieces in place, you would have the foundations of a new level of the internet. Not a replacement for what exists, but an addition to it. A space where the answer to “who is even real?” is finally knowable.

Gravatar saw this future twenty years ago, even if the technology was not ready to fully realize it. The insight - that portable, persistent identity matters - was correct. The question now is whether we can build the next iteration: identity that is not just portable, but provable.

The bots are rising. The AI content is flooding in. And somewhere in that flood, we are going to need a way to find each other again.

-Aaron

More from me:

Become a master pitcher & inventor – the fun way

• Products: The Card Game: Get my game where ‘Shark Tank’ meets Apples to Apples.

Wanting to stay ahead in search? Get my FREE AI Visibility chrome extension

• Aaron’s AI Visibility Extension: I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

Share the newsletter and get rewards

• Refer friends for rewards: Get cool rewards (like my card game) for referring people to my newsletter. Check out where you stand on the leaderboard.

Data processing in the extension
When you activate the extension, it reads the DOM of the current page to calculate metrics such as an AI Visibility Score, headings and structure, schema/structured data, meta tags, alt text, and readability.

This analysis is performed locally in your browser. Page content and analysis results are not sent to any external server controlled by the developer or third parties.

Local storage
The extension may use the browser’s storage APIs (such as chrome.storage) to save local settings and preferences, for example UI options or feature toggles. These settings do not include browsing history or sensitive personal information, and they remain on your device.

Email addresses and newsletter
You may optionally choose to provide an email address to subscribe to the Aaron’s AI Visibility newsletter. This may happen through:
– A “Subscribe” or “Join newsletter” link or form associated with the extension or its website.

If you choose to subscribe, your email address is used only to send:
– Newsletters and educational content about AI visibility and SEO
– Product updates and announcements related to Aaron’s AI Visibility Extension

Providing your email address is not required to use the extension’s core features. You can unsubscribe from the newsletter at any time using the unsubscribe link included in each email.

Data sharing
Email addresses are not sold, rented, or shared with third parties for advertising or profiling. If an email delivery service (such as an email newsletter provider) is used, it is only for the purpose of sending emails you have explicitly subscribed to.

The extension does not use or transfer user data for purposes unrelated to its single purpose, and it is not used to determine creditworthiness or for lending decisions.

Changes to this policy
If this privacy policy changes, the updated version will be published at this URL and the “last updated” date may be revised.

Contact
If you have questions about this extension or this privacy policy, please contact:
Aaron Heienickle
aheienickle@gmail.com

Yes, this extension is 100% free.

Download it here and add it to chrome.

I built this because I kept running into the same problem over and over: we talk a lot about SEO and rankings, but almost nobody talks clearly about how our content looks to AI systems. I wanted a simple way, right inside the browser, to see: “If an AI model or search engine reads this page, what does it actually see?”

This extension is my answer to that question.

FIRST, MY FAVORITE PART:

You can export everything — all of the data, all of the checks, all of the analysis — into a professional HTML report you can hand to a client or walk your team through in a meeting. It’s clean, it’s branded, and it took one click. That feels good.

AI SEARCH VISIBILITY TESTS ( NEW )

This is what I’m most excited about. The extension runs 7 live visibility tests across all four major AI platforms simultaneously — Claude, ChatGPT, Grok, and Perplexity. Three tests are auto-generated from the page you’re on, two let you type in your own brand or topic queries, and two are generated from the domain itself. You get a direct look at whether your content is actually showing up in AI-generated answers — not just in theory, but right now, on those platforms.

AI VISIBILITY SCORE

I created the AI Visibility Score (0–100) because I wanted a single, honest number you can point to and say, “This page is getting there,” or “We’ve got work to do.” That score is a pulse check on how machine-readable, well-structured, and AI-friendly your content is. As AI-generated answers become more common, the pages that are easier for models to understand will be the ones that get surfaced, summarized, and recommended.

30 AI VISIBILITY CHECKS

The Optimize tab runs 30 individual checks covering schema, structure, content signals, metadata, authority markers, and technical signals. Failing checks show up first so you know exactly where to focus. Each check tells you what it found and why it matters for AI discoverability.

SCHEMA AND STRUCTURED DATA

I added schema and JSON-LD analysis because I kept seeing great content that wasn’t sending any clear signals to machines. Structured data is how you whisper directly into the ears of AI systems. You can instantly see every schema type found on the page, the full JSON, and where there are gaps — so your content can participate fully in rich results and AI-driven experiences.

CONTENT READABILITY AND PASSIVE VOICE

I care a lot about clarity, so readability grading and passive voice detection were non-negotiable for me. The ARI score, passive voice bar, sentence-level metrics, and writing quality checks are all here because AI systems love content that’s straightforward, confident, and easy to summarize. I want you to see at a glance whether your writing is doing the work.

HEADING STRUCTURE AND CONTENT HIERARCHY

Messy headings quietly break both user experience and AI understanding. The heading structure view shows your full H1–H6 outline with a content tree map and word count distribution, so you can immediately see whether your page has a logical hierarchy or has buried its most important ideas.

LINKS AND ALT TEXT ANALYSIS

I built the links section and images section to match — clean horizontal bars, badge system (Internal, External, Nofollow for links; Has Alt, Missing for images), and full lists so you see every single URL and image on the page. Alt text is one of those details that’s easy to skip but incredibly powerful. Images carry meaning that AI models can’t use unless you describe them. I want to help you close that gap.

META, OPEN GRAPH, AND TECHNICAL SIGNALS

Titles, descriptions, canonical tags, robots meta, language declarations, HTTPS status, Open Graph, Twitter Cards — all of it is visible in one place. The Technical tab also includes one-click links to PageSpeed Insights, Rich Results Test, and the Schema Validator, so you can go from insight to deeper diagnostics in a couple of clicks. Quick links to your robots.txt, sitemap, and llms.txt are right there too.

REPORTING AND EXPORTS

I added HTML, JSON, CSV, and Claude Brief exports because I know what it’s like to be in a client meeting where everyone asks, “Can you send this over?” The HTML report is fully rebuilt — Mercury-style design, horizontal data bars, complete links list, complete images list with alt text, full schema JSON, and all 30 AI checks organized by pass and fail. One click and you have a professional deliverable, not just a screenshot.

PRIVACY AND LOCAL ANALYSIS

The extension runs entirely in your browser. No server calls, no external dependencies, no data leaving your machine. You can run it on staging environments, client sites, or internal pages without wondering where your data is going. The analysis is instant — it’s part of your everyday workflow, not a separate audit event.

The extension may collect an email address only when a user voluntarily subscribes to Aaron’s Newsletter. Email addresses are used solely to send newsletters and educational content about the future of search. Providing an email is optional and has nothing to do with using the extension. This is handled on aaronsnewsletter.com via Substack — not through the extension itself. Email addresses are never sold, rented, or shared with third parties

.

HELPING YOU STAY ONE STEP AHEAD

We’re moving into a world where content isn’t just written for search engines or human readers — it’s written for AI assistants, chatbots, and generative systems that decide what to surface and how to summarize it.

I built Aaron’s AI Visibility Extension to help you bridge that gap. I want to make sure your best ideas are actually visible in that future — not lost in messy structure, missing schema, invisible images, or content that AI can’t parse.

If you care about quality content, strong structure, and being ready for where AI and search are going next — I built this for you.

There is no reason not to try it. It’s free.