GEO and the Token Tax: How to Stay Visible to AI Assistants

Ben cohen

Picture a buyer asking an assistant for the best option in your category. It gives a shortlist of three, compares features, mentions reputations, and the buyer picks one. Your website never got a visit. Did you win or lose that deal? Your analytics can't tell you.

Ido

And that's the whole episode, really. Welcome to Be Visible. I'm Ido, and this is Ben. Today the question is how a brand stays visible when the thing standing between the customer and the web is an AI assistant.

Ben cohen

Start with what actually changed. In classic search you held a position, and everyone could see it. Now ChatGPT, Claude, Gemini and Perplexity build a custom answer on the fly, so the same question can give two different answers.

Ido

Right, it's probabilistic. So a fixed rank stops being the number that matters. You look at brand mentions, citations, share of voice inside answers, and sentiment. One of the tool reviews we read put it nicely: rank isn't really something to look at anymore, it's mentions and citations.

Ben cohen

But is this really different from SEO? Because Google's public line is that optimizing for AI doesn't need fundamental changes.

Ido

Both are true. The foundations carry over: crawlable, fast, clear, trustworthy. GEO adds layers. Can the system understand who you are, believe you, and pull a clean answer out of your page? Last time we talked about the three audiences, the human, the machine, and the agent. This is the machine part.

Ben cohen

And the token tax. I built a tool that shows what a site looks like to an AI once you strip the styling and scripts. Some page builder sites turn into a wall of nested divs before you reach one sentence about what the company does.

Ido

Which is where our restaurant analogy comes in. Your website is the restaurant, with the lighting and the furniture. Clean markdown is the menu. The AI doesn't need the decor, it needs to know what's on offer.

Ben cohen

So, the menu. Does adding an llms.txt file suddenly make a site visible?

Ido

No. Jeremy Howard of Answer.AI proposed it in September 2024. It's a markdown file at the root with a site name, a short summary, and curated links. Not every URL like a sitemap does, just what matters. The companion, llms full dot txt, bundles the full cleaned content, so a system can take it in one request.

Ben cohen

What does the evidence say?

Ido

Mixed. An agency called dev5310 submitted theirs to Search Console on February 2, 2026. The next day Google AI Mode cited it as the first source for a brand query. Their rankings didn't move. And they say themselves it's one agency, four queries, an 18 day window. Search Engine Land tracked 10 sites and found 8 of 9 showed no measurable change in AI traffic. Google says it's not a ranking signal.

Ben cohen

So it's useful when there's strong content and schema behind it, and it's not magic.

Ido

Exactly. One detail I loved from their Cloudflare logs: OAI SearchBot and ChatGPT User both hit the file. SearchBot is OpenAI's indexer, building the knowledge base. ChatGPT User is the live one, fetching because a real person asked something right then.

Ben cohen

Which matters, because a live fetch means a customer is in the middle of deciding. But what about cloaking? Mueller at Google and Canel at Bing pushed back in February 2026 on separate markdown pages just for bots.

Ido

Their objection was serving bots something humans never see, and doubling the crawl load. My read, and it's interpretation, is that the line is between a hidden mirror of every page and a public, honest summary anyone can open. That second one isn't settled by any official statement, so be careful.

Ben cohen

Then schema. It's the ID badge. Organization, author, article.

Ido

With one rule: it has to match what's visible. Fix posts published by admin, give real author names, include dateModified. FAQ markup with two or three sentence answers is easy to extract. But schema claiming prices or credentials the page doesn't show is worse than none.

Ben cohen

And then agents. What happens when the visitor is an agent acting for a person?

Ido

Then understanding isn't enough. It needs labeled forms, visible pricing, predictable navigation, clear permissions, maybe an API. A site an AI can read but can't operate is incomplete. And check robots.txt, because some security plugins block GPTBot, ClaudeBot, PerplexityBot, or Google Extended by default.

Ben cohen

So what can someone do this week? Pick five to ten buyer questions without your brand name in them. Ask ChatGPT, Perplexity and Gemini the same ones. Write down how you're described, who gets cited, and which links get credit.

Ido

A spreadsheet works. Tools like Otterly or Scrunch automate it, and LLMagnet is one example, tracking AI bot visits, prompts, and generating llms.txt. Test any of them with the same prompts. Then make your core facts explicit on the page, add accurate schema, and publish original material worth quoting.

Ben cohen

And what not to do: don't keyword stuff for models, don't mass produce thin AI pages, don't paste a sitemap into llms.txt, and don't count citations without checking accuracy.

Ido

My prediction, and it's only an opinion: hosts and CMSs will serve clean machine versions by default, so nobody hand builds them.

Ben cohen

Mine: AI visibility reports will sit next to analytics, and the question will shift from are we mentioned to are we described correctly.

Ido

The future of visibility isn't only getting people to visit. It's making sure machines can discover, understand, trust, represent, and eventually act on your brand correctly.

Ben cohen

Today, ask three AI systems the same customer question and compare. If this helped, follow Be Visible and send it to someone with a website.