LLMs.txt Has No Proof — Stop Billing Clients for It

TL;DR: A satirical file about fictional cats passed every test used to “prove” llms.txt works — crawled, indexed, cited, and endorsed by ChatGPT. The author’s point: the evidence bar for GEO tactics is so low that a file about a Tuxedo cat named Odd cleared it. Operators spending budget on these rituals deserve harder questions from their agencies.

The Joke That Became a Web Standard

Mark Williams-Cook got fed up watching conference decks treat “an AI bot fetched it” as proof of anything. So he invented cats.txt: a plain-text file placed at your domain root to formally declare the cats associated with your website — names, job titles, breeds, and a mandatory affection metric called PurrLevel scored out of 10. He wrote a proper specification, published it on his blog, and seeded a LinkedIn article introducing cats.txt as “the missing standard for SEO and GEO.” His goal was simple: run the exact same four “proofs” being used to sell llms.txt on something so transparently absurd that no one could keep a straight face. Within days, other SEOs had adopted the standard. Someone even built catstxt.org, a cleaner rival implementation. The experiment had acquired community momentum — which is, ironically, more traction than most real standards achieve in their first two weeks.

This is worth paying attention to if you run a high-spend acquisition program. Whether you’re managing forex acquisition campaigns or scaling player volume in regulated iGaming markets, the budget you allocate to speculative visibility tactics comes directly out of the budget available for things with measurable return. The cats.txt episode is a clean stress test for how your agency evaluates new GEO tactics before invoicing you for them.

The Four “Proofs” — and Why None of Them Hold

Williams-Cook tested cats.txt against each argument routinely presented as evidence for llms.txt.

1. “The LLM bots crawled it.” PerplexityBot, GPTBot, ClaudeBot, and Googlebot all dutifully fetched a file describing his cats’ professional responsibilities. A crawler fetching a file tells you nothing about whether the contents are read, weighted, or acted upon. Fetching things is the entire job of a crawler. The postman touching your gate is not an endorsement of what’s in your bins.

2. “Google indexed it.” Google indexed cats.txt and offered to surface it in Search Console with “indexing and ranking data.” Being in the index means a URL exists and contains words. It is not a verdict on truth or usefulness. Google has been indexing arbitrary text files since before most people selling GEO services owned a smartphone.

3. “ChatGPT returned information only in my file.” This is the strongest-looking argument and deserves the most scrutiny. When a model surfaces a fact from your file, it is using ordinary retrieval-augmented generation — it ran a search, landed on a page that ranked because it was indexed, and read it. The file functioned as a web page, not as a trusted special source. Google’s AI Overview confidently reported that the cat Odd is a “Render Cat,” a Tuxedo with a PurrLevel of 5/7, who “chases the cursor, pounces on stray pixels.” Every word invented, every citation real. The mechanism was not reverence for the file format. It was standard grounding applied to whatever ranked.

4. “ChatGPT itself says it works.” Two weeks after cats.txt launched, asking ChatGPT whether it could help you rank in LLM-driven systems produced an enthusiastic yes — complete with talk of “structured signals for machines” and how AI systems would “trust, summarize, and cite your content more accurately.” This is word-for-word the pitch made for llms.txt, delivered on behalf of a file about how much a Maine Coon named Byte enjoys being stroked. A language model endorsing a tactic is not evidence the tactic works. It is evidence that the surrounding text on the internet says the tactic works, and the model is returning the average of that discourse.

The Convergence Problem Is the Real Issue

Williams-Cook names the underlying mechanism the convergence problem. When you ask a model whether a GEO tactic works, it does not run an experiment. It returns the most common thing written on the subject. The web became thick with confident posts about llms.txt, so the model converged on that consensus and reflected it back as a considered opinion. By the time cats.txt had accumulated enough enthusiastic write-ups, ChatGPT endorsed it on the same basis. When the discourse caught up and admitted the gag, the model reversed its verdict — nothing about the file had changed, only the surrounding text.

This matters because the same convergence dynamic governs the next tactic, and the one after that. Any new GEO ritual that gets enough breathless coverage will receive ChatGPT’s endorsement within weeks. That endorsement is not a signal to act. It is a signal to ask harder questions. The marketing audit process at any serious agency should include a documented mechanism standard before a new tactic touches client budget — not a screenshot of what ChatGPT said when asked.

What Google and Ahrefs Actually Found

Williams-Cook grants llms.txt every benefit of the doubt before making his argument, but the data he parks is worth surfacing here. Google’s John Mueller stated plainly that no AI system currently uses llms.txt, and that consumer LLMs fetch pages for training and grounding but none of them fetch the llms.txt file. Ahrefs ran the numbers across 100,000 domains and found the file is largely ignored by the crawlers it is meant to court — with no measurable citation advantage for sites that add one. The mechanism people are paying to implement does not appear to fire. This is not a fringe finding. It is two of the most credible data sources in the industry reaching the same conclusion independently.

For operators running paid acquisition in competitive verticals — iGaming player acquisition, crypto exchange growth, or law firm intake campaigns — the opportunity cost of chasing undocumented tactics is real. Budget directed at speculative files is budget not directed at conversion rate work, landing page testing, or paid media optimization with documented return curves.

What This Means for Performance Marketing Operators

The cats.txt experiment is not an argument against preparing for AI search. It is an argument against treating preparation as proof of results. If adding llms.txt costs you thirty minutes and zero ongoing budget, the downside is negligible and the optionality is real — on the day a provider documents genuine support, your implementation is already live. Williams-Cook himself says as much. The problem is the invoice attached to that implementation, and the slide deck that uses four observations — crawled, indexed, cited, endorsed — to justify it.

For operators spending $10K or more a month on acquisition, the stakes on bad reasoning are not academic. A GEO tactic invoiced as a “proven lever into AI answers” should produce documented evidence of that mechanism before it touches budget. “The bots crawled it” does not clear that bar. “ChatGPT confirmed it helps” does not clear that bar. Neither does a Google index entry for a file about a British Shorthair named Pixel with a PurrLevel of 8.

The sharper test is whether your agency can point to a provider — OpenAI, Anthropic, Google — that has documented using the file during inference or answer generation. As of the date this article was reported, none have. That is the standard. Until it is met, the tactic belongs in the “low-cost option, unproven return” bucket, not in the “this is how we’re going to improve your AI visibility” pitch.

Operators who want a grounded look at where their current acquisition spend is performing against documented mechanisms should start with a structured review of what is actually moving the needle. Precision targeting across paid channels has documented return curves. Content that ranks and converts has documented return curves. A text file that passes the same four tests as a joke about fictional cats does not.

The cats, at least, were honest about being made up.

Originally reported by Search Engine Journal, August 2026.