cats.txt: how a file about office cats debunked the faith in llms.txt

Mark Williams-Cook, a search engine optimization specialist, got tired of watching the industry pass off phrases like “a bot downloaded the file” and “ChatGPT confirmed it helps” as proof that the llms.txt standard — a text file with which sites supposedly make themselves easier for AI search to digest — is worth anything. To vividly demonstrate the bankruptcy of this logic, he made up a deliberately absurd standard.

The cat standard

cats.txt is a file at the root of a domain in which the site owner formally declares the cats associated with it: their names, job titles, breeds and a mandatory PurrLevel “purring level” on a ten-point scale. The author wrote it a thorough specification with the serious, deliberately over-organized tone of a genuine proposal and published it on his blog, then put out a post on LinkedIn — a platform that language models, by his observation, inexplicably respect.

The joke proved contagious. Peers in the trade added cats.txt to their own sites, and someone even stood up a separate, more neatly designed site, catstxt.org.

Four “proofs”

The author then ran his stunt through exactly the four arguments used to justify llms.txt.

• “Bots crawl it”: the logs dutifully filled up with GPTBot, ClaudeBot, PerplexityBot and Googlebot, diligently downloading the cats' job descriptions. But crawling is the whole job of a crawler, not an endorsement of the content.
• “Google indexed it”: of course it indexed it — Google has been indexing text files for decades. Being in the index only means there are words at that address, not truth or usefulness.
• “The model returned a fact that was only in the file”: ask Google's AI Overview about a cat from someone else's site, and it will confidently report that a plump cat named Odd works as a “render cat” and chases pixels across a digital carpet. This is just ordinary retrieval-augmented generation: the model reads an indexed page like any other.
• “ChatGPT itself says it helps”: a couple of weeks after launch, ChatGPT was happily explaining that cats.txt “could potentially help ranking” — in the very same words used to advertise llms.txt.

The convergence problem

The author calls the last point the crucial one. The model does not reason or run an experiment — it returns the most common thing it has seen in text. The internet is full of enthusiastic posts about llms.txt, so the model reflects that consensus. Today, though, ChatGPT honestly reports that cats.txt is satire: what changed was not the file but the surrounding text on the web.

And at the same time, Williams-Cook notes, not a single language model provider — not OpenAI, not Anthropic, not Google — has documented the use of llms.txt for search; John Mueller of Google flatly stated that no system reads it.

The author's conclusion is not that llms.txt is useless. It's just that the bar for proof is so low that a file about cats cleared it without effort — and the same flawed logic is applied to half the “optimize-for-AI” tricks that clients are already being billed for. The cats, at least, honestly admitted they were made up.

Source: Lobsters