The llms.txt evidence most agencies present to clients, that AI bots crawled the file, Google indexed it, an LLM repeated content from it, and ChatGPT confirmed it works, applies equally to a text file declaring fictional office cats, their job titles, and their purr ratings. That is the finding of an experiment documented in Search Engine Journal, and it has direct implications for any brand currently budgeting for GEO tactics.
What Is cats.txt, and How Did It Get “Validated”?
cats.txt is a deliberately absurd web standard invented by a Search Engine Journal contributor who grew frustrated watching the industry treat four observable events as proof that llms.txt improves AI-search performance. To make the problem undeniable rather than arguable, they invented a satirical specification requiring websites to formally declare their office cats: names, breeds, job titles, and a mandatory “PurrLevel” score out of 10. The file was published at a domain root (the top-level folder of a website), seeded with a LinkedIn post promoting it as “the missing standard for SEO and GEO,” and then run through the exact four validation tests the industry uses for llms.txt. It cleared all four.
The experiment acquired its own momentum. A technical SEO named Dave Smart added cats.txt to his own site, making him an early adopter of a standard built to be nonsense. Someone else built catstxt.org, a cleaner, better-organized implementation of the joke specification. Neither of those developments were planned by the original author.
Is the llms.txt Evidence Strong Enough to Act On? The Four Tests, Examined
The cats.txt file passed each of the four proofs most commonly used to support llms.txt investment. The Search Engine Journal article breaks down what each test actually measures, versus what proponents claim it proves:
- AI bots crawled the file. PerplexityBot, GPTBot, ClaudeBot, Googlebot, and others all fetched the cats.txt file. The article notes that a crawler (an automated program that requests web pages) fetching a file tells you nothing about whether its contents are read, weighted, trusted, or acted upon. Fetching files is the default behavior of every crawler. The article describes it plainly: “The postman touching your gate is not an endorsement of the contents of your bins.” catstxt.org even built a live log viewer so anyone can watch AI bots dutifully crawling fictional cat job descriptions in real time.
- Google indexed it. Google indexed cats.txt and offered Search Console data for a file asserting that a British Shorthair named Pixel works as a “GUI Purrfectionist” with a PurrLevel of 8. The article points out that Google has indexed plain-text files since before most llms.txt proponents owned a smartphone. Indexing confirms a URL exists and contains words. It is not a verdict on usefulness or accuracy.
- An LLM returned details that only appeared in the file. When Dave Smart added cats.txt to his site, Google’s AI Overviews began citing it, confidently describing his declared cat Odd as a “Render Cat,” a Tuxedo with a PurrLevel of 5/7, who “chases the cursor, pounces on stray pixels, and stashes them on the digital carpet.” The explanation is ordinary RAG (retrieval-augmented generation, the process where an AI model searches the web and reads whatever ranks before forming an answer). The model read cats.txt as a regular webpage, not as a special trusted source. It ranked because it was indexed, per the previous test. That is the file working as a web page, not as an authoritative standard.
- ChatGPT endorsed it when asked. Roughly two weeks after cats.txt launched, asking ChatGPT whether the file could help with search rankings produced: “Yes, cats.txt can potentially help you rank in both search engines and LLM-driven systems.” The model elaborated about “structured signals for machines,” “better understanding leading to better visibility,” and how AI systems could “trust, summarize, and cite your content more accurately.” That is, word for word, the pitch typically made for llms.txt, delivered for a file about how much fictional cats enjoy being stroked.
What Is the Convergence Problem, and Why Should Marketers Care?
The convergence problem is the term the Search Engine Journal article uses for the mechanism that explains all four tests at once. When you ask a language model whether a tactic works, the model is not reasoning, running experiments, or weighing sources. It returns the most common thing it has seen written on the subject. When the internet was full of enthusiastic posts about cats.txt, ChatGPT said cats.txt works. When the discourse shifted and people explained the joke, ChatGPT reversed its position. Nothing about the file changed. Only the surrounding text on the web changed.
The article quotes Google’s John Mueller on llms.txt directly: “No AI system currently uses llms.txt… The consumer LLMs / chatbots (the ones that SEOs want traffic from) will fetch your pages, for training and grounding, but none of them fetch the llms.txt file. Maybe they will tomorrow? Maybe I’ll win in the lottery tomorrow?”
The article also references Ahrefs research across 100,000 domains showing that llms.txt is “largely ignored by the crawlers it is meant to court,” with no measurable citation advantage for sites that add one.
What This Means for AI-Search Visibility
The practical takeaway from cats.txt is simpler than a technical debate about file formats: most GEO “proof” circulating right now is pattern-matching dressed as causation. Here is a way to see it clearly. Imagine that every time you put away your umbrella, the rain eventually stopped, and someone concluded your umbrella was causing the rain to stop. The four tests for llms.txt follow the same shape: events that would happen regardless of whether the underlying mechanism does anything. Bots crawl everything. Google indexes text files. RAG pulls content that ranks. ChatGPT reflects whatever the internet says most often.
In practical terms, a brand’s AI-search performance depends on whether AI systems actually cite, recommend, or surface that brand to users asking relevant questions. The levers that plausibly affect that output, including content quality, structured data (standardized markup that helps machines parse page content), and presence in sources that AI models weight as authoritative, are all testable through observation that is independent of the tactic itself.
One thing the source article does not address, but is worth noting: the convergence problem cuts both ways. If a model’s endorsement of a tactic reflects the current consensus of internet text, then brands that consistently produce cited, authoritative content in their category may indirectly shape what models say about relevant questions over time. That is a plausible and testable GEO lever. Asking a model whether a file format works is not, because the answer reflects web consensus, not model behavior.
Hingewise’s read: the signal worth tracking is not “did a bot crawl my file?” but “does AI output about my category include my brand, and for which queries?” That question is answerable with structured observation, and the answer does not change the next time the discourse shifts.
Before You Commission a New GEO Tactic, Check These Things First
- Ask whether any major LLM provider (OpenAI, Anthropic, Google) has documented using this signal in their systems, not just in community discussion.
- Check whether the “proof” being offered would apply equally to an empty or fictional file by the same logic.
- Separate what a crawler did (fetched, indexed) from what the model actually did (cited your brand in a relevant answer to a real user question).
- Ask whether any ChatGPT endorsement of the tactic came before or after widespread industry promotion. Per the convergence problem, the timing explains the answer.
- Look for research at scale. The Ahrefs study of 100,000 domains on llms.txt, referenced in Search Engine Journal, found no measurable citation advantage. A single site’s experience is not a controlled test.
- Prioritize tactics with outputs you can observe independently: whether your brand appears in AI answers to specific queries, whether those appearances change after a specific action, and whether any change holds over time.
The cats.txt episode is unlikely to close the llms.txt debate. Adoption will continue, and future documentation from LLM providers may yet identify conditions under which the file influences model behavior. What the experiment does establish is a more useful bar for evaluating GEO claims: does the evidence hold up if you run the same test on something designed to be false? If yes, the evidence is not evidence. That question is worth asking about every new tactic before the invoice arrives.
Source: Search Engine Journal, “How Cats.txt Showed LLMs.txt Evidence Is GEO Astrology,” 2025. (searchenginejournal.com)
Lam Nguyen is the founder of Hingewise, an agency focused on AI-search visibility and GEO strategy.