Search Engine Journal reported on 7 August 2026 that Mark Williams-Cook invented a satirical file standard called cats.txt and used it to show that the evidence usually offered for llms.txt proves nothing. The cats.txt file is a plain text document listing office cats with job titles, breeds and a PurrLevel affection score from 1 to 10, and the report says it satisfied every one of the four proof points that get cited as confirmation that llms.txt works.
That is the whole finding, and it is narrower and sharper than the headlines around it suggest. The argument is about the quality of the evidence, not about whether llms.txt is good or bad. Search Engine Journal's piece does not claim llms.txt damages a site. It claims the four observations people point at when they defend llms.txt would happen for any file sitting on a web server, including one about cats.
What Mark Williams-Cook actually built
According to Search Engine Journal's 7 August 2026 report, Mark Williams-Cook did not stop at making the argument in a post. He published a formal specification for cats.txt, written the way a genuine standard would be documented, and then wrote a LinkedIn article promoting its adoption. The file contents are exactly what the name suggests: office cat names, invented job titles, breeds, and a PurrLevel score out of 10 describing how affectionate each cat is.
The construction is what makes the cats.txt experiment function as a test rather than a joke. Because the standard was made up, there is no mechanism by which a cats.txt file could improve anything for the site hosting it. So if the observations that supposedly demonstrate llms.txt effectiveness also show up for cats.txt, those observations are not measuring effectiveness. They are measuring the fact that a text file exists at a URL and is reachable.
The four proof points, stated the way the source states them
Search Engine Journal listed four observations that the cats.txt file satisfied, and these are the same four that circulate as evidence for llms.txt.
- AI crawlers fetched the file, which was visible in server logs.
- Google indexed the file.
- ChatGPT returned cat details that existed nowhere except inside the file.
- ChatGPT itself confirmed that cats.txt could help with ranking.
Why each of the four proof points is circular
The reading of each observation below is analysis, not a sentence by sentence claim from Search Engine Journal.
The first proof point, a crawler hit in the server log, records a request. A request tells you a piece of software asked for a URL and got a response. It carries no information about whether the response changed a model output, was retained, or was weighted at all. Any file that returns a 200 status can produce this line in a log.
The second, indexation, means Google stored a plain text document it found. Google indexes text files as a matter of course. Indexation says the document was crawlable and not blocked. It does not say the document influenced anything, and it is not a signal that ranks the site the file sits on.
The third is the one that feels strongest and is the most misleading. When ChatGPT repeats a cat name and a PurrLevel that appear only inside the file, it shows the assistant fetched a URL and read the text. That is retrieval. It is the same behaviour as fetching any page on the domain. It does not establish that the file was preferred over other pages, that it was used when a user did not point at it, or that the file changed how often the site gets cited in ordinary answers.
The fourth, ChatGPT agreeing that the file helps, is the weakest of all. Search Engine Journal notes that ChatGPT told Williams-Cook cats.txt could help with ranking and visibility. A language model producing an agreeable answer about a file it has just been shown is a statement about how the model handles a leading question, not a statement about search systems. The model has no privileged view of Google's ranking pipeline or of its own retrieval weighting.
What John Mueller said about llms.txt
The article also quotes Google's John Mueller, who said "no AI system currently uses llms.txt" and added that the consumer chatbots SEOs want traffic from do fetch pages, for training and grounding, but that none of them fetch the llms.txt file itself. That is a direct statement from Google about the current state, and it sits alongside the cats.txt result rather than depending on it.
The four claims and what each one supports on its own
The table below lists the four proof points exactly as Search Engine Journal reported them on 7 August 2026, with a plain reading of what each observation can support by itself. The right column is analysis, not a quotation from the source.
| Observation cited as proof | What it supports on its own |
|---|---|
| AI crawlers fetched the file, seen in server logs | A request was made and answered. Nothing about whether the content was used. |
| Google indexed the file | The file was crawlable and text based. Indexation is not a ranking outcome. |
| ChatGPT cited details found only in the file | The assistant read that URL. It does not show the file was chosen over other pages. |
| ChatGPT said the file could help ranking | The model answered a question agreeably. It has no visibility into ranking systems. |
What would count as evidence instead
Search Engine Journal's piece frames the problem as observations that would occur whether or not the underlying claim were true. The fix for that, in ordinary measurement terms, is a comparison rather than a sighting.
- A controlled comparison: two comparable groups of pages or sites, one with the file and one without, measured over the same window.
- A measured change in citation rate: how often the site is named or linked in assistant answers before and after, on a fixed prompt set that does not point the assistant at the file.
- A holdout: a set of properties deliberately left untouched, so any movement in the treated set has something to be compared against.
- Prompts that a real person would type, rather than prompts that hand the assistant the URL and then treat the retrieval as discovery.
None of those designs appears in the reported experiment, and the article does not claim to have run them.
What the source does not establish
Search Engine Journal's report does not test llms.txt against a control group. It does not measure whether sites with llms.txt get cited more or less than sites without it. It does not show that llms.txt harms a site. It does not say llms.txt will never be adopted by any AI system in future, and Mueller's statement is explicitly about what systems do currently rather than a permanent position. The article also does not say how long the cats.txt file was live, how many crawler hits it received, or which specific crawlers appeared in the logs.
The claim that survives all of that is narrow and still useful: the four observations in circulation cannot distinguish a working standard from a fabricated one, so they cannot be used to argue that llms.txt works.
Running the same test on any AI visibility claim
The structure of the cats.txt experiment transfers to almost any pitch that arrives with screenshots attached. The question to ask is whether the evidence offered would look identical if the claim were false.
- Write down the observation the vendor is showing, without the interpretation attached, then ask what else could produce it. A log line, an index entry and a retrieval all have mundane explanations.
- Ask what the comparison group is. If there is none, the demonstration is a sighting.
- Check whether the assistant was handed the URL during the demonstration. If it was, retrieval was arranged rather than earned.
- Ask for the measurement that would have looked different if the thing did not work, and treat a missing answer as the answer.
Concretely, a marketer can pull this apart in their own account without buying anything. Server logs or a log file analyser will show which AI user agents requested which URLs and how often. Google Search Console's page indexing report will show whether a text file was indexed, which is a separate question from whether any page ranks better. A fixed list of 20 to 50 prompts, run monthly without naming the URL, gives a citation baseline that a screenshot cannot. Work on AI SEO and generative engine optimisation is easier to defend when the baseline exists before the change does.
What this means for Thai marketers
Search Engine Journal's 7 August 2026 article says nothing about Thailand, Thai language prompts or the Thai market. There is no Thai angle in the source, and reading one into it would be inventing detail.
What does travel is the commercial situation. llms.txt currently shows up on proposals and scopes of work as a deliverable, sometimes priced, in Thailand as elsewhere. If a supplier is charging for it, the four proof points are very likely the evidence attached. Adding the file costs little and the article gives no reason to believe it causes harm. Paying a meaningful fee for it, on the strength of a log line and a screenshot, is a different decision. A buyer is entitled to ask what the comparison was, and to accept honest uncertainty as an answer. That question belongs in the same conversation as the rest of a Thailand SEO scope, alongside the items that do have measurable outcomes attached.
FAQ
Does this mean llms.txt is useless?
The source does not say that, and it is worth resisting the jump. Search Engine Journal's article argues that the four proof points commonly cited as evidence are worthless as evidence, because a fabricated standard cleared all four. It does not test llms.txt against a control, and it does not claim the file harms a site. Google's John Mueller is quoted saying no AI system currently uses llms.txt, which is a statement about the present rather than a prediction.
Is any of this specific to Thailand?
No, and the source does not mention Thailand at all. The experiment and Mueller's comment are both about how AI systems and Google handle the file generally, with no country or language breakdown given.
My server logs show AI bots fetching my llms.txt file. Is that proof it works?
No, and this is precisely the observation the cats.txt experiment was built to defuse. A log line records that a crawler requested a URL and received a response. It says nothing about whether the content was retained, weighted or used in any answer, and the same log line appeared for a file listing office cats.
What should I ask a vendor selling AI visibility work?
Ask what the comparison group was and what would have looked different if the tactic had not worked. If the answer is a screenshot of an assistant repeating text after being handed the URL, that is retrieval of a page rather than evidence of improved visibility. A fixed prompt set measured before and after, without naming the URL, is the cheapest honest baseline available.
The short version
A file about office cats passed every test that llms.txt advocates use to prove their case, which means those tests cannot tell a real standard from a fake one. The correction is not cynicism about AI search, it is asking for comparisons instead of sightings, and being willing to say that a question is still open. If you want a second opinion on what an AI visibility proposal actually promises, and what could be measured to check it, Relevant Audience is happy to look at the numbers with you.
Reporting by Search Engine Journal, published 7 August 2026.
