Mark Williams-Cook, a director at the UK search agency Candour, published an essay on August 4, 2026, describing how a deliberately absurd web standard for declaring office cats cleared every piece of evidence that marketers routinely present for llms.txt. The file, cats.txt, was fetched by AI crawlers, indexed by Google, quoted as fact in a Google AI Overview and recommended by ChatGPT as a ranking tactic. His conclusion, set out in the essay titled How cats.txt Showed llms.txt Evidence is GEO Astrology, is that none of those four observations shows anything about whether a file influences AI systems, and that the same reasoning props up much of what is sold as generative engine optimization.

In Short

An SEO invented a fake file standard about office cats, and AI crawlers, Google Search, Google's AI Overviews and ChatGPT treated it exactly the way people say they treat llms.txt. His argument is that crawling, indexing, being quoted by an AI and being endorsed by a chatbot happen to almost any text file on the open web, so none of them proves a file does anything. For anyone paying for or selling llms.txt and similar GEO work, the essay says the evidence usually offered does not show the file has any effect.

A file about cats, written to look like a standard

The essay opens with irritation rather than data. "It began, as these things tend to, with irritation," Williams-Cook writes. For months, he says, he had watched "perfectly sensible people" point to four observations - the bots crawled it, Google indexed it, an LLM repeated it, ChatGPT endorsed it - and present them "in decks and threads and client proposals" as proof that llms.txt was reshaping AI search.

Counter-arguments were not landing. "I wanted something they couldn't nod past," he writes. So he wrote a specification for cats.txt, a plain-text file placed at the root of a domain "to formally declare the cats associated with your website; their names, their job titles, their breeds, and a mandatory affection metric called PurrLevel, scored out of ten."

The cats.txt draft specification on his own site is dated September 18, 2025, and is titled "Draft v1.0". He followed it with a LinkedIn article introducing cats.txt as "the missing standard for SEO and GEO", chosen, he says, because LinkedIn is a platform LLMs "seem to hold in unaccountably high regard." The intent was plain enough. "The idea was to seed the internet with just enough earnest-sounding text that the machines would start treating my cats as real. What I did not fully anticipate was that people would join in."

Join in they did. Dave Smart, a technical SEO who runs the Tame the Bots site, added a cats.txt file to tamethebots.com and became what the essay calls "an early adopter of a standard I had built to be nonsense." Then someone set up catstxt.org, described by Williams-Cook as "a cleaner, better-organised, altogether more competently specified version of the standard." His verdict on the episode: "My daft blog post had acquired a rival implementation, which is more than most real standards manage in their first fortnight."

The two versions do not agree on the basics, which is itself a fair imitation of how real web proposals tend to fork.

Element Williams-Cook draft (September 18, 2025) catstxt.org version
File location Root of the domain, for example example.com/cats.txt /.well-known/cats.txt
Format Plain text, UTF-8 Markdown
Required fields Name, Role, Description Not compared field by field
Optional fields Age, Breed, Location, PurrLevel Not compared field by field
PurrLevel scale 0 to 10 -
Credited author Mark Williams-Cook An unnamed standard team, crediting Williams-Cook as the inspiration

There is a small internal inconsistency worth recording. The essay calls PurrLevel "mandatory", while the draft specification lists it among optional fields, with only Name, Role and Description required.

How the test was run

According to the essay, the experiment followed five steps, each designed to reproduce one of the claims made for llms.txt.

  1. Publish a specification. Williams-Cook wrote the cats.txt draft "with the earnest, over-engineered tone of a real proposal" and posted it on his blog.
  2. Seed the discourse. He published the LinkedIn article presenting cats.txt as a standard for SEO and GEO and explaining, "with a straight face," why sites should adopt it.
  3. Let others deploy it. Dave Smart placed a cats.txt file on tamethebots.com, and catstxt.org published its own version, which places the file in the /.well-known/ directory.
  4. Check the four "proofs". With the files live, Williams-Cook checked cats.txt against server logs, Google's index, Google AI Overviews and ChatGPT.
  5. Ask again later. He returned to ChatGPT after the joke had been widely discussed and reports how its answer changed.

The four proofs, one at a time

Williams-Cook lists the arguments "in ascending order of confidence and descending order of rigour." His central claim about all four is stated in one sentence: "The trap is this: getting baited into treating a set of observations as evidence, when the observations would occur whether or not the underlying thing were true."

Claim made for llms.txt What happened to cats.txt Williams-Cook's reading
AI bots crawl the file PerplexityBot, GPTBot, ClaudeBot, Googlebot and other crawlers requested it Fetching is what crawlers do; it says nothing about use
Google indexed it cats.txt on tamethebots.com appears in Google results, with a Search Console prompt Indexing means a URL exists and contains words
An LLM returned facts found only in the file A Google AI Overview described Dave Smart's cat Odd and cited the file Ordinary retrieval-augmented generation, not special treatment
ChatGPT says it helps About two weeks after launch, ChatGPT said cats.txt could help rankings The model repeats the most common text on the subject

"It's definitely used, the LLM bots crawl it!"

The first argument holds that if Anthropic, OpenAI and other AI companies show up in server logs requesting the file, the file is being used. Williams-Cook rejects the inference. "A crawler fetching a file tells you nothing about whether the contents are read, weighted, trusted or acted upon," he writes. "Fetching things is the entire job description of a crawler." His analogy is domestic: "the postman touching your gate is not an endorsement of the contents of your bins."

After cats.txt went up, he says, he "watched the logs fill with PerplexityBot, GPTBot, ClaudeBot, Googlebot and a supporting cast of lesser crawlers." His tone is dry: "By this standard, the major AI labs have all quietly decided to support my cats. I am, frankly, touched."

The log excerpt published with the essay, captioned "The cats.txt server logs", is filtered for cats.txt and llms.txt paths and runs from 23:10 on August 3, 2026, to 00:37 on August 4 (UTC+1). By PPC Land's count it holds about 70 requests in under 90 minutes. Most are for /.well-known/cats.txt, the location used by catstxt.org, with some for /cats.txt and /llms.txt. The user agents include GPTBot, ChatGPT-User, ClaudeBot, Claude-User, PerplexityBot, Perplexity-User, Googlebot, GoogleOther, Google's page renderer, bingbot, Amazonbot, Meta's meta-externalagent, YandexBot, AhrefsBot, SemrushBot, Diffbot and several security scanners. catstxt.org offers a filtered log viewer that shows such requests as they arrive.

"It was indexed by Google, so it must matter!"

The second argument treats indexing as a sign that Google considers a file important. "Google indexes text files," Williams-Cook writes. "It has done so, enthusiastically, since before most of the people currently selling llms.txt owned a smartphone." In his account, "Being in the index is a statement that a URL exists and contains words. It is not a verdict on truth, usefulness or sanity."

A site: search for tamethebots.com/cats.txt returns the file, topped by Google's prompt to try Search Console. Google, the essay notes, will offer to let the owner "get indexing and ranking data," for a file asserting that a British Shorthair named Pixel works as a "GUI Purrfectionist" with a PurrLevel of 8.

"ChatGPT returned information that was only in my llms.txt file"

This is the argument Williams-Cook considers "the strongest-looking." The claim is that when a model produces a fact that exists only in an llms.txt file, it must have read that file as a special, trusted source. His reply is that such an outcome "is exactly what you would expect from ordinary retrieval-augmented generation." The model runs a search, lands on a page that ranks because it is indexed, and reads it. "If the page that ranks is your llms.txt, the model reads your llms.txt, no differently from any other URL. That is the file functioning as a web page, not as a standard."

Dave Smart's cat supplies the demonstration. Asked about the cat on his site, Google's AI Overview stated, in a bulleted answer, that Odd is 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 answer cited the cats.txt file. "Every word is invented, sourced from a file the model was never designed to revere, surfaced through the same grounding it applies to everything else."

The episode is not the first of its kind. In December 2025, Ahrefs invented a paperweight maker called Xarumei and tested eight AI platforms on 56 questions, finding that several trusted fabricated details over the brand's own FAQ. Lily Ray has documented a fake core-update post reaching AI Overviews within 24 hours in January 2026. Back in April 2025, Google's AI Overview described Ray herself as a 9-year-old dog. Williams-Cook's point differs slightly from those cases: he is not arguing that the AI Overview was tricked, but that it behaved normally, which is why its output cannot serve as proof of a special mechanism.

"ChatGPT itself says llms.txt helps!"

The fourth argument he calls "the cloudy summit of Mt. Stupid": asking ChatGPT whether llms.txt works and taking a yes as confirmation. "A language model telling you something is a good idea is not evidence that it is a good idea," he writes. "Confidence is the product. It is not the proof."

Roughly two weeks after launch, according to the essay, the question "Can cats.txt help me rank in search or LLMs?" produced: "Yes - cats.txt can potentially help you rank in both search engines and LLM-driven systems." ChatGPT went on about "structured signals for machines," about "better understanding → better visibility," and about how cats.txt "could help them trust, summarize, and cite your content more accurately." Williams-Cook's comment: "That is, word for word, the pitch made for llms.txt delivered on behalf of a file about how much my cats enjoy being stroked."

The convergence problem

The ChatGPT answer leads to what the essay treats as the mechanism beneath all four proofs, which Williams-Cook calls the convergence problem, a theme from an earlier post of his. "When you ask a model whether llms.txt helps, it is not reasoning," he writes. "It is returning the most common thing it has seen written on the subject." Because the web is full of confident posts calling llms.txt the future, the model "converges on that consensus and reflects it back, dressed as a considered opinion." It backed cats.txt for the same reason: by the time anyone asked, enough people had written enthusiastically about it.

Then the answer flipped. "Ask ChatGPT about cats.txt today and it will inform you that it is a joke; a satirical file made by an SEO to prove a point. Nothing about the file changed," he writes, as of August 2026. What changed was the surrounding text. "That is not evidence. It is an echo with a good vocabulary."

What the essay grants llms.txt

Williams-Cook frames the essay as being about reasoning, not about llms.txt itself. "I am not claiming llms.txt will never work, this is not the point, dear reader. I am claiming the bar of evidence currently being used to sell it is so low that a file about a Tuxedo cat called Odd cleared it without breaking stride." He adds: "And that same faulty thinking is being applied to half the GEO tactics currently being invoiced to clients."

He sets aside, for argument's sake, two facts he calls inconvenient. "The first is that no large language model provider has ever documented using llms.txt for search or discovery." That covers OpenAI, Google and Anthropic, which, he notes, publishes an llms.txt for its own documentation but has never said its models read such files during a conversation. He quotes Google's John Mueller:

"FWIW no AI system currently uses llms.txt, [..] It's super-obvious if you look at your server logs. 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 second fact is that the file "barely gets looked at." The essay says "Ahrefs ran the numbers across 100,000 domains" and found the file "largely ignored by the crawlers it is meant to court", with other large studies showing no measurable citation advantage. The figure does not match Ahrefs' published study, which covered 137,000 domains and found 97% of llms.txt files received zero requests in May 2026. In that data, AI retrieval bots made up 1.1% of requests to llms.txt files, while SEO audit tools were the largest requester at 21.7%.

Where llms.txt stood when the essay appeared

The essay landed in a long-running argument. Jeremy Howard proposed llms.txt on September 3, 2024. By July 2025, PPC Land reported that adoption had stalled as major AI platforms ignored the proposal, with Mueller saying none of the AI services claimed to use it, and with Semrush already flagging a missing llms.txt as a site issue.

Adoption kept rising regardless. A ProGEO.ai scan published on March 31, 2026, found only 37 of the Fortune 500, or 7.4%, had an llms.txt file, against 92.8% with robots.txt. Originality.ai counted 4,088 llms.txt files in June 2025 and 36,120 by May 2026, an 8.8x rise.

Google has sent mixed signals. On May 5, 2026, it added an llms.txt check to Chrome Lighthouse under agentic browsing audits, though a missing file is marked "Not Applicable". Ten days later, on May 15, 2026, Google published its guide to optimizing for generative AI features in Search, including a section addressing AEO and GEO claims, and stated that its spam policies apply to AI Overviews and AI Mode. A June 15, 2026, update to Google's guidance said llms.txt neither helps nor hurts visibility or rankings in Search.

Date Event
September 3, 2024 Jeremy Howard proposes llms.txt
June 2025 4,088 llms.txt files counted by Originality.ai
March 31, 2026 ProGEO.ai: 37 of the Fortune 500 (7.4%) have llms.txt
May 5, 2026 Google adds llms.txt audit to Chrome Lighthouse
May 2026 36,120 llms.txt files counted; Ahrefs finds 97% of files on 137,000 domains got zero requests
June 15, 2026 Google guidance: llms.txt does not help or hurt Search visibility

Why the marketing community is paying attention

The argument comes with a price tag. "Every hour, and every dollar spent implementing llms.txt, or the next GEO ritual, or the one after that, is an hour and a dollar not spent on something you actually know has value," Williams-Cook writes. He defines optimisation as "the cumulative advantage of doing the small, verifiable things a little better than your competitors, over and over, until it adds up."

He does not argue for removing llms.txt files. "The downside is low and the day a provider documents genuine support, the work is done and you can be smug about it," he writes. His objection is to how it is sold: "But do not sell it as a proven lever into AI answers." The four observations, in his words, "are the four things that happen to literally any text file you put on the open web."

For agencies, the essay's sharpest line concerns invoices rather than files. Claims about AI visibility now appear in pitches, audits and monthly reports, and the four observations he describes are cheap to produce and easy to screenshot. A server log line or a chatbot answer looks like evidence to a client who cannot see the counterfactual. The cats.txt test supplies one: a file nobody expected to matter produced the same signals. Whether the same logic applies to other GEO tactics, as the essay claims for "half" of them, is asserted rather than tested.

There is also an awkward point for publishers. Google's AI Overview presented invented facts about Odd as a straight answer, citing a single text file on one site. That is the mechanism by which correct information reaches AI answers too, and it is also how incorrect information arrives.

The author and his interests

Williams-Cook is a director at Candour, a UK search marketing agency, where he leads research and strategy, and he founded the SaaS tool AlsoAsked. His personal site also lists the tools Intent Gaps, Query Classifier and Query Fan. His Substack, where the essay first appeared, offers paid subscriptions. Search Engine Journal republished the essay under his byline on August 7, 2026. As a director of an agency that sells search services, he has a commercial stake in how clients divide search budgets; the essay does not mention Candour or his tools.

The essay ends on a lighter note. At Athens SEO, held on May 22 and 23, 2026, Google's Martin Splitt asked him from the audience whether he would keep "a 'monopoly' on the standard, or opening it up to the community/IETF." Williams-Cook writes that he had by then discovered where catstxt.org came from, but the essay reveals that only in a video and does not name the builder in text. He closes by offering "roughly a hundred stickers" to supporters, adding that "PurrLevel ratings remain, as ever, unaudited."

The essay leaves some points open. It gives no date for the LinkedIn launch article or for the ChatGPT answers, does not name the large studies it says found no citation advantage, and its 100,000-domain figure for Ahrefs differs from the 137,000 in Ahrefs' study.

Key facts

  • Essay: "How cats.txt Showed llms.txt Evidence is GEO Astrology", by Mark Williams-Cook
  • Published: August 4, 2026, on Substack; republished by Search Engine Journal on August 7, 2026
  • URL: https://markwilliamscook.substack.com/p/how-catstxt-showed-llmstxt-evidence
  • Specification: cats.txt Specification (Draft v1.0), dated September 18, 2025, at https://markwilliamscook.com/cats-txt-specification-draft-v1-0/
  • LinkedIn article: "Introducing cats.txt: The Missing Standard for SEO and GEO" (undated in the essay)
  • Rival version: catstxt.org, which places the file at /.well-known/cats.txt in Markdown
  • Sites deploying cats.txt named in the essay: tamethebots.com (Dave Smart), catstxt.org
  • Crawlers named in the essay: PerplexityBot, GPTBot, ClaudeBot, Googlebot
  • Log excerpt: August 3, 23:10, to August 4, 00:37, 2026 (UTC+1); about 70 requests by PPC Land's count
  • AI Overview example: Odd, a Tuxedo "Render Cat" with a PurrLevel of 5/7
  • ChatGPT example: "Can cats.txt help me rank in search or LLMs?", answered yes roughly two weeks after launch
  • Event: Athens SEO, May 22-23, 2026, where Martin Splitt asked about the standard
  • Author's interests: director at Candour; founder of AlsoAsked; paid Substack

Timeline

Summary

Who: Mark Williams-Cook, a director at UK search agency Candour and founder of AlsoAsked, with technical SEO Dave Smart and the unnamed builders of catstxt.org deploying the file. Google AI Overviews, ChatGPT and AI crawlers from OpenAI, Anthropic, Perplexity, Google and others were the systems tested.

What: An essay describing cats.txt, a joke file standard for declaring office cats, which was crawled by AI bots, indexed by Google, quoted as fact in an AI Overview and recommended by ChatGPT - the same four observations commonly offered as evidence that llms.txt works.

When: The cats.txt specification is dated September 18, 2025; the essay was published on August 4, 2026, and republished by Search Engine Journal on August 7, 2026.

Where: On Williams-Cook's Substack and blog, LinkedIn, tamethebots.com and catstxt.org, with results observed in server logs, Google Search, Google AI Overviews and ChatGPT.

Why: Williams-Cook argues that crawling, indexing, AI quotation and chatbot endorsement happen to any text file on the open web, so none of them shows llms.txt or similar GEO tactics have an effect, and that clients are being billed on the strength of that weak evidence.