How Cats.txt Showed LLMs.txt Evidence Is GEO Astrology

Aug 07, 2026 07:00 PM - 3 weeks ago 321

I sewage tired of watching the manufacture dainty “an AI bot fetched it” and “ChatGPT said it helps” arsenic grounds that llms.txt does anything, truthful I invented a modular called cats.txt: a matter record successful which you formally state your agency cats, their jobs, their breeds, and really often they purr. I wrote a specification, published it connected my blog, and did a LinkedIn station explaining why you should decidedly adopt it, because arsenic we each know, ample connection models emotion LinkedIn. Then, I checked it against the nonstop 4 “proofs” group mention for llms.txt. It passed each four. It was crawled by the AI bots. Google indexed it. LLMs returned specifications astir a feline that exists obscurity but the file. ChatGPT confirmed, astatine length, that cats.txt could thief maine rank. None of which is grounds of anything, which was alternatively the point.

I americium not claiming llms.txt will ne'er work. This is not the point, beloved reader. I americium claiming the barroom of grounds presently being utilized to waste it is truthful debased that a record astir a Tuxedo feline called Odd cleared it without breaking stride. And that aforesaid faulty reasoning is being applied to half the GEO tactics presently being invoiced to clients.

How A File About My Cats Came To Be A ‘Web Standard’

It began, arsenic these things thin to, pinch irritation.

For months I had been watching perfectly sensible group constituent astatine 4 observations: the bots crawled it, Google indexed it, an LLM repeated it, ChatGPT endorsed it, and coming them, successful decks and threads and customer proposals, arsenic impervious that llms.txt was softly reshaping AI search. None of it was impervious of anything. But statement by counter-argument only gets you truthful far; group motion on and past spell backmost to their slides. I wanted thing they couldn’t motion past. I wanted to tally the aforesaid 4 “proofs” connected thing truthful transparently ridiculous that nary 1 could dress the tests meant anything.

So I invented a standard. cats.txt: a plain-text record you spot astatine the guidelines of your domain to formally state the cats associated pinch your website; their names, their occupation titles, their breeds, and a mandatory affection metric called PurrLevel, scored retired of 10. I wrote a due specification for it, pinch the earnest, over-engineered reside of a existent proposal, and published it connected my blog. Then, because I cognize arsenic good arsenic anyone which level LLMs look to clasp successful unaccountably precocious regard, I wrote a LinkedIn article introducing cats.txt arsenic “the missing modular for SEO and GEO” and explaining, pinch a consecutive face, why you should decidedly adopt it.

The thought was to seed the net pinch conscionable capable earnest-sounding matter that the machines would commencement treating my cats arsenic real. What I did not afloat expect was that group would subordinate in.

The joke was legible, that was ever the point, and truthful the SEO organization picked it up and ran pinch it, precisely because they could spot wherever it was going. My beautiful internet-peer Dave Smart (a genuinely fantabulous method SEO) added a cats.txt to his ain site and became, to his eternal credit, an early adopter of a modular I had built to beryllium nonsense. And past the point took connected a life of its own: Someone went disconnected and group up catstxt.org, a cleaner, better-organized, altogether much competently specified type of the standard: evidently the activity of personification who knew what they were doing, and conscionable arsenic evidently not me. My daft blog station had acquired a rival implementation, which is much than astir existent standards negociate successful their first fortnight.

With the record live, the spec published, the LinkedIn station seeded and different group cheerfully piling in, each that remained was to cheque cats.txt against the nonstop barroom the manufacture uses to certify llms.txt. Reader, it cleared it.

A Word Of Genuine LLMs.txt Fairness First

I do not overmuch attraction whether llms.txt works, will work, aliases really agelong it takes to get there. For the magnitude of this argument, I americium happy to parkland 2 inconvenient facts and assistance the thought each use of the doubt.

The first is that nary ample connection exemplary supplier has ever documented utilizing llms.txt for hunt aliases discovery. Not OpenAI, not Anthropic (who people 1 for their ain docs and person still ne'er said their models publication it during a conversation), and not Google. Google’s John Mueller has been astir arsenic blunt arsenic a hunt advocator gets:

“FWIW nary AI strategy presently uses llms.txt, [..] It’s super-obvious if you look astatine your server logs. The user LLMs / chatbots (the ones that SEOs want postulation from) will fetch your pages – for training and grounding, but nary of them fetch the llms.txt file. Maybe they will tomorrow? Maybe I’ll triumph successful the lottery tomorrow?”

John Mueller, Google

The 2nd is that moreover wherever it is deployed, it hardly gets looked at. Ahrefs ran the numbers crossed 100,000 domains and recovered that the record is, successful practice, mostly ignored by the crawlers it is meant to court, a uncovering since echoed by different ample studies showing no measurable citation advantage for sites that adhd one. So the system group are paying for does not look to fire. Fine. Park that too.

Assume the assemblage is retired connected some counts and assistance the thought the astir generous proceeding imaginable. The problem I really want to talk astir is not llms.txt astatine all. It is the reasoning being utilized to take sides it.

The Faulty Thinking

The trap is this: Getting baited into treating a group of observations arsenic evidence, erstwhile the observations would hap whether aliases not the underlying point were true. It is the intelligence balanced of concluding your umbrella causes the rainfall to stop, because each clip you put it distant the rainfall does yet stop.

llms.txt is simply a convenient example. The aforesaid surgery concatenation of conclusion is being applied to almost each caller GEO maneuver invented connected a monthly basis, and the mobility is ever the same, “Should we do this thing, aliases should we not?” – which makes the value of the answers alternatively important.

Here are the 4 “proofs” I support being shown, successful ascending bid of assurance and descending bid of rigour.

1. ‘It’s Definitely Used, The LLM Bots Crawl It!’

The first argument: You tin spot Anthropic crawling it, you tin spot OpenAI crawling it, the bots move up successful your logs, truthful the record is being used.

A crawler fetching a record tells you thing astir whether the contents are read, weighted, trusted, aliases acted upon. Fetching things is the full occupation explanation of a crawler. Bots petition much aliases little everything you time off lying around; the postman rubbing your gross is not an endorsement of the contents of your bins.

To beryllium the point, I put up cats.txt and watched the logs capable pinch PerplexityBot, GPTBot, ClaudeBot, Googlebot and a supporting formed of lesser crawlers, each diligently requesting a record describing the master responsibilities of my cats. By this standard, the awesome AI labs person each softly decided to support my cats. I am, frankly, touched.

The catstxt.org website moreover offers a filtered log viewer, if you want to watch each that crawling action live.

2. ‘It Was Indexed By Google, So It Must Matter!’

The 2nd argument: The record was indexed by Google, which proves Google considers it important, because why would Google scale thing that didn’t matter?

Google indexes matter files. It has done so, enthusiastically, since earlier astir of the group presently trading llms.txt owned a smartphone. Being successful the scale is simply a connection that a URL exists and contains words. It is not a verdict connected truth, usefulness, aliases sanity.

cats.txt is, naturally, indexed. Google will moreover connection to fto you declare it successful Search Console and “get indexing and ranking data,” pinch the straightest of faces, for a record asserting that a British Shorthair named Pixel useful arsenic a “GUI Purrfectionist” pinch a PurrLevel of 8.

3. ‘ChatGPT Returned Information That Was Only In My LLMs.txt File’

The 3rd statement is the strongest-looking, and truthful deserves the astir care. The declare is that a exemplary produced a truth that existed only wrong the llms.txt file, and truthful must person publication the record arsenic a special, trusted source.

The problem is that this is precisely what you would expect from mean retrieval-augmented generation. The exemplary runs a search, lands connected a page that happens to rank because it is indexed (see: erstwhile argument), and sounds immoderate is connected it. If the page that ranks is your llms.txt, the exemplary sounds your llms.txt, nary otherwise from immoderate different URL. That is the record functioning arsenic a web page, not arsenic a standard.

Consider Dave. Lovely Dave. A real, method SEO of bully opinionated put a cats.txt connected his site, becoming an early adopter of a modular I had built to beryllium nonsense. Ask Google astir the feline that lives connected his tract and the AI Overview will show you, successful a assured bulleted answer, that Odd is simply a “Render Cat,” a Tuxedo pinch a PurrLevel of 5/7, who “chases the cursor, pounces connected stray pixels, and stashes them connected the integer carpet.” It cites the cats.txt file. Every connection is invented, originated from a record the exemplary was ne'er designed to revere, surfaced done the aforesaid grounding it applies to everything else.

4. ‘ChatGPT Itself Says LLMs.txt Helps!’

The fourth, the cloudy acme of Mt. Stupid. You inquire ChatGPT whether llms.txt works; it tells you yes, that it tin astir apt help, you should do it, and you return that arsenic confirmation from the horse’s mouth.

A connection exemplary telling you thing is simply a bully thought is not grounds that it is simply a bully idea. It is grounds that a awesome woody of matter connected the net says it is simply a bully idea, and the exemplary has learned to manus that matter backmost to you pinch full composure. Confidence is the product. It is not the proof.

Roughly 2 weeks aft launch, you could inquire ChatGPT, “Can cats.txt thief maine rank successful hunt aliases LLMs?” and receive: “Yes — cats.txt tin perchance thief you rank successful some hunt engines and LLM-driven systems.” It went on, unprompted, astir “structured signals for machines,” astir “better knowing → amended visibility,” and astir how, for AI systems, cats.txt “could thief them trust, summarize, and mention your contented much accurately.” That is, connection for word, the transportation made for llms.txt delivered connected behalf of a record astir really overmuch my cats bask being stroked.

The Convergence Problem

This past 1 is not simply funny. It is the system underneath each four, and it is worthy naming: the convergence problem.

When you inquire a exemplary whether llms.txt helps, it is not reasoning. It is not moving an experiment, consulting a source, aliases weighing evidence. It is returning the astir communal point it has seen written connected the subject. The web is heavy pinch assured posts declaring llms.txt the future, truthful the exemplary converges connected that statement and reflects it back, dressed arsenic a considered opinion. It endorsed my cats for precisely the aforesaid reason: by the clip anyone asked, capable group had written enthusiastically astir cats.txt that the mean of the sermon said “yes.”

Ask ChatGPT astir cats.txt today, and it will pass you that it is simply a joke; a satirical record made by an SEO to beryllium a point. Nothing astir the record changed. What changed is the surrounding matter connected the internet: the sermon caught up, admitted the gag, and the exemplary dutifully converged connected the new most-common answer. The exemplary was ne'er assessing the standard. It was, and ever is, taking a moving mean of what everyone other is saying. That is not evidence. It is an echo pinch a bully vocabulary.

Why Any Of This Matters

I americium not doing this purely for sport, though I will admit the athletics is excellent.

There is simply a existent costs hiding nether the comedy. Every hour, and each dollar spent implementing llms.txt, aliases the adjacent GEO ritual, aliases the 1 aft that, is an hr and a dollar not spent connected thing you really cognize has value. That is what the “O” successful SEO is meant to guidelines for. Optimization is the cumulative advantage of doing the small, verifiable things a small amended than your competitors, complete and over, until it adds up. It is not chasing a record that gets crawled, indexed and confidently endorsed by a strategy that will reverse its verdict the infinitesimal the sermon shifts underneath it.

So, by each means, adhd an llms.txt if it makes you consciousness prepared for a early that whitethorn arrive. The downside is low, and the time a supplier documents genuine support, the activity is done, and you tin beryllium smug astir it.smugness is the champion kind. But do not waste it arsenic a proven lever into AI answers, and do not constituent astatine “the bots crawled it” aliases “ChatGPT said it helps” arsenic though either condemnation contained a fact. It doesn’t. Those 4 observations are the 4 things that hap to virtually immoderate matter record you put connected the unfastened web, including 1 describing a Maine Coon named Byte who hunts stray zeroes and ones crossed the server racks.

The cats, astatine least, were honorable astir being made up. I stay unconvinced the aforesaid tin beryllium said for everything other being sold this year.

I did, however, bask astatine this year’s Athens SEO, an assemblage mobility aft my talk from Martin Splitt, asking me, since inventing cats.txt, whether I would beryllium keeping a “monopoly” connected the standard, aliases opening it up to the community/IETF. He didn’t cognize that I had since discovered wherever catstxt.org had travel from:

https://cdn.searchenginejournal.com/wp-content/uploads/2026/08/2026-08-05_Video_1_default.mp4

Thank you to Iva Jovanovic for capturing this beautiful infinitesimal connected video.

The cats.txt draught specification is here. The LinkedIn announcement that started it is here. PurrLevel ratings remain, arsenic ever, unaudited.

More Resources: