In the summer of 2026, a British SEO consultant published a file on his site called cats.txt — an invented "standard" in which you officially declare the cats associated with your website: names, breeds, purr level. He wrote it as a formal specification, announced it on LinkedIn as "the missing standard for SEO", and waited.
What happened next should change how you buy any optimization service: the absurd file passed every test used in viral posts to prove that new tactics "work". AI bots fetched it (GPTBot, ClaudeBot, PerplexityBot). Google indexed it. And ChatGPT ended up citing the fictional cat from the file — because the page ranked in classic search. The experiment is public↗, data included.
The moral isn't about cats. It's that the same "evidence" was simultaneously selling real services for real money — llms.txt optimization, specifically. This article gives you the method: how to verify an SEO claim before you pay for it.
How a myth becomes "official truth":The mechanism
The phenomenon has a predictable pattern — we call it convergence:
- The real observation. Someone sees something true in their logs: "AI bots fetched my llms.txt!". The fact is real; the interpretation isn't — bots fetch any public file.
- The viral post. The observation becomes "proof": a screenshot, a tone of revelation, "the standard Google doesn't want you to know".
- The chorus. Other authors repeat the conclusion without repeating the verification. Each new post cites the old posts, not the source.
- Convergence. AI models, trained partly on public discourse, start repeating the chorus — and "even ChatGPT confirms it" becomes the closing sales argument. The circle closes: the myth cites itself.
At every stage, one thing was missing: the control test — the question "would something obviously absurd also pass this test?". Cats.txt demonstrated the answer: yes, it would.
The control test:15 minutes before any invoice
Four questions, in the order that cuts fastest:
What does the official source say — not "an article about" the source?
For any claim about Google: read Google's documentation and official statements, not interpretations. On llms.txt, John Mueller said explicitly that no AI system uses it — the statement was public WHILE "llms.txt optimization" services were being sold.
Would the evidence also pass on something absurd?
This is the cats.txt test. "Bots fetched it" — bots fetch everything. "It's indexed" — Google indexes cat files too. "It shows up in ChatGPT" — anything that ranks classically does. Evidence that validates the absurd validates nothing.
Who wins if you believe it?
Does the viral post sell something? The course, the service, the "compliance audit for the new standard"? Interest doesn't automatically falsify the claim — but it tells you where to press harder with questions 1 and 2.
Can it be measured on YOUR site, within a month?
Every real lever leaves measurable traces: impressions, positions, clicks, in Search Console, with a start date. "Be patient, it's long-term strategy" is sometimes true — but never a reason for there to be NO intermediate metric. Ask for the metric and the date you'll read it together; a refusal is your answer.
A real-time case study:How llms.txt was sold
The timeline, with what was verifiable at each step: the file appeared as a proposal in 2024; by 2025, "llms.txt optimization packages" were on sale, while Google's official statements that the file is ignored were already public. In 2026 the cats.txt control test showed how little the usual evidence proves, and on June 15, 2026, Google added a note to its guide to generative AI features↗: llms.txt files aren't needed for Google Search and don't move visibility or rankings in either direction, though keeping one for other services is fine. At every step, the four questions above would have stopped the invoice.
Two more quick applications of the same test: "guaranteed SEO results in 7 days" fails question 4, and "the AI detector says your text is generated" fails question 2.
The myths of 2026, through the control test
Five claims that circulated in 2026, each run through the four questions above. Two of them were ours, and we corrected both on October 29, 2026.
- "Google has a patent that detects AI-written text." We spread this one ourselves, in three of our articles. The first version of our March 2026 spam update analysis gave application 20260089127 as that patent, though it is another company's filing on another subject, and it placed the volatility on March 8–22 from tracker data, while Google's Search Status Dashboard puts the update on March 24–25. Question 1 would have caught both: the patent record and the dashboard were public.
- "High Discover impressions signal strong E-E-A-T." Ours too: our advanced Search Console guide said so until October 29, 2026. Google's rule for Discover↗ is plain: indexed content within the content policies can appear, with no promise that it will, and its Discover page never mentions E-E-A-T. It fails question 1.
- "Search Console shows nothing about AI." Since August 2024, Google's log of documentation updates↗ has recorded that AI Overviews are counted in the Performance report, with clicks, impressions and position logged the same way as for featured snippets; that note clarified the method, it did not add a report. The separate view came on June 3, 2026, when Google announced generative AI reports↗, open to every site since August 31: impressions only, with no clicks and no queries. A report that promises your Google AI citations query by query must get them elsewhere, and question 1 asks where.
- "AI answers need AEO/GEO tricks of their own." The Google guide linked in the timeline above asks site owners to put effective SEO ahead of such tricks, and for Google Search it names three that can be skipped: content cut into chunks for machines, AI text files without a use (llms.txt included) and mentions that aren't authentic. It fails question 1, and question 3 shows who gains from the opposite claim.
- "No schema, no AI citations." Google's page on AI features↗ sets no extra requirement and no special markup for AI Overviews or AI Mode: the page must be indexed and eligible to show a snippet. Structured data still does its usual job, rich results, provided it matches what the page shows. It fails question 1.
What we do differently — and what you can copy for free
We're not immune to myths; we're procedurally disciplined: nothing enters our articles or reports without an official source or our own dated measurement, and when a hypothesis of ours fails the evidence — it gets retracted publicly, not repainted. The method is fully copyable and free: the four questions above, applied in 15 minutes, before any signature.
FAQ.PROTOCOL
Frequently Asked Questions
Corrections and updates
- October 29, 2026. A new section runs five claims from 2026 through the control test, two of them our own mistakes corrected the same day, and the llms.txt timeline now cites Google's documentation instead of unnamed studies.
