Google says llms.txt does nothing for search. We checked our own 84 posts.

Two documents and one setting matter.

The first is “AI features and your website”, last updated December 10, 2025. Its position on eligibility is flat: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” To be shown as a supporting link, “a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements. There are no additional technical requirements.”

The second is “Optimizing your website for generative AI features on Google Search”, announced on May 15, 2026. It contains a mythbusting section, and the myths are the product catalogue of a whole industry.

On llms.txt the wording leaves no room: “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn’t use them.” And then, for anyone hoping it at least cannot hurt: “It’s completely fine if you decide to create and maintain LLMS.txt files (or other similar files) for other services or systems that use these files. Doing so will neither harm nor help your site’s visibility or rankings in Google Search, as Google Search ignores them.”

On chunking: “There’s no requirement to break your content into tiny pieces for AI to better understand it.” And, in the same passage, “There’s no ideal page length.”

On writing style: “You don’t need to write in a specific way just for generative AI search.” On schema: “Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.” On mentions: “seeking inauthentic ‘mentions’ across the web isn’t as helpful as it might seem.”

There is also a warning aimed squarely at vendors: “Be wary of third-party tools that promise ranking success or claim to use ‘internal’ Google metrics. No third-party tool has access to our internal ranking or AI systems.”

And a trap worth reading twice, because it sounds like sensible advice for AI Mode. Google describes a “query fan-out” technique, where it issues multiple related searches across subtopics to build an answer. The obvious move is to build a page for each of those subtopics. Google’s guide says that doing it “primarily to manipulate rankings or generative AI responses in Google Search violates Google’s scaled content abuse spam policy.”

The third thing is not advice, it is a switch. Google’s Search Console help page for the Search generative AI control states: “As of August 31, 2026, we’ve rolled out this control to all websites worldwide.”

It governs whether your content can appear in AI Overviews, AI Mode and generative features in Discover. The default is to include you. Google also says the control “isn’t used as a ranking or inclusion signal affecting other parts of Search”, so switching it off does not demote you in ordinary results, and switching it on does not lift you.

The reason to look is simpler than any optimization. Somebody on your team, or an agency before us, may have turned it off. Five minutes in Search Console settles it, and no amount of content work overcomes a property that has opted out.

We turned the guide into seven rules, each one a claim Google explicitly rejects, and ran them on October 15, 2026 across all 84 posts then live, sentence by sentence, body text and FAQ answers: llms.txt presented as helping, advice to chunk, advice to write for AI specifically, special schema presented as required, chasing mentions, a page per query variation, and any tool claimed to read Google’s internals. The checker flags sentences; a person then read every flagged post in full. The rules, every flagged sentence and the verdict for each post are published with this post: the method, the per-post results and the flagged sentences.

Nineteen of the twenty-four present it as optional, unproven or ignored by Google. The wording now in those posts reads, “an optional summary for compatible tools, not a ranking factor or citation guarantee”, and “llms.txt is an optional discovery file worth 2 of 66 audit points, not a citation requirement”, and, in our state of AI search piece, llms.txt appears in a list of things that “got sold hard in 2024 and 2025” and do not hold up. One more only reports which sites publish the file: four of the twelve contractor sites in one study do.

Google had already said in December 2025 that no new AI text files are needed, so we take no credit for agreeing. The four that implied it helps are the part that matters.

All four are our own studies, which is the uncomfortable part, because studies are what other people quote.

Our broken-schema study described llms.txt, in a link, as “the second highest-impact AEO move after schema”. That was our own ranking of effort, it was never measured, and Google has stated that Search ignores the file. The line is gone.

Our AI-blocking study listed shipping an llms.txt file among the AEO work that makes a site a citable source. No engine documents reading the file for answers, so it is no longer on that list.

Our speed study told owners to ship llms.txt in the same rebuild that fixes their speed, to “get both wins”. The speed advice stands; the llms.txt half of it is gone.

The fourth is subtler. In our study of 50 ChatGPT queries about Calgary marketing agencies, we reported the traits shared by the nine firms that received the bulk of the citations, and one line reads “Every top-cited firm had: llms.txt at the root. 8 of 9.” The same post then told readers to ship llms.txt and predicted that firms that did would move into the citation set. With nine firms and no control, that inference is not available from our data, and no engine, OpenAI included, documents reading the file. We removed the advice and the prediction and added a dated note to the observation, rather than deleting it, because the observation is what we recorded.

The other rules fired on sentences in eleven posts. Read in full, none of those sentences advises what Google rejects: most state Google’s position or recommend earning genuine coverage. Two are worth describing, because they show what pattern matching gets wrong. Our post on law firm search suggests a page per practice area, which the checker caught under the fan-out rule. Reading it back, a practice area is a distinct service with distinct buyers, not a variation on a query, and Google’s objection is to generating pages for query permutations. False positive. The chunking rule also caught our post on web reading myths, which quotes eye-tracking research about big chunks of unformatted text for people, not machines. Another false positive, and worth saying so, because a checker that is never wrong is a checker that is not looking hard enough.

Yes, and our audit tool still gives it 2 points out of 66, which ties with FAQPage parity for the smallest weight of any check we run.

The reason is narrow. Google says it ignores the file, and that is now settled. No other engine documents using it either. OpenAI publishes almost nothing about how ChatGPT selects sources, and what it does publish is a line about multiple factors and a note that placement is not guaranteed. So the honest case for llms.txt is not that it works. It is that it costs almost nothing, it is a plain summary of what a site covers, and if a future system reads it you have one. That is a two-point bet, not a strategy, and anyone charging you for it as a strategy should be asked what evidence they have.

The stronger version of the same point: the peer-reviewed research on generative engine optimization, the 2024 paper that gave the field its name, tested content-level tactics like adding citations, quotations and statistics, and it ran its main experiment by simulating an engine over the top five Google results. Its own replication on a live engine was 200 queries with sources supplied as uploads. Useful as a starting point, mostly not reproduced by later benchmarks, and a long way from a file at your domain root.

Nothing about the work, quite a lot about the invoice. Google’s line is that optimizing for its generative features “is optimizing for the search experience, and thus still SEO”. If an AEO proposal in front of you is priced as a separate discipline and its deliverables are a file, some chunking and a schema pack, the guidance above is the answer to it.

Check the switch first. The Search Console control is the only item on this page that can take you to zero, and it has existed for every site since August 31. Look before buying anything.

Judge a vendor by what they will not claim. Google’s warning about tools claiming internal metrics is a usable test. Ask what the tool measures, where the number comes from, and what it would look like if the tactic did not work. We publish our audit’s thirteen checks and their weights for the same reason.

Keep doing the boring half. None of the guidance retracts the things that were always true: be indexed, be eligible for a snippet, say who wrote it, answer the question the page claims to answer. In our view, those basics matter far more than anything exotic.

Google’s positions are quoted from two pages on developers.google.com, “AI features and your website” (last updated December 10, 2025) and “Optimizing your website for generative AI features on Google Search” (announced May 15, 2026, last updated July 10, 2026), plus the Search Console help page for the Search generative AI control, all checked against the live pages in October 2026. Quotations are verbatim and the pages are linked above. Google updates these pages without changing their URLs, so a future reader should check the wording against the date stamps.

Meridian15’s own audit ran on October 15, 2026 over the 84 posts then live, on the body text and FAQ answers split into sentences. It is pattern matching, so it finds phrasing rather than meaning: it will miss an efficacy claim worded unusually, and its automatic hedge test is strict (a link or heading that names the file counts as unhedged), so every flagged post was read in full by a person and the verdicts are that reading. Outside llms.txt, every flag turned out on a full read not to be advice Google rejects, and the two clearest misfires are described above rather than hidden. The method, the rules and the results are published with this post.

One thing this post does not do is test whether llms.txt works for any engine. It reports what Google states, what other operators do and do not document, and what our own archive says. We know of no controlled experiment on the file, and we have not run one.

• llms.txt explained: what the file is, who reads it, and what it does not do.
• The state of AI search in 2026: what the evidence supported before this guidance landed.
• AEO vs SEO vs GEO: the distinction, now that Google has given its own answer.
• The free AEO audit: thirteen checks, sixty-six points, weights published.