The llms.txt myth

There is a new line item working its way through agency proposals right now. It is called llms.txt: one markdown file at the root of your website, sold as the switch that gets your brand into ChatGPT's answers. The pitch writes itself. robots.txt told search engines where to crawl; llms.txt, the story goes, tells the AI what to say about you. Setup fees. Monthly "AI-readiness" retainers. Guaranteed inclusion. We are a studio that sells AI visibility, and we are telling you not to buy this.

This is a reality check, and we will keep it honest at both ends. The file itself is harmless, and even mildly useful. We will say exactly what it is good for. But the claim being invoiced, that AI systems read llms.txt and rank you for it, is contradicted by the only two kinds of witnesses that count: the people who operate the machines, and the server logs that record what the machines actually fetch.

Section OneA quarter of the web left a note on the door

The idea is seductive because it is tidy. A proposed convention says: put a clean, curated markdown summary of your site at /llms.txt, so that language models read your version of the facts instead of piecing one together. One file, total control of the narrative. Who wouldn't want that?

The convention borrows its credibility from its ancestors, and that is exactly where the borrowed part shows. robots.txt and the XML sitemap earned their place because the crawlers' operators committed to reading them, documented that commitment, and confirmed the behavior in everyone's server logs for two decades. llms.txt has the syntax of a standard and none of the substance of one: no operator commitment, no documented consumer, no confirmed fetch behavior. It is a standard-shaped proposal, waiting for a machine to adopt it. So far, none has said it will.

And it spread anyway. When Ahrefs examined 137,210 domains for a study published in June 2026, 28 percent had an llms.txt file in place. More than a quarter of the sample had left a note on the door for the machines.

Ahrefs, L. Linehan & X. Guan, "We Analyzed 137K Sites: 97% of llms.txt Files Never Get Read," June 15, 2026.

Adoption, however, measures belief. It does not measure whether anyone, or anything, reads the note. For that, you have to ask two questions: what do the operators of the AI systems say, and what do the access logs show? Both answers are now on the record.

Section TwoThe people who run the machines say no

Start with the most direct statement available. In June 2025, Google's John Mueller wrote, in public: "FWIW no AI system currently uses llms.txt." Not "Google doesn't." No AI system.

John Mueller (Google), Bluesky post via @johnmu.com, June 17, 2025.

His colleague Gary Illyes went further, drawing the parallel to the long-dead keywords meta tag: the 1990s relic that let websites declare their own relevance until search engines concluded that self-declared relevance is worthless, and stopped reading it. "It's very easy to draw a parallel between 1990's keywords meta tag and this," Illyes wrote. Whatever machines end up trusting, it will not be a self-authored file of claims about yourself.

Gary Illyes (Google), Bluesky, August 2025, via Search Engine Roundtable; Search Engine Journal, "Google Says LLMs.txt Comparable To Keywords Meta Tag."

In July 2025, Illyes said it without hedging on Google's own Search Central Deep Dive podcast: Google doesn't support llms.txt and isn't planning to. And Google's official documentation on AI features now states plainly that you don't need to create new machine-readable files, AI text files, or markup to appear in its AI features. There is no special file the machine is waiting for.

Google Search Central Deep Dive podcast, Gary Illyes, July 2025; Google Search Central, "AI features and your website," guidance last updated December 10, 2025.

What about the other AI vendors? Here honesty requires precision: as of August 2026, no major AI vendor has made a verifiable public commitment to reading other sites' llms.txt files. Anthropic publishes one for its own documentation. But publishing a file is evidence of writing, not of reading. No operator statement, anywhere, confirms consumption.

Note, too, what none of these statements say. Nobody said "llms.txt hurts you." Nobody said the machines punish it. The file is not dangerous; it is inert. Keep that distinction. It matters for the honest part of this story, further down.

A ranking factor that no ranker reads is not a ranking factor. It is a ritual.

Section ThreeThe server logs agree

Operator statements can be spin. Server logs cannot. The Ahrefs study checked what actually happened to those 137,210 domains' files, and the result is the bluntest number in this story: in the month of May 2026, 97 percent of published llms.txt files received zero requests. Not zero requests from AI bots. Zero requests from anything.

Ahrefs, L. Linehan & X. Guan, "We Analyzed 137K Sites: 97% of llms.txt Files Never Get Read," June 15, 2026.

The 3 percent that did get fetched tell an even better joke. Of their requests, only 19.5 percent came from named AI tools (GPTBot first, Claude Code second). More of the traffic came from SEO-audit crawlers and generic bots than from any AI bot. Read that again: the busiest readers of llms.txt are the SEO tools checking whether you have one.

A second study, run on entirely different data, lands in the same place. The analytics firm limy.ai logged 515,382,577 LLM-bot events across its measured sites over a 90-day window. Of that half-billion machine visits, exactly 408 targeted an llms.txt file. Roughly one in 1.3 million.

limy.ai, LLM crawler log analysis: 515,382,577 bot events over 90 days, 2026.

Section FourWhy the myth sells anyway

If the evidence is this one-sided, why is llms.txt still being invoiced? Because it is a perfect product, even though it is a useless lever. A file is visible. A file is billable. A file can be delivered by Friday and screenshotted for the client report. Actual visibility work means rewriting pages until they state plainly what you do, fixing the contradictions between your website and everywhere else the machines look, and earning mentions that corroborate your claims. That work is slow, and it resists the screenshot.

If you are holding such a proposal right now, two questions dissolve it. First: which AI system reads this file? The honest answer, on the record above, is: none confirmed. The burden of proof sits with the person invoicing you. Second: what will you measure to show it worked? A file's existence can be screenshotted; a citation in an AI answer has to be earned by the pages the machines actually do read. A vendor with real work to sell survives both questions. A ritual does not.

We have skin in this game, in the uncomfortable direction. Our business would be easier if the myth were true: a one-file fix we could sell at scale. It is precisely because we measure AI visibility that we cannot: nothing in the logs, ours or anyone's, shows the file doing what the invoice says it does.

Section FiveWhat the file is actually good for

Now the honest counterweight, because the file itself deserves better than its salesmen. Writing an llms.txt forces a discipline most businesses have never performed: putting what you do, for whom, at what price, in one page of plain language, and keeping it current. That inventory of your own facts has real housekeeping value. Most companies cannot produce it, which is usually the actual visibility problem.

If you do write one, treat it as the one-page inventory of your verifiable facts: legal name, what you sell, for whom, where, at what price, and what makes you provably different. Then do the part with actual leverage: make your public pages (the ones the machines demonstrably crawl) say the same things, in the same plain words, with no contradictions. The inventory is the rehearsal. The pages are the performance.

So our advice is not "never touch it." Our advice is: write it yourself, in an hour, for free. It costs nothing to publish, it harms nothing if ignored, and if some future system ever does read it, you are already there. What it will not do, on today's verified evidence, is make any AI system see you, cite you, or rank you. The file is fine. The invoice is the problem.

The lie was never the file. The lie is "the AI reads it and ranks you for it."

In ClosingThe boring work is the lever

What actually moves AI visibility is everything the llms.txt pitch lets you skip. Pages that state plainly who you are and what you sell. The same facts, consistent, everywhere the machines can look. Independent mentions that corroborate what you claim about yourself. Machines can only act on what they can clearly read. That is true of the models deciding which brands to recommend, and, as we keep documenting, of the agents now working inside companies too. There is no file-shaped shortcut around it. There is only the clarity work, and the good news is that it is entirely doable.

If you want to know where you actually stand, you can measure how visible your brand is to the machines at /signal-index/, or write to us at /contact/.