A year ago, llms.txt was the file that was going to get your brand cited by AI. Drop it in your root domain, the pitch went, and ChatGPT, Gemini, and Perplexity would finally read your site the way you wanted. New data says the file is mostly sitting there, untouched.
Ahrefs looked at 137,000 domains. In May 2026, 97 percent of the llms.txt files they found received zero requests. Not a trickle. Zero.
The Numbers Do Not Support the Hype
Of roughly 38,000 domains that had a valid llms.txt file, only about 1,100 saw any traffic to it at all. And here is the part that should end the argument. AI retrieval bots, the actual systems that build answers in ChatGPT or Gemini, accounted for just 1 percent of the requests that did land.
So who was reading these files? SEO audit tools at 21 percent. Unidentified bots at 14 percent. Generic web crawlers at 13 percent. The file is being checked by the tools that check files, not by the AI engines it was built to feed.
Google's John Mueller put it plainly. The file is "not done for search." He called it a "temporary crutch, perhaps to save some tokens" for AI coding assistants. That is a long way from the promise that llms.txt was your ticket into the answer box.
I am not writing this to mock anyone who added the file. It costs ten minutes and breaks nothing. The problem is the story that got attached to it, the idea that a single text file is the lever that moves AI visibility. That story sold a lot of audits. The cold data just retired it.
Why This Keeps Happening
Every shift in search produces a new artifact that promises to be the shortcut. Meta keywords had their moment. So did a dozen schema tags that were going to guarantee rich results. llms.txt is the latest entry in a long line of files we wanted to believe in because they are easy.
The discomfort is that AI visibility does not have a ten-minute fix. From building GEOflux, I watch which brands actually get pulled into AI answers, and the pattern is boring in the best way. They get cited because the open web already treats them as a credible source. Real mentions, consistent positioning, specific and verifiable claims on pages that other people link to and quote. The model is summarizing a reputation it found, not obeying a config file you left at the door.
There is a control instinct underneath the file too. Something you place on your own server feels like a thing you command, and AI visibility is uncomfortable precisely because it is not under your command. It is decided out in the open, by what other people say about you and what the model already trusts. A text file is the illusion of a steering wheel on a car someone else is driving.
I made this case before, that AI search will not cite you just because you asked it to, and the llms.txt data is the cleanest proof yet. The engine is not looking for your instructions. It is looking for evidence that you matter.
There is a quieter signal in the same data. The one group genuinely using llms.txt is AI coding tools, pulling it to save tokens while they work. That is a real use case. It is also a completely different job from getting your brand named in a buyer's answer. The file found a purpose. It just was not the purpose it was marketed on.
What to Do With the Ten Minutes
Keep the llms.txt file if you already have one. It is harmless, and the standard could still mature. Just stop counting it as an AI visibility strategy, because the request logs say it is not one.
Spend the real effort somewhere measurable. Pick the ten prompts a buyer would actually type before choosing your category, and read what ChatGPT and Gemini say back. If your brand is missing or described wrong, that absence is your true ranking, and no file fixes it. That is a baseline you can act on.
Then work the inputs that move the answer. Publish claims specific enough to quote. Earn mentions on sources the model already reads. Make your positioning consistent enough that an engine summarizing the web keeps landing on the same description of you. When nine out of ten brands are still invisible inside AI answers, the gap is not a missing text file. It is a missing presence in the sources the model trusts.
If you want one concrete move that beats the file, look at where your category gets discussed and make sure your strongest, most specific claim is present there in language a model can lift. A clear statistic, a named result, a definition you own. Pages that state something quotable get quoted. Pages that hedge get skipped, file or no file.
The llms.txt episode is a useful tell. When a fix sounds too simple for a problem this big, the request logs usually agree. Visibility in AI is earned in the open, not declared at your front door.