Let me tell you about a file that has generated more conference talks, LinkedIn carousels, and breathless agency emails than almost anything else this year. The file is called llms.txt, it lives in your site root next to robots.txt, and the promise attached to it is intoxicating: place this little markdown document on your server, and the great machines of artificial intelligence — ChatGPT, Gemini, Perplexity, Claude — will read it, understand your site more deeply, and cite you more often.
It is a beautiful idea. Clean, simple, technical, the kind of thing an SEO consultant can implement in twenty minutes and bill for an hour. And I understand completely why people want it to be true.
But the truth is, we have to be honest with each other. After more than a year of this file existing in the wild, the evidence has come in — and it is not flattering.
What llms.txt was supposed to do
The premise, when it first appeared, was reasonable enough. We already have robots.txt, which tells search crawlers where they may and may not go. So why not a parallel file — llms.txt — that hands large language models a tidy, curated summary of your most important pages, written in clean markdown they can parse without wading through your navigation menus, your cookie banners, your JavaScript?
A site owner would list their key URLs, add short descriptions, maybe a paragraph explaining what the business does, and the AI systems would, in theory, prefer this distilled version over crawling the messy live site. Less noise, more signal. Everyone wins.
Naturally, the SEO industry seized on it. Plugins appeared overnight. Generators sprang up. "Add an llms.txt file" became a line item in every GEO audit, mine included for a while, I will admit it.
And then the data arrived
Here is where the story turns, and it must be said plainly. The major AI companies — OpenAI, Google, Anthropic, Meta, Mistral — have not adopted it. Their crawlers do not request the file. Across one analysis of more than 515 million LLM bot traffic events, the share of requests touching /llms.txt was statistically negligible. The bots that actually matter for your AI visibility are, for the most part, simply not asking for it.
Think about what that means for a moment. You can write the most exquisite llms.txt file in the history of the format, place it lovingly in your root directory, and the very systems it is designed to feed will walk straight past it without knocking.
Google has been the most direct of all. Gary Illyes, speaking at Google Search Central Live, said the company does not support llms.txt and has no plans to. And John Mueller — whose word carries a certain weight in this world — went further. He compared the file to the old keywords meta tag.
If you have been doing this long enough, that comparison should make you wince. The keywords meta tag was the thing where you stuffed your target terms into the HTML head, and search engines, after a brief honeymoon in the 1990s, learned to ignore it entirely. Why? Because it is controlled by the site operator and therefore trivially gamed. Anything I can declare about myself, with no external verification, is something I will eventually exaggerate. Mueller's point is that llms.txt sits in exactly the same trap. It is a self-declaration. And self-declarations, in a world drowning in spam, are worth roughly nothing.
There was even a slightly comic episode where llms.txt files showed up on some Google properties, and the GEO crowd erupted — "See! Google uses it!" Mueller's response, when asked whether this represented an endorsement, was a single word. No. The files appeared because Google's internal content management system had quietly added support, and some teams simply had not bothered to remove them. An accident, not a signal.
But wait — doesn't somebody use it?
To be fair, and I want to be fair, the picture is not perfectly black. There are scattered reports of smaller AI tools and certain agentic frameworks fetching the file. Some developers swear they have seen their llms.txt pulled. A handful of documentation-heavy sites — software companies, mostly — find genuine value in it as a way to feed clean reference material to coding assistants.
So I will not tell you the file is harmful. It is not. It costs you nothing but the twenty minutes of writing it, and on a documentation site it may even help a developer's AI assistant find the right API page. That is a real, if narrow, use case.
What I am telling you is something different. I am telling you that llms.txt is not a GEO ranking signal. It is not the lever that gets you cited by ChatGPT or surfaced in Google's AI Overviews. Anyone selling it to you as that lever is either behind on the evidence or hoping you are.
So where should that twenty minutes go instead?
This is the part that actually matters, because the danger of the llms.txt hype is not the file itself — it is the opportunity cost. Every hour spent polishing a document the machines ignore is an hour not spent on the things that genuinely move AI citations.
And what moves them is not mysterious. AI systems cite content they can extract cleanly and trust. That means writing in a way where a single paragraph answers a single question completely, so a model can lift it whole. It means real expertise on the page — named authors, genuine experience, specific numbers instead of vague claims. It means your brand being mentioned, discussed, and linked across the wider web, because the models weight corroboration from sources they did not get from you.
The verità, the truth, is almost old-fashioned. The things that earn AI citations in 2026 are the same things that earned human trust in 2006: be clear, be credible, be useful, and be talked about by people who are not you. There is no markdown file that shortcuts any of that.
How the confusion took hold
It is worth pausing to ask why this particular file gained so much traction, because understanding that protects you from the next one. The appeal of llms.txt was never really technical — it was psychological, and it rode on a genuine anxiety. Business owners watched their Google traffic soften as AI Overviews and chatbots began answering questions that once sent people to websites. They felt the ground shifting, and they wanted something to do about it. Something concrete. Something they could implement on a Tuesday and feel, for a moment, that they had taken control back.
Into that anxiety arrived a file with a reassuring name and a tidy implementation. It looked official. It sat next to robots.txt, which everyone trusts and understands, and so it borrowed that file's credibility by association. And it offered the one thing frightened people most want — a sense of agency in the face of a confusing change. None of that, unfortunately, makes it work. A remedy that calms you is not the same as a remedy that cures you, and the gap between those two things is precisely where a great deal of wasted effort lives in this industry.
There is also a quieter dynamic. A whole layer of tools and consultancies had a commercial reason to promote llms.txt, because it was easy to sell — a clean deliverable, simple to explain, simple to implement, simple to invoice. The genuine work of earning AI visibility is messy and slow and hard to package neatly. A file is a product. Authority is not. And so the file got marketed with an enthusiasm its evidence never earned. I do not say this cynically; I understand the incentive. But you, on the receiving end, should know the incentive exists.
The pattern underneath all of this
If you zoom out, llms.txt is the latest entry in a very long catalogue. Every few years our industry falls in love with a technical artifact that promises to replace the hard work of being genuinely good. The keywords meta tag. The "submit your site to 500 directories" era. The schema-stuffing phase. Each time, the appeal is the same: a discrete, controllable file or action that we can complete and check off, instead of the uncomfortable, never-finished work of earning authority.
I have a certain affection for these episodes, honestly. They reveal something human. We would all prefer a clean technical fix to a messy human one. But search engines, and now AI engines, are built precisely to see past the things we declare about ourselves and toward the things others confirm about us.
So here is my honest recommendation, the one I would give a friend. If you run a documentation-heavy site and you have a spare twenty minutes, sure, add an llms.txt — it might quietly help a developer's assistant, and it does no harm. For everyone else? Skip it without guilt. Take that time and put it into one genuinely excellent, citable page instead. The machines that matter are not reading your llms.txt. But they are reading your content. Give them something there worth quoting.
If you want to know whether your actual content is the kind AI systems will cite — not whether you have the right file in your root directory — that is the question worth answering. It is also exactly the question Licheo was built to help you answer. Contact us, and we will show you where you genuinely stand.