A founder asked me last month whether he had “done GEO” yet. His evidence was a file. He had added an llms.txt to his site, the way a blog post told him to, and he wanted to know if that was the work or just the start of it. It was a fair question, and the honest answer is one almost nobody selling llms.txt setups will give you. The file is fine. It is also doing far less than he thought, and probably less than you think too.
If you have not run into it yet, llms.txt is a proposed standard: a plain text file you put at the root of your site, like robots.txt or a sitemap, that lists your most important pages and describes what your company is, in language meant for an AI model rather than a search crawler. The pitch is clean and appealing. Give the engines a tidy, curated map of you, and they will understand you better and cite you more. I add one to every site I run. I just do not pretend it is the lever.
What the file is supposed to do versus what engines actually do with it
The theory is that an AI engine, when it wants to answer a question about your company, will look for your llms.txt, read your curated summary, and prefer it. In practice, the load-bearing word in that sentence is “will,” and it is doing a lot of unearned work.
As of now, none of the major AI engines have publicly confirmed that they read llms.txt when answering a user’s question. Not ChatGPT, not Perplexity, not Google’s AI answers. The standard exists, plenty of sites have adopted it, and the engines have mostly stayed quiet about whether they consume it. That silence matters. It means adding the file is a bet on future behavior, not a documented input to current behavior.
This is the same distinction that runs underneath how AI engines actually decide what to cite. The things we can observe moving citations are retrieval of live pages, clear and verifiable claims, and what trusted third-party sources say about you. A self-authored summary file sitting at your root is, at best, one more place you describe yourself, and self-description is the weakest evidence on the web that anything about you is true.
Why “I added llms.txt” is the new “I submitted my sitemap”
There is a familiar shape to this. In old-school SEO, people would submit a sitemap to Google and feel like they had done SEO. The sitemap helped Google discover pages. It did nothing to make those pages rank. The two got confused constantly because the sitemap was concrete and easy, and the actual work was diffuse and hard.
llms.txt is in that exact spot for AI search. It is concrete, it takes an afternoon, and it produces a file you can point at. So it gets mistaken for the work. But the file does not change what third parties say about you, it does not correct a wrong label, it does not earn you a citation on a buying-intent query. Those are the things that move AI visibility, and none of them live in a text file you wrote about yourself.
The risk is not that llms.txt hurts you. It almost certainly does not. The risk is that it gives you a false sense of completion, so you stop at the easy thing and skip the hard thing that would have actually worked.
The narrow, real value (so add it, just for the right reasons)
I keep recommending llms.txt despite all of the above, because it does have value. It is just not the value in the sales pitch.
It is cheap insurance on a real possibility. If the engines do start consuming llms.txt as a ranking or grounding input, and that is a plausible direction, the sites that already have a clean one are ready on day one with no scramble. The cost of being early is one afternoon. The cost of being late could be a quarter.
It forces a useful exercise. Writing a good llms.txt makes you state, in plain language, what category you are in, who you are for, and which pages are the canonical answers about you. Most companies have never written that down cleanly. Doing it well surfaces exactly the confusion an engine would also have about you, which is genuinely worth knowing.
It is a clarity signal that costs nothing. Even if no engine reads the file as a special input today, a tidy, accurate, well-structured statement of what you are is the kind of content that helps when it does get read, anywhere. Clarity is the one thing that helps you in AI search no matter which mechanism turns out to matter.
What it does not do is substitute for fixing the sources. If AI search is describing you wrong, the fix is a source-layer operation, the patient work in how to fix the way AI search describes your brand, not a file you control. A label the engine learned from other people’s pages will not be overwritten by a summary you wrote about yourself.
What I got wrong
The first time I took llms.txt seriously, I over-invested in it. I built an elaborate one for a site, carefully ordered, every important page summarized, the positioning tuned, and I half-expected it to move citations on its own. Then I measured. For weeks I tracked the same buyer queries across engines, the way measuring AI search visibility requires, and watched for the file to do something.
It did nothing I could detect. The citations that did come in over that period traced cleanly to other causes: a third-party page that mentioned the product, and a clear, liftable proof section on the site itself. Not the file. The engines that improved their answers did so on their own recrawl and refresh schedule, the one I described in how long it takes to get cited, and they would have done it with or without my llms.txt.
So I changed where it sits in the order of work. Now I ship llms.txt as hygiene, near the end, and I never describe it to a client as the thing that will get them cited. It is the tidy last step, not the first move, and definitely not the whole program.
Why this matters
The honest position on llms.txt is not “skip it.” It is “add it, and do not mistake it for the work.” It is a sound, low-cost bet on where the standards may go, and a useful forcing function for stating clearly what you are. It is not a citation lever, because the levers that actually move AI search are about what others say about you and how verifiable your own claims are, none of which a self-authored file can fake.
If a vendor is selling you an llms.txt setup as your AI search strategy, that is a useful tell. It means they have optimized for the part that is easy to deliver and easy to show, and skipped the part that is hard and slow and actually changes the answers buyers see.
If you want to know what is genuinely moving the answers about your company, and what is just tidy housekeeping, that separation is the first thing a paid audit makes clear, and the full method is on the methodology page.