llms.txt: what this file really does for your AI visibility
It is sold as the robots.txt of the AI era. Google has said plainly that it does not use it—and that is not a reason to delete it.
TL;DR llms.txt will not move you up the rankings: Google says it does not use it. Keep it as a verifiable fact sheet generated from your site, and fix your pages first.
In mid-September, while auditing my own site, I hit a problem in a file I thought was exemplary. My llms.txt, meant to give AI assistants a clean, reliable version of what I do, still listed city pages that no longer existed.
Four dead URLs, removed and redirected long ago, and eight live pages missing from the list. The document meant to stop machines from inventing things was serving stale information—with all the authority of an official file at the domain root, where nobody thinks to re-read it.
That episode says almost everything about llms.txt. It is not a visibility lever, and Google has confirmed it. It is a mirror of your site: useful when accurate, harmful when it drifts, and pointless when the pages it summarizes say nothing concrete.
Why llms.txt will not get you cited by answer engines
The idea is seductive. robots.txt tells crawlers where they may go; llms.txt, proposed in 2024, is a Markdown text file at the site root that summarizes for language models who you are and where to find the essentials. A business card for machines, in short.
Since then, the file has been sold as the missing piece of AEO—answer engine optimization. Add it, and ChatGPT, Perplexity, or Google will finally recommend you. The promise is simple, and it does not hold.
What Google wrote in June 2026
The clearest answer came from Google itself. In its guide to optimizing for generative AI features , updated in June 2026, it states that you do not need to create new machine-readable files, AI text files, or Markdown to appear in its results.
That applies to classic search as well as AI Overviews and AI Mode, which rely on the same ranking systems as everything else. An llms.txt does not penalize you, but it does not move you up a slot either. On the assistant side, I know of no public commitment that promises more, and nothing forces a bot to read this file instead of your pages.
In other words, a vendor billing you for an llms.txt as an AEO strategy is selling a line of code at the price of a lever. The file has a use—but not the one people claim.
1. Write verifiable facts, not sales copy
If the file is not a lever, why keep one? Because it remains the only place where you can state, on a single page, the facts an assistant is most likely to invent about you: your prices, lead times, languages, and service area.
Explicit rules against invention
Mine includes a line I consider the most useful in the whole file: do not invent discounts, subscriptions, or guarantees—followed by published rates and how they are billed. That is not marketing; it is a list of facts anyone can check against an answer.
Tone matters as much as content. A file full of adjectives—leader, innovative, must-have—gives a model nothing to quote. A legal name, an address, three services described in one sentence each, and a price grid do: those are facts that survive being pulled out of context for citation.
2. Generate the file from the same source as your site
The real risk of an llms.txt is not that it is ignored. It is that it is read when it no longer matches the site. Written by hand one day and forgotten the next, it drifts with every page added, renamed, or removed—with nothing to alert you.
The file that drifts in silence
That is exactly what happened to me: my German city pages had been reorganized, old URLs redirected, and the file kept citing the retired versions. It now lists 91 URLs, none dead, because it is checked on every audit the same way as the sitemap.
Ideally, the URL lists in the file are produced by the same code that generates your pages and sitemap. If not, build it into your routine: whenever a page changes address, the file changes with it—in the same action, on the same day.
That is the rule I apply for clients, who all have their own now. A photographer's site in Sierre got its file during the migration to Webflow, and the principle is the same everywhere: the file follows the site, never the reverse.
Add a last-updated date at the top of the file too. It is not only for machines—it reminds you the document has an age, and a file several months old on a site that has changed in the meantime deserves a re-read before anyone trusts it.
3. Put the effort on the pages engines actually read
What decides a citation is your page content and how it is structured. An answer engine does not see your design; it reads your HTML—heading hierarchy, sections, links, and data that say what is a service, a price, or an address.
Semantic HTML before any sidecar file
A title that is actually an <h1>, navigation inside <nav>, main content isolated in <main>: those tags are vocabulary a machine understands without effort. On my own site, an audit showed the primary title on 39 pages sat outside the main content zone, and the fix was a single line.
That is the work that makes a site citable by AI , far more than a file engines may ignore. Google also notes that structured data is not a gate to its AI answers: it helps you be understood and earn rich results; it does not replace a page that actually answers.
A good llms.txt is a symptom of a well-run site, not the cause
An accurate, up-to-date, factual file does not make a site visible. But a site that has one often reflects a wider discipline: stable URLs, published facts, structure built to be read. That discipline is what engines reward—not the file that summarizes it.
One source of truth for the whole site
Pages, sitemap, structured data, llms.txt: ideally, everything starts from the same content and updates together. That is how I build an SEO and AEO strategy : facts first, stored in one place, then every surface that exposes them—for humans and machines alike.
If you only have one hour for AI visibility this week, do not spend it on llms.txt. Spend it on your main service page: does it clearly state what you do, for whom, where, and at what price? The file can come later, and it will take ten minutes.
Ready to make your site a source AI can verify?
An llms.txt will never do your pages' job, and that is fine. Treat it as an honest ID card, generated from your site and re-read whenever something changes, and keep your energy on the content engines actually read. It is less spectacular than a miracle file, and it works.
👉 Want to know what AI assistants actually retain about your site? Get in touch