There is a lot of confident writing about llms.txt by people who have never shipped one, and a lot of dismissal by people who have never tried. We generate one for six sites from live data. Here is what it is, what we put in ours, and the honest answer on whether it does anything.
Take everything you would say to a journalist in thirty seconds — who you are, what you cover, which pieces
are your best work, how you want to be credited — and put it in a text file at
/llms.txt. That is the whole idea. The format is Markdown, the audience is a language model, and the
premise is that an assistant summarising your site does better with a curated map than with whatever it scraped.
It is explicitly not a permissions file. robots.txt says what may be crawled; llms.txt says what is worth reading. Publishing one grants nothing and blocks nothing.
-full variant is the territory: the whole catalog in plain text for an assistant that wants the data rather than the highlights. We publish both — ours has the {len(DATA) if False else MCP['_checked']} tools grouped by category.Ours is generated from the same database that drives the site, so it can't drift out of date. The sections, and why each is there:
Our illustration of the three routes we publish. They differ in effort, not in honesty — all three carry the same attribution and disclosure.
The honest answer is that we cannot prove it, and neither can anyone else. No major assistant publishes llms.txt as a retrieval or ranking input. Anyone showing you a traffic chart and attributing it to a text file is selling something.
What we can say is narrower and, we think, more useful. Writing ours forced a decision we had been avoiding: what do we actually claim, and which of our pages would we be comfortable seeing quoted back at us? That turned out to be a short list. It also surfaced an inconsistency — we were describing far more reviews as "hands-on" than Daan had genuinely tested, and fixing that in the file meant fixing it across the site.
So: publish one, spend an hour, and treat any traffic effect as a bonus. The reason to do it is that it makes you write down what you stand behind. If your answer to "which pages would I want cited?" is "all of them", the file isn't the problem.
You can read ours and take whatever is useful.
A plain-text file at the root of your site that tells AI assistants what the site is, which pages are worth citing and how to attribute you. It is a proposed convention, not a standard anyone enforces — think robots.txt in spirit, but written for reading rather than crawling permission.
Honestly: nobody outside the AI labs can prove it. No major assistant documents llms.txt as a ranking or retrieval input. What we can say is that it costs almost nothing to publish, it forces you to write down what you would want cited, and that exercise improved our own pages. We would not spend a week on it.
No, and the difference matters. llms.txt is passive: a file an assistant may read if it happens to fetch it. An MCP server is active: the assistant connects and asks questions, and you answer in real time. One is a note on the door; the other is a doorbell.
The claims you are willing to stand behind, with links. Ours lists our best-of guides, the reviews Daan actually tested hands-on (kept separate from the researched ones), the machine-readable catalogs and the attribution line. What should not go in it: everything. A file that lists every page is a sitemap, and you already have one.
Written from running llms.txt on six sites since 2026. We make no claim that it drives traffic, because we cannot measure that honestly. Some links elsewhere on this site are affiliate links; there are none in this article. How we review →
New dossiers, cost-traps we found, and what actually changed in AI tooling — no hype, no sponsored-disguised-as-advice. Unsubscribe anytime.