The question people ask is always the same, and it deserves a direct answer before anything else: no, there is no evidence it works.
llms.txt started as a community proposal in 2024, modelled on robots.txt and sitemap.xml: Jeremy Howard published it on September 3, 2024. The idea is reasonable: hand a model a clean summary of the site instead of making it reconstruct one from the HTML. But a reasonable idea and an adopted standard are two different things, and right now this is the first. For scale: robots.txt dates from 1994 and the IETF only formalised it in 2022, as RFC 9309.
What goes inside an llms.txt file?
A Markdown text file served at `yoursite.com/llms.txt`. There is not much in it:
- NOIA title — the company name
- NOIOne paragraph — what you do, in two or three lines, no adjectives
- NOIA list of links with one line of description each — the pages that matter
- AIAn optional section for what is secondary
That is all. No markup, no schema, no validation. If you can write a bulleted list you can write an llms.txt. Under the proposal the title is the only required part; the "Optional" section is a convention for material a model can skip.
Which pages belong in llms.txt?
This is the only part you can get wrong, and also the only part where the file becomes useful.
The temptation is to include everything. That is the mistake: if the file lists forty pages it is not summarising anything — it is duplicating the sitemap, which already exists and does that job better.
If your llms.txt lists everything, you are not summarising. You are duplicating the sitemap.
The rule we use: only what you would say out loud to a client in the first two minutes. Who you are, what you do, three or four pages that answer the real questions. Ten links is already the ceiling; ours only reaches it because every guide exists in two languages.
Ours, at length
# As Growth
> As Growth is the first Italian Growth Partner in Vibe Marketing: AI does the
> heavy lifting (audits, production, data), people do strategy, copy and
> control. Founders: Alessandro Iannuzzi and Sara De Vita.
## Pages
- [Home](https://www.asgrowth.io/): method, case studies, team, FAQ
- [Diagnosis](https://www.asgrowth.io/diagnosi): 7 questions, answer within 48h
- [Blog](https://www.asgrowth.io/blog): guides in Italian
- [English blog](https://www.asgrowth.io/en/blog): the same guides in English
That is the opening of the real file, not a simplified example. Below it sit two more sections: the main guides — each with the Italian and English link on the same line, because the site is bilingual and otherwise a model sees the same content twice without realising they are the same thing — and the contact details.
Even so it stays under thirty lines, and that is as far as it makes sense to go. If yours grows past that, the question is not how to organise it better: it is what to take out.
Is llms.txt worth adding to your site?
One thing first: the companies building the models use it. Anthropic and OpenAI both publish one for their developer docs. Publishing one is not the same as reading yours, and for its AI Overviews Google says you do not need AI text files.
We keep ours for three reasons, in decreasing order of honesty.
It costs half an hour and does no harm. It is a static file: it does not touch rendering and does not affect traditional SEO in any way. The ratio between cost and possible benefit justifies half an hour even at low odds.
Writing it is a useful exercise. Forcing yourself to explain what you do in two lines and to pick four pages out of eighty surfaces things about the site. A couple of times it made us realise an important page did not exist.
If it were adopted, arriving late costs the same. There is no first-mover advantage in a static file. That is the weakest reason, and we say it for completeness.
What to do first
Here is the actual point of this piece, and the one recommendation we will make without reservation.
If you are considering llms.txt because you want to be cited by AI, the file is the last thing to do and the least effective. The things that work are elsewhere: content that answers a question in an extractable way, correct structured data, and, plainly, a site the crawlers can read.
How you become citable is in the GEO guide; how to write to be the answer rather than a result is in the AEO one.
Write your llms.txt on a Friday afternoon, once the rest is done. In that order it makes sense. The other way round, it is time spent to feel productive.
Frequently asked questions
What is the llms.txt file?
It is a Markdown text file at the root of a site that summarises in one page what the company does and which pages matter, meant to be read by a language model. It began as a community proposal in 2024, modelled on robots.txt: same idea, readable instructions for whoever passes through.
Do AI engines actually read llms.txt?
There is no evidence they do. None of the major engines has publicly confirmed using it, and no independent data shows an effect. Anyone telling you otherwise with confidence is going beyond what is known: it is a community proposal, not an adopted standard.
Where does the llms.txt file go?
At the root of the domain, reachable at yoursite.com/llms.txt and served as plain text. It does not need declaring anywhere and does not conflict with robots.txt or the sitemap, which remain the files crawlers genuinely read.
What is the difference between llms.txt and robots.txt?
robots.txt tells crawlers what they may visit and has been a respected standard for twenty years. llms.txt tries to tell a model what is important on the site, with no confirmed adoption. The first controls access, the second tries to steer comprehension.
How many pages should go in an llms.txt?
Few, ten links at most. List forty and you are not summarising, you are duplicating the sitemap, which already exists and does that better. The practical rule is to include only what you would say out loud to a client in the first two minutes.
Is it worth doing?
Yes, but last. It costs half an hour, does no harm, and writing it forces you to clarify what you do and which pages matter. If you are doing it to get cited by AI, though, extractable content and structured data come first and count for far more.
Want to know whether AI mentions you?
We run the check across 30 real questions in your sector and tell you where you stand against your competitors. Seven questions, founders answer within 48 hours.
ApplyKeep reading
Founder of As Growth, an AI-first growth partner. He runs growth and strategy across partner businesses: OroEtic, Giojello, Equity Plus and others. Apply if you want to work with us.