Write an llms.txt a model can actually use
Optional — everything below can be filled in by hand. We read up to 50 pages, sort them into sections and fill in the fields. Only the URL goes to our server; everything you type stays in the browser.
The business, not the tagline. This becomes the H1.
What you do and who for. This is the line a model reads to decide what this site is.
Optional. Hours, service area, what you do not do. Plain sentences — no headings.
Paste a list of pages instead
Here is a finished one, for a practice that does not exist. Read a URL above or start typing on the left and it becomes yours — the file builds as you go.
# North Austin Dentistry
> A family dental practice in North Austin, taking new patients, open Saturdays.
We serve North Austin, Cedar Park and Round Rock. We do not take Medicaid.
## Services
- [Invisalign](https://example.com/services/invisalign): What it costs, how long it takes, and who it suits
- [Dental implants](https://example.com/services/implants): Single tooth and full arch, with recovery times
- [Emergency dentistry](https://example.com/services/emergency): Same-day appointments, what counts as an emergency
## Answers
- [What we charge](https://example.com/fees): Prices for common treatments, and the payment plans we offer
- [Insurance we take](https://example.com/insurance): The plans we are in network for, and what we do not accept
- [New patient FAQ](https://example.com/faq): First visit, parking, what to bring, how long it takes
## About
- [The practice](https://example.com/about): Who works here, how long we have been in North Austin
- [Book an appointment](https://example.com/contact): Hours, address, phone, and online booking
Upload it so it answers at /llms.txt, served as text/plain. Since v2 of the spec a subpath counts too — /docs/llms.txt covers the pages under it. Then check the live file — a file that is served as HTML, or from the wrong path, is a file nothing reads.
What to put in it, and what to leave out
The temptation is to list everything. Resist it. The file is read by something that has to decide, in one pass, what this site is for — and a hundred links tell it less than ten to twenty do.
Three rules that hold up:
- The summary does most of the work. One sentence that names what you do, who you do it for, and where. “A family dental practice in North Austin, taking new patients” beats a paragraph of positioning.
- Link the pages that answer a question, not the pages that exist. A pricing page, an FAQ, a comparison, a service page with real detail. Your privacy policy is not on this list.
- Describe each link. The text after the colon is free context and most files skip it. “Invisalign: cost, timeline, and who it suits” is worth more than the URL on its own.
Why the draft is shorter than your site
Reading a site takes a minute and gets you a file. It also gets you the temptation this whole format exists to resist: we found forty pages, so here are forty links. So the draft stops at eighteen, six to a section, and everything else sits under a list you can add from one at a time. The blog, the homepage and anything the classifier could not place start out of the file rather than in it.
The two fields worth your attention afterwards are the summary and the descriptions. A crawl can only copy what a page already says about itself — and most meta descriptions were written for a search result, not to tell a model what the page answers.
It is the smallest item on most reports
We should say this plainly, since we also sell the thing that writes these for you: an llms.txt is rarely the reason an engine names somebody else. When we run the buyer questions against a site and it loses, it usually loses because there is no page answering the question at all — not because the index pointing at the missing page was formatted wrong.
Publish it anyway. It takes fifteen minutes and it is one of the few things on the list that is entirely within your control. Then find out what is actually costing you.
Questions
See who the engines recommend instead
12 buyer questions across ChatGPT, Perplexity and Gemini. You get the questions your client loses, who won them, and what is missing from the site. No score, no signup wall.