What is an llms.txt file?
llms.txt is a convention proposed in 2024 (llmstxt.org): a Markdown file at the root of a site that says in a few lines who you are and which pages matter, for AI assistants and agents that read the web. A title, a description in a blockquote, then sections of annotated links.
Emerging convention, adoption not guaranteed: no major AI assistant has announced reading llms.txt systematically, and its effect on your citations cannot be measured today. It costs ten minutes, cannot hurt, and documents your site cleanly; that is why we recommend it, not as a magic formula.
How the draft is built (v1, September 2026)
Every step is deterministic and reads only what your pages already publish:
- Reading robots.txt: location of your sitemap, and respect of the rules set for our robot (SEOForge-GEO-Test).
- Reading the sitemap (200 URLs at most; for an index, the first 2 child sitemaps). Without a sitemap: the links of the homepage (20 at most).
- Selection of at most 10 pages by the shape of their URL: same site, HTML pages, no parameters, no pagination or utility pages (cart, login, legal notice), closest to the root first.
- One request per page to read its <title> tag and meta description. The body of the pages is neither analysed nor stored.
- Assembly in the llmstxt.org format: "# Site name" (og:site_name or the homepage <title>), "> description" (homepage meta description, if any), one "##" section per first URL segment. Total budget: 20 seconds.
What the generator never does
- Invent a description: a page without a meta description is listed with its title only.
- Summarise or rephrase your pages: titles and descriptions are copied as-is.
- Override your robots.txt or read non-public pages.
- Keep the content of your pages: only the tested address, the date and counters are recorded.
Limits
- Selection by URL shape is a reasonable bet, not a reading of your strategy: the pages that matter to you are not necessarily the shortest.
- A site rendered only with JavaScript may expose neither title nor description in its HTML: the draft will be poor, and that is a signal in itself (see the GEO test).
- The file is not published for you: you place it at the root of your site.
- The llms-full.txt format (full content) is not generated: it calls for writing, not extraction.
This draft vs a strategic llms.txt
The generator extracts; the GEO Audit decides. The difference fits in three words: curation, priorities, disambiguation.
| Generated draft | GEO Audit - €490 excl. VAT | |
|---|---|---|
| Pages listed | Up to 10 pages chosen by URL shape | The pages that carry your offer, chosen with you, in the order an assistant should read them |
| Descriptions | Existing meta descriptions, copied | One answer-first line per page, written to be cited |
| Identity | Site name as published | Disambiguation sentence (namesakes, sector, area), consistent with your structured data |
| Consistency | Standalone file | llms.txt, robots.txt, Organization schema and pages aligned; llms-full.txt where relevant |
| Follow-up | Snapshot at generation time | File generated from your routes or content, updated with the site |
Frequently asked questions
Do AI assistants really read llms.txt?
Not in a guaranteed way. llms.txt is a convention proposed in 2024 by Jeremy Howard (llmstxt.org), adopted by tools and documentation sites, but no major assistant (ChatGPT, Claude, Perplexity, Gemini) has announced reading it systematically. It replaces neither robots.txt nor the sitemap. We recommend it because it costs little, cannot hurt and documents your site; not because it guarantees citations.
Where do the draft's descriptions come from?
Only from your pages: the <title> tag and the meta description of each, copied as-is. A page without a meta description is listed with its title only. The generator does not summarise, rephrase or invent anything.
Why are some of my pages missing?
The draft lists at most ten pages besides the homepage, chosen by URL shape (closest to the root first), after excluding documents, pagination and utility pages. Pages forbidden to our robot by your robots.txt are respected. Add the missing ones by hand: that is exactly what a draft is for.
Where and how do I publish the file?
At the root of your site, at /llms.txt, served as plain text (Content-Type text/plain, UTF-8), reachable without a redirect and not blocked in robots.txt. On WordPress, a file dropped at the root or a static-files plugin is enough; on a custom site, a dedicated route lets you generate it from your pages and never let it go stale.
What should I add to the draft to make it truly useful?
Three things extraction cannot guess: a disambiguation sentence after the title (who you are, what you are not, your area), a priority order that reflects your offer rather than your site tree, and descriptions written to answer a question rather than for Google. Also remove what does not help a hurried reader.
What do you keep from a generation?
The URL entered and the final URL, the date, the number of pages found, listed and described, the source used (sitemap or links), the warnings, the duration and, for 30 days, a truncated, irreversible fingerprint of your IP address to detect abuse. Neither the generated file nor the content of your pages is stored; the draft is rebuilt on demand for one hour (cache). This data is deleted after 24 months. If you request the file by e-mail, the generation is attached to your request.
An llms.txt is useless if AI crawlers cannot read your site: check it with the GEO test.