The essentials
- In an hour: a useful llms.txt for a small business site takes an hour to write, is published at the root and is updated with every significant publication.
- Contents: a title, a one-sentence summary, a description of the business, then sections of annotated links to 20 to 200 reference pages.
- Expectations: no demonstrated effect on citations in ChatGPT or the AI Overviews in 2026; the value lies in the clarity, the preparation and the agents already reading it.
- Prerequisite: allow the AI crawlers in robots.txt, without which the file serves no purpose.
The llms.txt introduces a site to AI assistants and agents in a condensed, readable form: a Markdown document placed at the domain root, saying what the company does and which pages to go to. Its definition and status are covered in llms.txt: definition ; this article describes the practical implementation, from selecting the pages to the final test.
Step 1: check the AI crawlers are allowed
A welcoming llms.txt means nothing if the robots.txt file forbids GPTBot, ClaudeBot or PerplexityBot. That is the first check, and it is often where the real visibility problem lies: many sites inherited blocking rules put in place in 2023 out of caution. Open the votre-domaine.fr/robots.txt address and confirm there is no ban on those agents. The crawlers concerned are detailed in indexing by AI.
Step 2: select the pages
| Category | Pages to include | Indicative number |
|---|---|---|
| Identity | About, team, legal notices, contact | 3 to 5 |
| Offers | One page per service or product, the pricing page | 5 to 15 |
| Proof | Case studies, reviews, references, certifications | 3 to 10 |
| Reference content | Pillar guides, definitions, comparisons | 10 to 150 |
| Optional section | News, blog, secondary content | Everything else |
The rule is to select, not to list everything: a file of two thousand links with no comments adds nothing to a sitemap. Each link gets a one-line description saying what the page contains, in the words a user would use.
Step 3: write the file
The expected structure is a Markdown H1 title, a summary as a blockquote, free paragraphs, then H2 sections containing lists of links. Here is the outline:
# Nom de l’entreprise
> Une phrase qui dit ce que fait l’entreprise, pour qui et où.
Deux ou trois paragraphes : activités, publics, spécialités, zone
d’intervention, date de création, éléments différenciants vérifiables.
## Offres
- [Nom de l’offre](https://www.exemple.fr/offre/) : ce qu’elle couvre, pour qui, à partir de quel prix.
## Resources
- [Titre du guide](https://www.exemple.fr/guide/) : la question à laquelle il répond.
## Optional
- [Actualités](https://www.exemple.fr/blog/) : publications récentes.The summary deserves the most care: it is the sentence an assistant will reuse to describe the company. It should be identical to the one on the About page, the Google Business Profile listing and the LinkedIn profile, so that every source agrees. That consistency weighs more than the file itself, as entity: definition.
Step 4: generate and publish
- By hand for a site of fewer than fifty pages: a text editor is enough, in thirty minutes.
- By script for a site with plenty of content: a script that walks the site's data and produces the file with titles and descriptions, re-run at every publication. That is the approach used on this site.
- By extension for WordPress, Shopify or another CMS: several modules generate the file automatically from the existing metadata.
- Publication: the file goes at the domain root, reachable at
votre-domaine.fr/llms.txt, served as plain text, with no authentication and no redirect. - The full variant: an
llms-full.txtfile containing the full text of your key pages in Markdown, useful for technical documentation, unnecessary for a small business site.
Step 5: test
Three checks are enough. Open the URL in a browser: the file should display as plain text, with no 404. Confirm each link returns a 200, since a dead link gives a poor impression. Finally, paste the file's address into an AI assistant and ask it to describe the company and its offers from that document. The answer you get is what the file actually communicates; if it is vague or incomplete, it is the summary and the descriptions you need to rewrite, not the structure.
What to expect from it
No assistant publisher has announced reading this file systematically to build its answers, and Google has said it doesn't use it for search. The tools that exploit it today are mainly agents and development assistants a user gives a site address to. The value therefore stays modest and indirect: the file costs an hour, carries no risk, forces you to state the offer clearly, and will be ready should the engines adopt it. The levers that really determine citations lie elsewhere: indexing, page structure and source consistency, described in structuring your pages for AI and appearing in LLM results.
Maintenance
- At every significant publication: add the link and its description, or re-run the generation script.
- At every change of offer: update the matching section, and the summary if the positioning shifts.
- Every quarter: check for dead links and the accuracy of the descriptions.
- After a redesign: regenerate the file completely, since the URLs have changed.
How GreenRed helps
Rather than juggling several tools, GreenRed's GEO and AI module brings these metrics together in a single dashboard, compares them over time and tells you which actions come first. You can try it free, with no card, from the Pricing.
Frequently asked questions
How do you create an llms.txt file?
By writing a Markdown document with a title, a one-sentence summary, a description of the business and sections of annotated links to your reference pages, then publishing it at the domain root at your-domain.com/llms.txt, served as plain text.
How many links should go in an llms.txt?
From 20 to 200 depending on the size of the site, each with a line of description. Selection counts for more than exhaustiveness: a file of two thousand links with no comments adds nothing to an XML sitemap.
Does llms.txt improve visibility in AI?
No demonstrated effect in 2026: assistant publishers haven't announced reading it systematically and Google says it doesn't use it. It is mostly read by agents a user gives an address to. Its value is clarity, preparation and a near-zero cost.
How do you test your llms.txt?
By opening the URL to check it displays as plain text, confirming each link responds, then pasting the file's address into an AI assistant and asking it to describe the company. The answer you get shows what the file actually communicates.