llms.txt: The New File AI Models Are Starting to Read

🇫🇷 Version française

After robots.txt for crawlers and sitemap.xml for indexing, a third file is showing up at the root of some websites: llms.txt. Its job isn't to manage permissions — it's to save a language model time when it's trying to quickly figure out who you are. Here's what it actually does, and whether it's worth your attention right now.

A language model answering a question about your company doesn't read your site the way a human does. It doesn't have the time or the compute budget to browse a hundred pages, wade through JavaScript, and navigate a menu before finding the useful information. llms.txt starts from that observation: offer, at the root of the domain, a Markdown summary of the pages that actually matter, designed to be swallowed in one read by a model rather than browsed like a website.

How is llms.txt different from robots.txt and the sitemap?

The three files coexist, but they answer different questions:

A sitemap lists everything, with no distinction: a strategic product page and a legal-notice page get the same status. llms.txt does the opposite: it selects, summarizes each piece of content in one line, and points to the most readable versions (often plain Markdown pages rather than script-heavy HTML).

What does an llms.txt actually look like?

The format is deliberately simple: a title, a short company description, then sections that group links by topic. Here's the typical, stripped-down structure:

# Company Name

> One sentence summarizing what you do and for whom.

## Documentation
- [Getting started guide](/docs/getting-started.md): the basics to get going
- [API reference](/docs/api.md): full list of endpoints

## Offer
- [Pricing](/pricing.md): plan and price details

## Resources
- [Blog](/blog.md): in-depth articles on the topic

The point isn't to duplicate your content — it's to give a model the map of it. Some sites go further with an llms-full.txt, a version that concatenates the full text content of key pages, useful for models that prefer to receive everything at once rather than follow links one by one.

Should you already adopt it in 2026?

Let's be direct: support for llms.txt remains uneven. Some AI tools and agents actively consult it, others ignore it completely and keep crawling regular HTML the way they always have. It is not — not yet — an officially adopted standard across every answer engine.

What changes the calculation is the setup cost: a text file of a few dozen lines, no risk to the rest of the site, no conflict with robots.txt or the sitemap. For a company that already has solid GEO fundamentals — structured content, an open robots.txt, accurate data — adding it is an hour of work that prepares for the future at zero cost today. For a company that doesn't have those fundamentals yet, it's clearly not the priority: a flawless llms.txt on a poorly structured site doesn't compensate for anything.

What matters more than the file itself

llms.txt remains an acceleration tool, not a foundation. A model with no access to this file can still cite you if it finds, through normal crawling, content that's clear, well-structured, and up to date. Conversely, a perfect llms.txt will never make up for a site where information is vague, contradictory, or hard to find. Prioritize content quality and structure first; the file comes second, as a bonus.