What is llms.txt?
A plain-text file placed at the root of a website offering a structured, human-readable summary of the site for large language models.
A worked example
A site publishes a file at /llms.txt listing its main sections and key pages in Markdown, each with a one-line description of what the page covers, so a model retrieving the site gets a clean map instead of trying to infer structure from parsed navigation.
Why it matters
The proposition is straightforward: models increasingly retrieve web content, parsed HTML is noisy, and a curated summary lets you influence how your site is understood. It is cheap to implement, cheap to maintain, and gives you a small amount of editorial control over the framing of your own site.
It is also worth being precise about what it is not. llms.txt is a proposed convention rather than a ratified standard. Adoption across model providers is uneven and none of the major ones has committed to honouring it as specified. It is not a ranking mechanism and it will not get you cited on its own.
Our position is that it belongs in the same category as a well-maintained sitemap: sensible housekeeping with a plausible upside, not a growth lever. Implement it, keep it current, and be sceptical of anyone charging meaningfully for it.
The nuance most people miss
The most common failure is publishing it once and letting it rot. A file that lists pages which no longer exist, or omits half the site added since, is worse than not having one. It actively misdescribes you. Treat it as a generated artefact rather than a static file: build it from the same data that produces your sitemap so it cannot drift. The second common failure is stuffing it with marketing language, which defeats the purpose. It should read as a factual map, not a brochure.
Common mistakes
- Treating it as a ranking mechanism rather than a discovery convenience
- Publishing it once and never updating it as the site changes
- Filling it with promotional copy instead of factual descriptions
- Using Markdown link syntax incorrectly, which some validators and accessibility audits will flag