What is llms.txt?
llms.txt is a proposed convention for a Markdown file served at the root of a website, at the path /llms.txt, that describes the site to language models. The proposal, published at llmstxt.org, suggests a title, a short summary, and sections of curated links with one-line descriptions, so that an AI system fetching the file can understand what the site is and where its important content lives.
It is analogous in spirit to robots.txt or a sitemap, but for orientation rather than crawl control: a short, human-readable map intended for a model rather than a crawler.
What does it actually do?
It gives an AI system that fetches it a fast way to orient: what the site is about, which pages matter, and what each contains, in a format models read well. For a system deciding which of your pages to retrieve for a question, or summarising your site, that map is useful.
What it does not do is guarantee anything. Whether a given engine fetches llms.txt, and what it does with it, varies and is not fully documented. It is not a documented ranking factor for search or for AI Overviews. Treat it as a low-cost courtesy to systems that read it, not as a lever.
How should you write one?
Write it as an honest map. Start with a level-one heading naming the site, then a short blockquote summarising what the site is for in plain language. Follow with level-two sections grouping your important pages, each as a Markdown link with a one-line description of what the page answers. Optionally, a final section can list secondary pages a model can skip when context is limited.
Describe pages by what they answer rather than by marketing language. A model reading the file is trying to route a question; the definition of a Digital FTE with cost, roles, and governance is more useful to it than the future of work starts here.
- H1: the site name
- Blockquote: one or two sentences on what the site is
- H2 sections: groups of links with one-line descriptions
- Optional final section: secondary pages
- Plain Markdown, served as text at /llms.txt
What mistakes should you avoid?
Avoid turning it into a pitch. A file full of superlatives and no useful descriptions helps nobody and reads as low-trust. Avoid listing every page; the value is curation. Avoid letting it go stale, because a model that follows a link to a moved page learns not to trust the file.
Avoid claiming it does more than it does. If a vendor tells you llms.txt will get you cited, ask for their evidence; the honest answer is that it helps systems that read it and that citation depends on the content itself.
How does ftes.ai use it?
ftes.ai serves an llms.txt at its root describing what the site is, the Digital FTE roles it sells, and where the definitional, pricing, governance, and library content lives, each with a one-line description of what the page answers. It is updated whenever the library changes so that it stays current with the content.
It is one small part of the site's AI-search discipline. The larger part is the content standard itself: answer-first, self-contained, entity-explicit, and verifiable, which is what actually earns citation once a system has found the page.
What does a labelled example of an llms.txt look like?
As a labelled example of the shape: a first line that is a level-one heading naming the site. A blockquote of one or two sentences saying what the site is for in plain words, such as a marketplace for hiring AI employees that own a job function inside the tools a company already runs. A level-two heading for the main product area, followed by a short list of Markdown links, each with a one-line description of what the page answers: the definition page described as the definition of a Digital FTE and how it differs from agents and copilots, the pricing page described as token-metered blocks of work with a hard stop.
Further level-two sections group other important pages, such as security, industries, and a curated set of the library's hub pieces and whitepapers, each described by what it answers. An optional final section lists secondary pages a model can skip. The whole file is plain Markdown served as text at the root, kept short enough to read in a minute, updated whenever the important pages change, and free of superlatives. The version ftes.ai serves follows this shape and is regenerated by hand when the library's hubs change.