llms.txt

What it is

One Markdown file, at /llms.txt, listing the pages of a site that are worth reading, each with a line saying what it holds. Optionally a second file, llms-full.txt, with much more of the content inline.

The idea comes from the same place as robots.txt: a model asked about your company has a limited budget of pages it can fetch, and HTML is an expensive way to spend it — navigation, scripts, markup, cookie banners, and the answer somewhere in the middle. A curated index of plain text is cheaper to read and easier to quote.

How honest to be about it

Completely, because this is the part most articles about llms.txt skip: no major AI crawler has committed to reading it. It is a proposal with real momentum and no guarantee behind it. Anyone telling you it will get your site into ChatGPT’s answers is selling something.

What makes it worth doing anyway is the cost. On this site the four files — English and Polish, short and full — are generated from the content collections at build time. Nobody writes them, nobody updates them, and a new post or a new page appears in the right one automatically. A build gate fails if a page in the sitemap is not linked from the index of its language. That is a rounding error of effort against an upside that might arrive.

If it had to be maintained by hand, my answer would be different.

What goes in it

The index is only useful if what it points at is also cheap to read, which is why the two go together here: every page the sitemap lists ships a plain-Markdown twin at the same address with a .md suffix, and every line of llms.txt names it. The headings in those files carry the same anchors as the HTML page, so a model that quotes one section can cite a URL that resolves for a human reader too.

Each language gets its own complete index, linking only its own pages and pointing at the other. One bilingual file would have made both halves harder to read and neither complete.

What llms.txt is not

It is not a ranking signal, it is not read by Google Search, and it does not replace either the sitemap or the structured data. It is also not a place for marketing copy: the file exists so a machine can find the sentence that answers a question, and padding it with adjectives makes it worse at exactly that.