Changes

v2 (August 2026)

The original llms.txt proposal was published in September 2024, when the idea that language models would routinely read websites was still speculative. Since then the community has taken to the proposal far more than I expected. Thousands of sites now publish an llms.txt file, documentation platforms generate one automatically, and coding agents use them reliably. That shift is what this revision reflects.

Adoption brought requests, and the commonest was discoverability. Given a page, how does an agent find its markdown version, or the llms.txt file that covers it, without guessing? v2 answers with standard link relations: rel="alternate" type="text/markdown" points to a page’s markdown version, and rel="describedby" points to the llms.txt file that covers it, provided as HTML <link> elements or an HTTP Link: header.

Practice also diverged from v1 in ways worth blessing. v1 specified one URL form for markdown versions, .md appended to the full page URL (page.html.md). Some publishing tools instead replace the extension (page.md), so v2 allows both. v1 permitted llms.txt files in subpaths without saying what that meant. v2 defines it: a file covers the pages under its path, and the most specific file applies. This is also what lets a site that only controls a path, such as a GitHub Pages project site, participate fully.

v1 said nothing about how llms.txt should be consumed, and described the llms_txt2ctx tool for expanding a file into an LLM context. v2 instead states the expectation directly: agents view or search the llms.txt to find what they need, then follow the relevant links, which should point to LLM-friendly content. The context-expansion tooling is no longer part of the proposal, and with it goes the special meaning of the Optional section, which told those tools what to omit. Optional sections are still allowed, and remain a useful convention for secondary links, but they no longer carry mechanical semantics. Finally, the background and examples now describe how agents actually use websites, rather than predicting that they might.