llms.txt is published
An llms.txt file at a well-known location is the agent equivalent of a sitemap. It tells AI ingesters what pages on your site are worth reading and gives them a single, deterministic starting point for crawling.
How the check decides
The check sends a GET request to /llms.txt, then to /.well-known/llms.txt, then to /docs/llms.txt. If any of those return a 2xx response it passes. If none do, it falls back to the same paths under llms-full.txt (the long-form variant). If still nothing answers, the check fails.
How to implement it
Create a plain-text file named llms.txt at the root of your site (or under /.well-known/) that lists the pages you want agents to read. Each line is either a markdown link, a section heading, or free text. The format is intentionally simple, agents can parse it without an HTML parser.
Pass
# Example Docs
## Getting started
- [Install](https://example.com/docs/install.md)
- [Quickstart](https://example.com/docs/quickstart.md)
## Reference
- [API](https://example.com/docs/api.md)
Served at https://example.com/llms.txt.
Fail
No file at any of the well-known paths, GET /llms.txt, GET /.well-known/llms.txt, GET /docs/llms.txt, and the llms-full.txt variants all return 404.
What we measured
Publishing the file and getting an agent to read it are two different problems, and this check only measures the first one.
We benchmarked a clean llms.txt discovery layer against a coding agent, five runs per arm, everything else held identical. When the agent read the file, it used 33% fewer tokens on the same page (177,735 against 266,591). But it only read the file when the prompt told it to. Across four other arms, a <link rel="llms-txt"> tag, a <link rel="alternate" type="text/markdown"> tag, a visible “For agents” footer link, and no signal at all, the agent fetched llms.txt 0 times out of 5.
Publish it anyway. It is the half you control, the savings are real once something does read it, and the ingesters that look for it by convention will find it. Just do not expect the file on its own to change how an agent behaves, and if you are the one writing the prompt, say where to look.