llms.txt v2 adoption: the first public baseline
Two days after the proposal’s v2 update, we captured an implementation baseline. It measures observable publishing patterns, not agent support or SEO impact.
Last updated:
The 12 August 2026 baseline starts with 113 hosts whose root response was non-HTML and began with an H1. That is a reproducible inclusion rule, not proof of full v2 conformance. The run then tests four common path-level locations and samples up to three linked pages per file for reachability, Markdown alternates, and rel="describedby".
Key takeaways
- 21 hosts exposed at least one distinct tested path-level file.
- 31 of 286 reachable pages advertised Markdown; four advertised
describedby.- The baseline measures implementation, not compatibility or ranking benefit.
What does the baseline measure?
The input is the successful subset of a fixed 219-host panel. The earlier weekly check found 113 root responses meeting the H1-first inclusion rule among 218 reachable hosts on 3 August. The v2 run used those 113 files, preserving a defined and public denominator.
For each host, the crawler tested /docs/llms.txt, /documentation/llms.txt, /help/llms.txt, and /api/llms.txt. It also sampled up to three HTTP links listed in the root file and inspected successful HTML responses for the two v2 relations.
The five measured results
- Files analyzed: 113.
- Hosts with a distinct tested path-level file: 21.
- Linked pages requested: 294, of which 286 were reachable.
- Reachable pages advertising Markdown: 31.
- Reachable pages advertising
rel="describedby": 4.
The same directory snapshot shows 90 of 113 root files using a blockquote. Median structure was 14,347 bytes, 7 H2 sections, and 76 list links. Raw aggregates are published in v2-adoption-stats.json and directory-insights.json.
What can we conclude?
Path-level files were observable before or immediately around the v2 update, while explicit describedby discovery remained rare in the sampled pages. Markdown alternates existed on a minority of pages. These statements are limited to this dated sample.
We cannot conclude that any agent uses the relations, that unadvertised Markdown does not exist, or that the features affect Search. Google says llms.txt is not a special Search signal. The SEO evidence guide separates those questions.
How the series will continue
Future runs will keep the same host input and test paths, publish their dates, and record methodology changes. New proposal revisions will be annotated rather than silently rewriting historical results.
The specialist reports cover subpaths, discovery relations, link health, and file size.
Is this a Web-wide adoption study?
No. It is a fixed, curated technology-oriented panel.
Why publish a zero result?
Zero is useful when the scope and method are explicit. It creates a baseline for later comparison.
Will later reports overwrite this result?
No. This article states the date. Current JSON snapshots can change, while the historical claim remains tied to the 12 August run.