What this check tests
If no llms.txt was found, the check is skipped. Otherwise it pulls every http:// or https:// URL out of the file text with a pattern, keeps the first 10, and requests each one with a 5 second timeout, following redirects. A URL is "resolved" when the final answer is a 2xx or 3xx status.
- The score is the share of checked URLs that resolved. All resolved is a pass, some is a partial, none is a fail.
- A file with no URLs at all is a fail.
- A timeout, a network error, or a 4xx or 5xx answer counts as broken. That includes a link your server blocks for the audit crawler (a 403 from a bot filter, for example), even if browsers load it fine.
- The pattern stops at whitespace and at
),>,],,and:after the scheme, so a URL containing a comma or a port number can be cut short. Any URL in the file counts, not only the ones in list items. - Only the first 10 matches are checked. A broken link on line 40 of a long file will not show up here.
Why it matters to agents
The file is a list of places an agent is told to go next. A link that returns 404 sends a coding assistant to an error page, and a link behind a bot challenge sends it nowhere. The link list is the point of the file, so stale entries cost more than a missing blockquote.
How to fix it
- Open the details of a failing result in your audit report. It lists each broken URL.
- For a 404, update the link to the page's new address or remove the entry. For a redirect chain that ends in an error, link straight to the final working page.
- For a 403 or 429 that only affects automated requests, check your CDN or firewall rules for the path. See edge agent access for how the audit looks for that.
- Keep the first 10 links the most important ones, and keep them absolute.
## Docs
- [Quickstart](https://example.com/docs/quickstart): Create a project
- [API reference](https://example.com/docs/api): Endpoints and auth
A quick local check, run from a shell, prints the final status for each URL:
grep -oE 'https?://[^ )>]+' llms.txt | while read -r u; do
printf '%s %s\n' "$(curl -s -L -o /dev/null -w '%{http_code}' "$u")" "$u"
done
Related: llms.txt exists, llms.txt follows the format, and the llms.txt validator, which checks a live URL's links.