- [ Home ](/) / - [ Blog ](/blog/) / - llms.txt link health: 294 URLs tested # `llms.txt` link health: 294 URLs tested A syntactically valid map can still lead an agent to dead resources. This benchmark measures reachability without confusing it with content quality. Last updated: August 28, 2026 On this page - [ How healthy were the sampled links? ](#how-healthy-were-the-sampled-links) - [ What the reachability test measures ](#what-the-reachability-test-measures) - [ A safe link-checking policy for CI ](#a-safe-link-checking-policy-for-ci) - [ Limits of an HTTP result ](#limits-of-an-http-result) A valid Markdown list is not enough. If the linked pages time out, redirect unexpectedly, or return errors, the file stops being a reliable route into the site. Our baseline reached 286 of 294 sampled URLs, a 97.3% observed success rate for this run. **Key takeaways** - 286 of 294 sampled URLs were reachable on 12 August 2026. - Reachability is not the same as relevance, freshness, or agent use. - External failures should be retried before they block a release. ## How healthy were the sampled links? The v2 research crawl selected up to three list links from each of 113 root responses meeting the H1-first inclusion rule. It requested 294 URLs and received a successful response from 286. Eight were not reachable within the test conditions. This is a sample, not a full audit of every listed link. The median verified file contains 76 list links, so checking three per file favors breadth across hosts rather than exhaustive depth inside a single large map. ## What the reachability test measures The script follows redirects, applies a seven-second timeout, and records whether the final response is successful. It does not log in, execute a browser session, judge textual quality, or assert that the destination matches its link description. A success can still hide a soft 404 or generic application shell. A failure can be temporary, rate-limited, geographically restricted, or caused by bot protection. That is why the result is published as an observation with a date and method. Use the [verified directory](/examples/) for the fixed input and the [validator](/validator/) for file-level syntax. Neither replaces human review of high-value destinations. ## A safe link-checking policy for CI Check internal links on every build because you control both ends. For external links, run a scheduled job, retry failures, and treat only persistent 404 or 410 responses as strong evidence of removal. Do not make every production push depend on hundreds of third-party hosts. Record the origin URL, final URL, status, and timestamp. Flag redirect chains separately from errors. When a canonical page moves, update the file rather than relying indefinitely on the redirect. For large files, sample on each run and rotate the sample, or run a full crawl less frequently. The [best-practice guide](/best-practices/) explains why a smaller curated list often costs less to maintain. ## Limits of an HTTP result HTTP 200 means the server answered. It does not mean the page is current, complete, safe, or useful to an agent. Conversely, one timeout does not prove a dead link. Preserve the raw result long enough to compare repeated runs before removing an entry automatically. The benchmark also says nothing about Google rankings. Link hygiene improves the reliability of the file for clients that read it; Search uses its own crawl and ranking systems. Review [how llms.txt works](/how-it-works/) and keep the scope aligned with the [v2 migration guide](/blog/llms-txt-v2-migration/). ### Should one broken external link fail production? Usually not. Retry it in a scheduled check and fail only on a policy you control, such as persistent 404 or 410 responses. ### Are redirects invalid in llms.txt? No, but direct canonical destinations reduce latency and ambiguity. Update stable redirects during maintenance. ### How often should links be checked? Match the content’s change rate. Weekly checks suit an active directory; stable documentation may need less frequent full scans. ## Sources - [ llmtxt.org, file-list proposal ](https://llmstxt.org/) - [ llmtxt.info, v2 benchmark ](https://llmtxt.info/v2-adoption-stats.json)