Skip to content
llllmslist.
← Back to directory

Transparent documentation checks · V2

How readiness is calculated

AI Resource Readiness measures publicly observable documentation quality and accessibility. It does not measure ChatGPT, Claude, and Gemini ranking, citation probability, model preference, factual correctness or guaranteed discoverability. The weights below are LLMSList product choices, not scientifically validated predictors.

The formula

Score = round(earned points ÷ assessed maximum × 100)

The default five components total 100 points. Each check earns its stated maximum multiplied by its measured fraction. Check points are rounded to two decimal places; the overall score is rounded to a whole number. If all 100 points were assessed, 82 earned points gives 82/100.

If destination checks are entirely inconclusive, their 10 points are excluded: 72 earned out of 90 assessed gives round(72 ÷ 90 × 100) = 80. That result is labeled provisional and the excluded points are shown. Unknown does not mean broken. A partly conclusive sample uses only working and confirmed-broken results for its fraction; unknown outcomes reduce the confidence label.

Every site’s detail page shows its actual rule weights, observations, arithmetic, score version and calculation time. Operators may configure different five-component weights that still total 100; the check weights scale proportionally and the profile is the authoritative record.

The five components

Accessibility

25 points

  • Current usable file

    25 points

    The latest primary llms.txt check succeeded with usable Markdown/plain text (HTTP 200 or a valid 304). A failed latest check earns zero current-access points even if older content was good.

Document structure

25 points

  • Project title

    10 points

    Exactly one useful H1, appearing as the first content block. Placeholder-only titles do not qualify.

  • Coherent sections

    5 points

    At least one H2 followed by content, with no jumps of more than one heading level.

  • Documentation file lists

    5 points

    At least one valid HTTP(S) documentation link inside a Markdown list.

  • Non-repeated headings

    5 points

    5 × unique heading texts ÷ all real headings. Case and whitespace variants are treated as duplicates. No headings earns zero.

Link quality

25 points

  • URL syntax

    5 points

    5 × valid HTTP(S) URL occurrences ÷ all valid, malformed, unsupported and unresolved occurrences. Relative URLs are resolved against the file URL.

  • Unique destinations

    5 points

    5 × unique destinations ÷ valid URL occurrences. Fragments are ignored when deduplicating pages.

  • Descriptive labels

    2.5 points

    2.5 × unique links with non-placeholder labels ÷ unique links. Generic labels such as “click here”, “link” and bare URLs do not qualify. This is a text heuristic.

  • Notes beside links

    2.5 points

    2.5 × unique links with explanatory notes beside a Markdown list link ÷ unique links. The first occurrence of a repeated destination supplies its label and notes.

  • Sampled accessibility

    10 points

    10 × working sampled destinations ÷ working plus confirmed-broken sampled destinations. HTTP 404/410 is broken; HTTP 403/429, timeouts and restricted checks are unknown. If links exist but there is no decisive sample result, these ten points are excluded. No documentation links earns zero, not a fictitious perfect success rate.

Freshness & confidence

15 points

  • Verification age

    10 points

    Last successful primary verification up to eight days old: 10 points. More than eight through 30 days: 6 points. Older or unknown: 2 points. The eight-day window gives weekly scheduling one day of grace.

  • Latest check succeeded

    5 points

    Five points when the latest primary verification succeeded; zero during an outage. Checking unchanged content today confirms current access, not that the author recently edited it.

Document context

10 points

  • Summary blockquote

    5 points

    A useful blockquote before the first H2, with at least 20 non-whitespace characters and not only placeholder text.

  • Introductory paragraph

    5 points

    A separate useful paragraph before the first H2 with at least 20 non-whitespace characters and not only placeholder text. Listing name, description and category fields earn no points.

Format compliance and quality are different

The llms.txt proposal requires a project/site H1. Summary, sections and annotated links are useful recommendations, not universal validity requirements. An H1-only document remains usable and earns the checks it satisfies. Missing optional sections may lose LLMSList quality points, but do not make accessible text invalid. Subpath files such as /docs/llms.txt are supported.

Markdown is parsed as a syntax tree. Headings and links inside fenced/indented code examples and images do not count as documentation structure. Repeated headings and links cannot inflate the score. Long files, large token counts and many sections earn no extra quantity bonus. llms-full.txt and WebMCP remain optional capabilities and do not add or remove points.

How link sampling works

We select up to five unique destinations deterministically from the document content hash and URLs. The sample is refreshed during verification, including when the file is unchanged. At most two distinct hosts are checked concurrently, with a 15-second total sample budget, a four-second per-request limit, bounded 64 KiB responses, and the same DNS/address/redirect controls as primary checks. Full response bodies are not stored or executed.

Each detail page lists the selected URLs, outcomes, check time and source snapshot. HTTP 2xx is working; only HTTP 404/410 is confirmed broken. Access restrictions, rate limits, server errors, response limits, timeouts and checks deferred by the budget remain unknown. A successful response can still be irrelevant or a soft error page; we do not claim semantic validation. A sample of five working URLs does not certify every link on a large website.

Availability, outages and confidence

No first score is published until the primary llms.txt has been successfully retrieved and analyzed. HTML error pages, empty files and unsuitable content types do not count as usable documentation. During an outage, current-access points drop to zero and the confidence label becomes low, while previously measured document-quality evidence is retained and clearly identified as historical. Invalid new content receives no score until corrected.

Confidence is high only when the current primary check succeeded within eight days and link measurements have no unknown outcomes. It becomes medium for older evidence through day 30 or inconclusive/partial link measurements, and low for a failed primary check, an invalid date or evidence older than 30 days. The label describes confidence in these measurements, not confidence that an AI will recommend the site.

Versions and weekly refresh

V2 replaces the old three-heading structure test, syntax-only link score and listing-field metadata score. Legacy V1 scores are explicitly labeled while awaiting recomputation and are not directly comparable with V2. Existing approved listings are picked up for V2 verification by the worker; new approvals are checked immediately. Successful checks repeat weekly, with bounded backoff on failures.

A file’s source URL, SHA-256 hash, processing version, last successful verification and score calculation time identify the evidence behind its result. Pages may lag a completed check by their cache lifetime. Token sizes estimate characters divided by four; token size is informational and does not improve readiness.