We assess the proof before a check can affect your score. Documented support matters. A popular idea alone is not enough.

Advisory and experimental checks do not change the score. The policy below explains the grades, limits and review rules.

Evidence Policy

Agent Lighthouse scores websites on their readiness for AI agents. A wrong audit is worse than no audit: it sends site owners to do work that has no effect, and it erodes trust in every other result. This policy defines what counts as proof, how audits are graded, and what each grade is allowed to do to a score.

The core rule

No audit contributes to a score unless its mechanism is proven at grade A or B.

Every audit must state its mechanism as a falsifiable causal claim — “GPTBot fetches /llms.txt before crawling”, not “AI systems prefer well-structured sites”. Vague claims cannot be proven or refuted, so they cannot be scored.

A grade licenses the mechanism it was proven for, and nothing wider. If the mechanism is about an artifact’s contents — an OpenAPI document’s servers array, a feed’s lastmod dates — the grade covers sites that publish that artifact. It says nothing about a site that publishes none, because no vendor documents a consumer for the absence. An audit about an artifact’s contents therefore returns notApplicable when the artifact is absent; only a present-and-defective artifact may fail. The grade-C row below records the same reflex from the other side: Chrome Lighthouse scores a missing llms.txt as notApplicable.

Evidence grades

Grade	Bar	Examples
A	Documented consumer behavior (a vendor doc states that a named agent reads the signal) or a ratified standard with known consumers	RFC 9309 robots.txt parsing; Anthropic documenting that `Claude-User` sends an `Accept` header preferring Markdown
B	Draft standard with meaningful adoption, or strong empirical evidence of effect	GEO paper citation-rate deltas; MCP authorization via RFC 9728 discovery
C	Community convention with partial adoption; plausible but unproven mechanism	`llms.txt` existence (published widely; no documented agent consumer — Chrome Lighthouse fetches it as an auditor and scores a missing file `notApplicable`; Google states Search ignores it)
D	Speculative or invented; no known consumer, no adoption evidence	security headers as “AI trust signals”; `agents.txt`; vendor-invented “AI trust score” meta tags

What grades are allowed to do

Grade	Scoring tier
A, B	Scored — contributes to the category and overall score
C	Informative — shown in the report with weight 0
D with an active draft-spec trajectory	Experimental — behind a flag, unscored
D otherwise	Rejected — not shipped; if previously shipped, deprecated with a dated public notice

Source requirements

Every claim cites a source id from sources.json — one registry, no scattered links.
Primary sources outrank secondary: vendor docs and spec text over blog posts over social media.
Every source URL is verified to resolve at research time; the registry records the access date.
Counter-evidence is recorded, not hidden. If Google states it ignores a file, the dossier for that file’s audit says so. Credibility comes from showing the evidence against our own checks.
CI runs a link checker over the registry. A dead link fails the build; a source older than 12 months triggers a re-review flag.

Re-review cadence

The AI-agent ecosystem changes quarterly. Every audit’s dossier carries a reviewed: date. An audit whose evidence is older than 12 months, or whose cited standard changes status (ratified, deprecated, abandoned), is re-graded. Downgrades below B remove the audit from scoring in the next minor release.

Deprecation process

Audits die publicly, not silently:

The dossier’s status changes to deprecated, with the evidence for removal cited.
The audit ships one more minor release as informative (weight 0) with a deprecation notice in the report.
The next major release removes it. The dossier stays published as the record of why.

History

2026-08-20 — policy adopted. Full review of all 207 v1 audits (11-agent code review + 12-domain evidence research, 400 sources) produced the current grade assignments.
2026-08-24 — evidence re-sweep on llms.txt and /.well-known/api-catalog. The grade-C llms.txt example was re-verified and holds: no documented agent consumer across 11 vendors, llmstxt.org v2, the IANA Well-Known URIs and Link Relations registries, and the IETF Datatracker; Google Search Central, updated 2026-07-10, still states Search ignores it. The policy was right and the audits were wrong — machine-discovery/llms-txt-exists (A / scored / 1.0) and machine-discovery/llms-txt-links-valid (B / scored / 0.6) were corrected to C / informative / 0 to match this row.
2026-08-24 — the grade-D example ai-catalog.json retired: it is no longer speculative or invented. Since 2026-06-17 it is the file defined by the Agentic Resource Discovery specification (draft v0.9; Google, Microsoft, Hugging Face and others under a Linux Foundation working group, Apache 2.0), and it has a documented first-party consumer client in huggingface/hf-discover, which performs automatic /.well-known/ai-catalog.json discovery from a website. Under this table’s own bars that is stronger evidence than the grade-C row above it, so keeping it as the canonical grade-D example was false on both halves.
2026-08-29 — the core rule gained its population clause: a grade covers sites that publish the artifact its mechanism is about. Four openapi-* audits (openapi-servers, openapi-endpoints, openapi-schemas, openapi-operation-ids — B / scored / 0.6 each, 2.4 combined) were charging fail at high or medium priority to every site with no OpenAPI document at all, on 41 of 41 corpus fixtures. No source documents a consumer that is worse off for the absence, and openapi-servers’ own recorded counter-evidence argues the opposite for the weaker case. Corrected to notApplicable on the absence; every verdict on a document that exists is unchanged.