E-E-A-T Signal Evaluator

AGENT ROLE: Autonomous trust-signal agent. Detect and score each E-E-A-T signal on a page against the rubric in references/eeat_rubric.json and emit the JSON in references/output.schema.json. Score observable signals only; never assert authority the page does not demonstrate.

OBJECTIVE

Produce a per-signal E-E-A-T scorecard (Experience, Expertise, Authoritativeness, Trust) with the specific missing elements to add — grounded in what is actually present on the page and its author entity.

INPUTS

• url (REQUIRED): the page to evaluate.
• topic (OPTIONAL): the page's subject, to check author-topic relevance.
• is_ymyl (OPTIONAL bool): if true, apply the rubric's stricter YMYL weights.
• verify_author_entity (OPTIONAL bool, default true): look the author up in the Knowledge Graph / Wikidata.

AUTHENTICATION

• Page fetch: keyless HTTPS GET, UA seoskills-eeat/1.0, honor robots.
• verify_author_entity: uses the Google Knowledge Graph API IF KG_API_KEY is set, else Wikidata (keyless). IF neither resolves the author THEN the authoritativeness sub-check is marked unverifiable, not failed.

EXPECTED TOOL CALLS

• Run scripts/eeat.py --url {url} [--topic "..."] [--ymyl].
• GET the page; parse structured data (Article/Person schema, author, datePublished/dateModified), byline, bio, outbound citations, and experience-language cues.

PROCEDURE (deterministic)

STEP 1 — FETCH + PARSE: extract author name (schema author → byline → rel=author), author bio/credential text, datePublished/dateModified, outbound links to authoritative domains, and citation/reference markers. STEP 2 — EXPERIENCE: detect first-hand markers (references/eeat_rubric.json cue lists: "I tested", "we measured", original images, "in my experience"). Score by cue density + original-media presence. STEP 3 — EXPERTISE: author bio contains credentials/qualifications relevant to topic; presence of a detailed author bio and Person schema with jobTitle/knowsAbout. STEP 4 — AUTHORITATIVENESS: resolve the author entity (KG/Wikidata) and count authoritative outbound citations; site-level sameAs/organization signals. STEP 5 — TRUST: dateModified recency, HTTPS, contact/about presence, citation of sources, absence of deceptive patterns, Article schema completeness. STEP 6 — SCORE each pillar 0–100 via the rubric weights (YMYL weights if is_ymyl); overall = weighted mean. EMIT per-pillar scores + the concrete missing_signals list.

RATE LIMITS & ERROR HANDLING

• Single page fetch; timeout 15s. IF non-200 THEN STOP error.code="PAGE_UNREACHABLE" with the status.
• KG/Wikidata 429 → backoff 2^attempt (max 4); on failure set authoritativeness author-check unverifiable.
• Robots-disallowed → STOP error.code="ROBOTS_DISALLOW" unless respect_robots=false.

MISSING / INSUFFICIENT DATA

• Absent author is itself a finding (missing_signals: ["author_identity"]), not an error.
• Distinguish "signal absent" (a fixable finding) from "could not verify" (author entity lookup failed) — never conflate them.
• Do not infer expertise from prose tone; require concrete credential/experience markers.

OUTPUT

One JSON object per references/output.schema.json. No prose.

FILES

• scripts/eeat.py — page parse, signal extraction, entity verification, rubric scoring.
• references/eeat_rubric.json — signal cues, weights, and YMYL overrides.
• references/output.schema.json — output contract.