seo-geo-answerblocks (M11)

AI engines cite passages they can lift whole, without surrounding context — so passage structure is the highest-leverage on-page AI signal. This module checks extractability; for who is allowed to fetch and cite the page, see references/ai-crawlers.md.

Inputs

Work from the PageSnapshot named in your dispatch envelope: read parsed from <run_dir>/pages/<slug>.json (headings[], text_sample, word_count) — passages come from the HTML file, not from a summary; Grep pages/<slug>.html for verbatim evidence; site artifacts live in <run_dir>/site/{robots.json,sitemaps.json,discovery.json}. Deterministic findings already emitted by audit.mjs are listed in <run_dir>/findings.deterministic.json — do not re-emit those ids; add model-judged findings only. If invoked directly with a URL/path and no snapshot exists, first run node "${CLAUDE_PLUGIN_ROOT}/scripts/snapshot.mjs" <target> --out "${CLAUDE_PLUGIN_DATA}/runs" and use the printed snapshot path.

Audits

Working from the PageSnapshot (parsed_rendered when render.used is not none, else parsed):

1. Heading → direct answer: for each H2/H3 that reads as a question (or implies one), is the first paragraph a direct, ~40-60 word answer — not a windup ("In this section we'll explore…")?
2. Self-contained passages: are top-level passages ~134-167 words and answerable on their own? Flag unresolved anaphora ("as mentioned above", "this", "the latter") that needs prior context to parse.
3. Structured forms: presence of lists, comparison tables, and explicit term definitions where the content warrants them — these are disproportionately quoted.
4. Question-shaped headings: do headings phrase real user questions (how/what/why/when), matching likely prompts?
5. TL;DR / summary blocks: an up-front summary or key-takeaways block the engine can extract verbatim.
6. Semantic completeness: can each passage answer its heading with zero external context — the strongest single correlate of AI citation, and the heaviest check in this module (it carries the most weight, but it never caps the score — see Findings).

Fixes

• PROPOSED: answer-block rewrites (heading → tight 40-60 word lead) and TL;DR drafts, generated from existing page content, surfaced one item at a time for the user to accept/edit. Each becomes a fix_preview diff only after acceptance.
• ADVISORY: structural suggestions ("split this 400-word passage", "add a definition list here") where intent is editorial.
• Never silently rewrite published prose, and never invent facts to fill a passage — if a claim is missing leave a clearly-marked TODO placeholder for the user. No finding here is auto.

Verification

• Method is manual_review. The deterministic part is reproducible offline: word-count windows (answer 40-60w; passage 134-167w), question-heading detection, anaphora openers, and TL;DR presence via node "${CLAUDE_PLUGIN_ROOT}/scripts/check-answerblocks.mjs" --snapshot <pages/<slug>.json> [--lang en|es|auto] (or --url <u> / --file <path>). Use manual_review as the finding's verification.method (the enum has no heuristic value); the script is the verification.reproduce command.
• Language coverage: the heuristics ship EN and ES lexicons (question words including ¿, windup openers, anaphora, TL;DR/Resumen markers). When the detected language is neither, the language-dependent fields come back null — never 0 — and the finding is manual_review, so a page in an unsupported language is never scored as a clean pass.
• Semantic completeness and anaphora resolution still require a judgment pass; when that pass or the rendered DOM is unavailable, status is needs_api (or manual_review), never a false pass.

Findings

Findings conform to schema/finding.schema.json; axis ai. Severity policy: 5 is reserved for catastrophic, eligibility-killing facts at site/template scope; 4 major · 3 moderate · 2 minor · 1 cosmetic · 0 informational — and only an established severity-5 fail in an active category can cap a score (references/scoring-model.md). Nothing in M11 is established, so nothing here ever caps.

• M11.heading.no_direct_answer — a question H2/H3 followed by a windup instead of an answer (fail, severity 3, fixable: proposed, confidence directional).
• M11.passage.unresolved_anaphora — a passage opens with "As mentioned above…" / "Esto…" and cannot be read standalone (warn, severity 2, fixable: proposed, confidence directional).
• M11.passage.overlong — a passage runs well past the extractable window (>167 words) with no internal structure to lift (warn, severity 2, fixable: proposed, confidence directional).
• M11.summary.missing_tldr — no up-front summary / key-takeaways block on a long-form page (warn, severity 2, fixable: proposed, confidence directional).
• M11.answers.ok — question headings are answered directly and passages are self-contained (pass, severity 3, confidence directional).
• M11.lang.unsupported — the page language is outside the EN/ES lexicons, so the passage heuristics cannot decide (manual_review, severity 0; language-dependent fields null). Each finding's evidence.observed quotes the page (heading text + first sentence, or the offending clause) and verification.reproduce is the runnable command above; expected_impact is banded + confidence-tagged with no naked percentage.

Honesty

• Google does not ask for AI-specific rewriting or chunking. Its generative-AI guidance points back at the same fundamentals — indexable, snippet-eligible, well-structured pages written for people — and it publishes no passage length, no "chunk size", and no answer-block format. Everything in this module is a readability and extractability argument that also helps human readers; say that, and never sell a rewrite as an engine requirement.
• Word-count windows are heuristics from observed citation patterns, not engine-published thresholds — they are directional, never established; flag a passage, don't hard-fail purely on length.
• Adding a "TL;DR" label or bolting question marks onto headings does nothing if the passage still cannot answer standalone. Semantic completeness is what carries weight; formatting theatre is a myth this tool does not ship.
• The heuristics are language-bound. On an unsupported language the honest output is manual_review, not a pass — a silent zero would read as "no problems found".