seo-geo-factdensity (M12)

Generative engines preferentially cite passages that are concrete, quantified, and attributable. This module measures how "citable" the page's prose is — facts, numbers, original data, and authoritative outbound links — not vocabulary. AI retrieval/citation context: references/ai-crawlers.md.

Inputs

Work from the PageSnapshot named in your dispatch envelope: read parsed from <run_dir>/pages/<slug>.json (anchors[] for outbound links, text_sample) and the HTML file for passage-level counts; 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. Statistic/number density per passage: tokenize the main content into passages (paragraph / <li> / heading-bounded block) and count numeric tokens — figures, percentages, dates, quantities, ranges. Flag long passages of pure assertion with zero numeric support.
2. Proprietary/original data: detect first-party-data signals — patterns like "our study", "our survey", "our data", "we analyzed", "we surveyed", "in our test", "internal data" — and note whether such claims are backed by a method/sample, a table, or a chart.
3. Outbound citations: count outbound links from the main content to authoritative sources (standards bodies, primary research, official docs, .gov/.edu, named publications); distinguish them from internal/nav/affiliate links.
4. Claim-without-source flags: detect strong factual or comparative claims ("the most", "fastest", "studies show", superlatives, hard numbers) that carry no inline citation or data reference, and mark each as a candidate for sourcing.

Fixes (fixable: advisory)

ADVISORY only — this module proposes nothing it would write. It produces a list of (a) claims that should carry a statistic or citation, (b) passages where an original-data callout (table, "our data" box, methodology note) would raise citability, and (c) unsupported superlatives to soften or source. The tool will NOT fabricate statistics, sample sizes, study results, or source URLs. Where a value is missing, it emits a clearly-marked TODO placeholder for the user to fill — never an invented number. (Findings here are fixable: advisory per the finding schema.)

Verification

• Deterministic counts plus manual_review: node "${CLAUDE_PLUGIN_ROOT}/scripts/factdensity.mjs" --snapshot <pages/<slug>.json> [--lang en|es|auto] (or --url <u>) returns numeric-token density per passage — split into all / years / prices / substantive — and the outbound-authority link count. Report the raw counts so a human can re-derive them; a page whose only "numbers" are a copyright year and a price is not fact-dense, which is why substantive is the count that matters.
• Language coverage: authority-host rules, superlative lexicons, and original-data phrases ship for EN and ES (including .gob.mx, .gov.uk, scielo.org). Outside those languages the language-dependent fields come back null, never 0, and the finding is manual_review.
• Judgment calls (is this claim "strong"? is this source "authoritative"?) require manual_review; do not auto-pass them.
• When the content tier or rendered DOM needed to count passages reliably is unavailable, status is needs_api, never a false pass.

Findings

Emit findings per schema/finding.schema.json; axis ai; fixable: advisory throughout. 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 M12 is established, so nothing here ever caps.

• M12.density.low_numeric_passages — multiple main-content passages with zero substantive numeric support (warn, severity 3, confidence directional).
• M12.citations.no_outbound_authority — the main content makes factual claims and links to no authoritative outbound source (warn, severity 3, confidence directional).
• M12.original_data.missing — a topic that invites first-party evidence (benchmark, survey, pricing study) carries none (warn, severity 3, confidence directional).
• M12.original_data.present — a first-party data claim backed by a method, sample, table, or chart (pass, severity 3, confidence directional).
• M12.claim.unsourced_superlative — a superlative/comparative claim with no inline citation or data (warn, severity 1, confidence speculative).
• M12.lang.unsupported — the page language is outside the EN/ES lexicons, so density and authority scoring cannot decide (manual_review, severity 0; language-dependent fields null). Each finding: evidence.observed quotes the exact passage/claim from the page; verification.reproduce is the runnable count above; expected_impact is banded + confidence-tagged (no naked %).

Honesty

• What Google actually asks for is unique, non-commodity content — information that is not already available on every other page on the topic, produced by people with first-hand experience. Numeric density and outbound citations are our proxies for that, not the thing itself; they are directional and they are never established. A page can be dense with numbers and still be commodity content.
• Refuse "AI-specific keyword" rewrites — there is no magic vocabulary that wins citations. Citability comes from extractable structure, verifiable facts, and demonstrable authority, not phrasing tricks.
• Never invent a statistic, sample size, or source to "fill" a flagged claim. Quantification only helps if it is true and attributable; a fabricated number is a liability, not a win.
• Density is a means, not an end — stuffing numbers into prose that does not warrant them is its own anti-pattern, and this module flags it as one.
• Outside EN/ES the honest answer is manual_review, not a pass: a zero count in a language we cannot tokenize is a gap in the tool, not a finding about the page.