evidence-density

Absorption-stage audit for whether cited pages can shape AI answers. Use when a brand is found but not quoted or its pages add little to answers. Measures extractable evidence genres and structural le

SKILL.md

GitHub

Evidence Density (absorption layer)

Selection and absorption are different outcomes. arXiv:2604.25707 measures the gap directly:

Platform Mean citations per prompt Mean influence of fetched pages
ChatGPT 6.88 0.2713
Google AIO 12.06 0.0584
Perplexity 16.35 0.0646

Perplexity cites nearly two and a half times as many sources as ChatGPT and absorbs each one roughly a quarter as deeply. Being cited is not the same as mattering, and this skill measures the second thing.

When to use

After crawl-reachability has built the page inventory. Run it standalone when a brand is found but never quoted — the symptom of a selection pass with an absorption failure.

Inputs

  • Required — a run workspace: preflight.json for the scope domain and reachability.json for the page inventory and saved HTML paths.
  • Not consumed — no network access. This layer re-scores markup that crawl-reachability already fetched, so it is re-runnable offline.
python3 scripts/score_evidence.py [--workspace PATH] [--json]

Procedure

Deterministic, in this order. Text statistics only — no model, no sampling, no wall-clock input, so the same markup always scores the same.

  1. Load pages and scope. Read preflight and reachability; note pages whose HTML is missing as a coverage limitation.
  2. Thin content. Score pages with too little substance to carry an absorbable unit.
  3. No evidence genre at all. Flag pages carrying none of the five measured genres (this is the finding that moves the paper's measured influence).
  4. The Q&A trap. Flag Q&A-shaped pages with no evidence genre underneath, and never recommend adding an FAQ.
  5. Structure. Measure heading hierarchy, list density and paragraph length as legibility proxies.
  6. Genre gaps across the site. Roll per-page genre coverage up to find the genres the site systematically lacks.
  7. Survey-graded genres. Check for prices/rates and visible dates — the extractable-fact genres arXiv:2607.14035 §7.2 names and §7.5 grades Moderate. Reported with the grade, never with an effect size.
  8. Marketing over evidence. Compare promotional language against concrete substantiation, and attach the retrieval-backfire caution to the fix.
  9. Boilerplate. Discount repeated chrome so it cannot inflate a page's apparent substance.
  10. Site-level rollup. Emit the absorption distribution and the genre matrices the report renders.

The one finding people get backwards

Adding an FAQ is the standard GEO recommendation. The paper measures Q&A formatting as the only negative evidence genre:

Genre Influence with Influence without Difference
Code 0.1747 0.0988 +76.9%
Numbers / statistics 0.1171 0.0725 +61.6%
Definition markers 0.1252 0.0795 +57.3%
Comparison content 0.1389 0.0894 +55.3%
How-to content 0.1296 0.0918 +41.2%
Q&A format 0.0947 0.1005 −5.7%

The mechanism the paper proposes: evidence genres create reusable support units, while Q&A formatting is a surface wrapper. Rephrasing a thin page as a question does not add anything to extract.

This skill therefore never recommends adding an FAQ, and flags Q&A-wrapped pages that carry no evidence genre underneath.

Checks and severities

Every check the procedure above can emit, with the claim level it is allowed to assert.

Check id Detects Severity Claim level
thin_content Pages below the site-calibrated word floor high 1
no_evidence_genre No definition, numeric, comparison, how-to or code content high 2
qa_wrapper_without_evidence Q&A packaging with no underlying evidence medium 2
unstructured_page Too few headings or paragraphs to segment medium 1
no_numeric_evidence Claims made with no figures anywhere on the site medium 2
no_comparison_content Nothing positioning the brand against alternatives medium 3
no_definition_content Nothing defining what the brand or product is medium 3
no_price_evidence No price, rate or cost anywhere on the site medium 2
no_date_evidence No published or updated date on any sampled page low 2
marketing_over_evidence Promotional hedges far exceeding extractable facts medium 3
low_absorption_site Site-wide readiness in the weak band medium 3
boilerplate_dominant Chrome outweighing substance medium 2

Severity is capped by claim level automatically. The three mechanism-level checks (no_comparison_content, no_definition_content, marketing_over_evidence) cannot exceed medium no matter how stark the observation, because the paper supports them as interpretation rather than as measured cause.

The two genre checks from the survey (no_price_evidence, no_date_evidence) are grounded differently from the rest: arXiv:2607.14035 grades prices and dates Moderate on its lever table (§7.5, Table 4) rather than reporting an effect size, so these findings cite the grade and are capped at claim level 2. Neither is a defect — a page with no price is harder to quote, not broken. A price or date is only useful when it is specific, current and attributed; the research is explicit that the goal is not to "add numbers" (survey §7.2), and a fabricated figure is worse than none. Much of this module's advice recommends editing page content, so every such recommendation also carries the survey's backfire warning (§7.4): in a controlled arena, body-only rewrites reduced top-20 presence by about 9%, top-10 after reranking by about 16%, and final citation by about 6%. Add the missing fact; do not rewrite the page around it.

Page-type exemptions

Absorption is not a universal goal. The skill exempts:

  • Legal pages (privacy, terms, cookies) — optimising these for citation is not desirable.
  • Hub and category pages — their job is routing, not explaining. Judging a storefront category page on prose depth is a false positive.

Exemptions come from AuditConfig.expectations_for(page_type) and are recorded in the output so a reviewer can see what was skipped and why.

Thresholds are derived, not hardcoded

The word-count floor is max(corpus bottom quartile of 170, half this site's median), capped at 600. A site of 2,000-word guides gets a higher bar than a site of 200-word product pages, so "unusually thin for this site" stays detectable either way. Every threshold reports its provenance in the finding's evidence string.

Output

04-evidence-density.json with per-page scores, the site rollup, the genre matrix and the findings.

Guardrails

Read-only; consumes saved HTML only and issues no requests of its own.

Grounding

§8.1 structure quartiles (words 11.44x, headings 12.46x, paragraphs 5.69x, list density 8.92x between top and bottom influence quartiles); §8.2 evidence genre uplift; §8.3 semantic role influence (definition 0.1531, comparison 0.1524, down to reference 0.0529); Fig. 6 word-count bins. The scoring model and its deliberate divergence from the paper's Equation (2) are documented in references/absorption-model.md.