Citebench Methodology v2

Status: LOCKED 2026-05-11. Empirically-grounded revision of v1. Research basis in methodology-research.md.

Updated 2026-05-11 (same day): Added per-page-type rubrics — see "Page-type-aware scoring" section. The article rubric below is unchanged; non-article page types use a subset of rules and different dimension weights.

Living document. AEO landscape shifts month-to-month — Google AIO algorithm updates, ChatGPT crawler changes, Perplexity-Reddit dynamics, new schema types Google starts citing, etc. Revisit weights quarterly (next: 2026-08) and after any major engine update. Score interpretation is "best current understanding," not "ground truth."

Page-type-aware scoring

Citebench detects page type and applies type-specific rubrics. A marketing homepage and a 3,000-word article shouldn't be scored identically.

Detected types:

• article — long-form editorial. Full v2 rubric applies.
• homepage — marketing homepage at root URL. Weights skewed to Source & Authority (40%) and Structural (25%); rules like "author byline" and "answer capsules under H2" are filtered out (N/A).
• landing — product or campaign landing page. Similar to homepage but with slight weight differences.
• machine-version — page serves AI-targeted markdown content to non-browser user agents. Awarded a positive signal (+5pt to Authority) for serving AI-friendly content; structural and answer-fit rules are mostly N/A because there's no HTML structure to evaluate.
• other — unclassified. Defaults to article weights but some rules skipped.

Per-type dimension weights:

Detection signals (in order):

1. Heavy markdown signature + no HTML semantics → machine-version
2. Schema.org Article/BlogPosting/NewsArticle → article
3. URL contains /blog/, /articles/, /news/, or date pattern (/YYYY/MM/DD/) → article
4. Root URL (/) → homepage
5. H1 + 3+ H2s + 5+ paragraphs → article (catches blog posts without schema)
6. Schema.org Product/Service/Offer → landing
7. Default → landing or other

Rules that filter out for non-article types:

• auth.byline — only article (homepages don't have authors)
• auth.publish-date — only article (homepages don't have publish dates)
• auth.outbound-citations — article + machine-version + other (marketing pages rarely cite externally)
• fit.answer-capsules — only article (homepages don't have H2-per-section structure)
• struct.paragraph-length — article + other
• And so on — see RULE_APPLICABILITY in citebench/lib/types.ts.

Non-applicable rules are shown in the UI marked "N/A for this page type" so users understand why they're skipped, rather than being hidden silently.

What changed from v1 (and why)

Core principle (unchanged)

Deterministic + explainable. No LLM at audit time. Every score line has a clear "here's exactly why." Transparency is the positioning wedge.

v2 rubric: 100pt across 5 dimensions

1. Source & Authority — 25 pt

• Author byline + Person schema + credentials/sameAs (LinkedIn/ORCID/Wikidata) — 8 pt
• Outbound citations to authoritative domains (.edu, .gov, named sources) — 6 pt
• Organization schema with sameAs chain (Wikipedia, Wikidata, social) — 5 pt
• Visible publish/update date (datePublished, dateModified) — 3 pt
• Wikipedia/Wikidata entity match for primary subject — 3 pt

2. Factual Density — 25 pt

• Original research / proprietary statistic presence — 8 pt (hardest to detect; use heuristics for "in our study," "we surveyed N," named methodology references)
• Quantitative claim density (numbers/percentages/dates per 1,000 words) — 7 pt
• Named-entity density (NER-detected people, orgs, places) — 5 pt
• Penalty: AI-fluff phrase density (see fluff-wordlist-seed.md) — up to −8 pt
• Fatal: AI self-identifier phrases ("as an AI language model," "as of my last training") — automatic disqualification

3. Answer Engine Fit — 20 pt

• Semantic completeness — answer capsule under each H2 (40–160 word self-contained declarative blocks opening with the answer) — 10 pt (largest single sub-component in the rubric)
• TL;DR / summary block in top 200–300 words — 4 pt
• FAQ section + FAQPage schema — 3 pt
• Lists/tables for enumerable content — 3 pt

4. Structural Citability — 20 pt

• Schema stack present (Article + FAQPage + BreadcrumbList + Organization + Person + topic-specific) — 8 pt
• Clean H1/H2/H3 hierarchy (one H1, no level jumps) — 4 pt
• Paragraph length appropriate (avg 40–80 words, no wall-of-text) — 3 pt
• Descriptive title tag with primary question/topic — 3 pt
• Internal linking with descriptive anchors (3+ contextual internal links) — 2 pt

5. Technical Cleanliness — 10 pt

• HTTPS — gate (fatal if missing)
• Renders core content without JS — 3 pt
• robots.txt does NOT block ChatGPT-User / Claude-User / Perplexity-User (live retrieval) — 3 pt (−5pt if blocked, NOT fatal)
• robots.txt does NOT block GPTBot / Google-Extended / ClaudeBot / CCBot (training) — 2 pt (−2 to −3pt if blocked, NOT fatal)
• TTFB <1s + LCP <2.5s + no crawl-time errors — 2 pt

Output the user sees (sample)

CITEBENCH SCORE: 73 / 100

✅ Source & Authority: 20 / 25
   • Author byline + sameAs: 8/8
   • Outbound .edu/.gov citations: 4/6
   • Org schema + sameAs: 5/5
   • Visible date: 3/3
   • Wikipedia/Wikidata entity match: 0/3 ← missing

⚠️  Factual Density: 14 / 25
   • Original data: 0/8 ← missing
   • Quantitative claim density: 5/7
   • Named-entity density: 5/5
   • AI-fluff penalty: −4 pt (detected: "delve into," "in today's fast-paced," "leverage")
   • Self-identifier: clean

✅ Answer Engine Fit: 15 / 20
   • Answer capsules: 7/10 (3 of 5 H2 sections have capsule structure)
   • TL;DR present: 4/4
   • FAQ + schema: 0/3
   • Lists/tables: 4/3

✅ Structural Citability: 16 / 20
✅ Technical: 8 / 10

Top 3 fixes (sorted by score lift):
1. Add original data or proprietary statistics in the first 600 words (+8 pt est)
2. Add FAQ section with FAQPage schema (+3 pt) + complete answer capsules in remaining H2s (+3 pt)
3. Add Wikipedia/Wikidata sameAs to Organization schema (+3 pt)

What Citebench cannot score (honest scope statement)

A URL-level audit cannot measure the strongest predictors of AI citation, which are off-page:

• Off-page brand mentions across the web (Ahrefs: r=0.66–0.74 with AI visibility)
• YouTube mentions (r=0.737 — strongest single signal Ahrefs found)
• Reddit presence (46.7% of Perplexity citations are Reddit)
• Cross-platform brand consistency

This is a known scope gap. Future versions could add brand-level scoring via search APIs; v2 stays URL-focused and is honest about it.

Platform divergence (informational)

The v2 score is an average expected citability across major engines. But:

• ChatGPT favors Wikipedia (47.9% of top-10), encyclopedic editorial
• Perplexity favors Reddit (46.7%), freshness, per-claim attribution
• Google AIO favors UGC + YouTube + multi-modal content, freshness
• Gemini favors .gov/.edu/institutional (26%)
• Only 11% of domains are cited by both ChatGPT and Perplexity

A perfectly Citebench-optimized URL likely won't be cited equally everywhere. The score targets the on-page features that can matter across all four. Future improvement: per-platform sub-scores.

Confidence and caveats

• Ahrefs' 75k-brand study and BuzzStream's 4M-citation analysis are the most defensible empirical anchors.
• Wellows' r=0.92 for multi-modal is implausibly high — treat individual r-values as directional, not literal.
• Most "studies" are vendor research without disclosed methodology. v2 weights are evidence-informed best guesses, not regression-derived coefficients.
• All claims sourced in methodology-research.md.

Implementation notes (carry into build phase)

All rules executable with:

• HTTP fetch + Cheerio HTML parsing
• DOM traversal for structure + paragraph stats
• JSON-LD schema parsing
• Wordlist matching for fluff bigrams (see fluff-wordlist-seed.md)
• robots.txt fetch + parse for crawler block check
• Open-source NER (spaCy/transformers.js) for entity density — runs in browser/Edge or via Vercel function
• External lookups for Wikipedia/Wikidata entity match (Wikipedia API, free, rate-limited)

Audit completes in <5 sec per URL. Free to run.