Kuroma
How Kuroma scores AI Readiness
Rubric version 2026-07-08. This page renders from the same definition the scoring engine runs, so what you read here is exactly what we ship.
Every factor below carries a verdict and the research behind it: controlled studies where they exist, large-scale observational data where they do not, and official platform guidance where vendors have spoken. When the evidence says a popular tactic does not work, we down-weight it and say so.
What this score can and cannot see. This audit measures citation readiness: how easily AI engines can crawl, extract, and quote these pages. It cannot see the off-site brand footprint (mentions, reviews, third-party coverage) that gates whether AI engines retrieve a brand at all; Kuroma's visibility scans and off-site signals track that side.
Validated against observed AI answers
We regressed observed citation outcomes from our own visibility scans (201,695 AI answers collected over 22 weeks) on the audit factor scores each brand held BEFORE those answers were generated. Validation is out-of-sample on a strict time split: the model is fit on earlier weeks and scored only on later weeks it never saw. The metric is AUC, the probability the model ranks a cited case above an uncited one; 0.5 means no better than chance, 1.0 means perfect ranking.
| Engine | Held-out AUC (later weeks, never seen in training) |
|---|---|
| ChatGPT | 0.84 |
| Grok | 0.76 |
| Claude | 0.68 |
| Google AI Overviews | 0.63 |
| Perplexity | 0.61 |
| Google AI Mode | 0.61 |
Corpus: 201,695 AI answers, 869,783 extracted citations, 22 weeks, validated 2026-07-07. Fitted cohort: 17 brands.
Gemini is excluded from this table for an honest reason: across our whole corpus it almost never cites a brand’s own domain (it cites retailers, press, and review sites instead), so "was the brand’s domain cited" is the wrong outcome to grade it on. We are building a third-party-coverage label for it.
- AUC measures ranking power of the factor set as a whole. It does not license causal claims about any single factor, and we do not reweight the rubric from these coefficients.
- The validated audits come from the March to June 2026 engine, whose measurement core the current rubric inherits with evidence-based reweighting. The current rubric accrues its own validation cohort every scan week.
- The split validates prediction across time for brands the model has seen, not prediction for brand types it has never seen. The cohort is 17 brands and growing; we will publish updated numbers as it does.
On-Page Content: 25% of the overall score
| Factor | Weight in category | Status | Why (with evidence) |
|---|---|---|---|
| Content Depth & Comprehensiveness | 26% | Evidence-backed | Deep, comprehensive pages are far more likely to be cited than shallow ones, and missing information is a leading diagnosed cause of zero-citation failures. Coverage of the topic matters, not raw length. Cross-engine citation-driver study, deep vs shallow content and freshness gatekeeping (arXiv 2605.25517) · Zero-citation failure taxonomy: contextual gap, intent divergence, information scarcity (arXiv 2603.09296) · Google Search Central guidance on AI features (no special schema needed, no ideal page length) |
| Question Coverage & FAQ Presence | 26% | Evidence-backed | Covering the sub-questions users actually ask is the single strongest diagnosed lever: contextual gaps and intent divergence explain most zero-citation failures. FAQ formatting by itself is not the lever; answering the questions is. Zero-citation failure taxonomy: contextual gap, intent divergence, information scarcity (arXiv 2603.09296) · Google Search Central guidance on AI features (no special schema needed, no ideal page length) |
| Structured Data Quality | 18% | Rescoped to the evidence | Rescoped to extractable data structures in the HTML itself (tables and lists), which measurably improve machine extraction. This is distinct from schema.org markup, which shows no AI-citation lift in controlled tests. GEO: Generative Engine Optimization (arXiv 2311.09735) · Ahrefs controlled schema test (1,885 treated pages vs matched controls) |
| AI Readability Score | 12% | Evidence-backed | Engines preferentially select easier, clearer text, and fluency rewrites improve citation rates in controlled sandboxes. A moderate, causal signal. GEO: Generative Engine Optimization (arXiv 2311.09735) |
| Heading Hierarchy | 11% | Evidence-backed | Heading-scoped sections match how retrieval systems chunk pages, so clean hierarchy helps at the retrieval and chunking stage even though it is neutral at the answer-generation stage. Anthropic contextual retrieval (self-contained passages cut retrieval failures) · GEO: Generative Engine Optimization (arXiv 2311.09735) |
| Media Accessibility | 7% | Rescoped to the evidence | No study shows alt text lifts AI citations, and major AI fetchers request no images. Kept at reduced weight for one honest purpose: a text alternative must exist for content that is locked inside media. Vercel and MERJ AI crawler study (569M GPTBot requests: zero JavaScript execution, no image fetches) |
Technical Optimization: 20% of the overall score
| Factor | Weight in category | Status | Why (with evidence) |
|---|---|---|---|
| Schema Markup Completeness | 5% | Near-zero weight for AI (kept for classic search) | Controlled tests show no AI-citation lift from schema.org markup, and Google states no special structured data is needed for AI features. Near-zero weight is kept only for classic rich results and Bing grounding. Ahrefs controlled schema test (1,885 treated pages vs matched controls) · Google Search Central guidance on AI features (no special schema needed, no ideal page length) |
| Page Performance Indicators | 10% | Deliberately down-weighted | Speed correlations with AI visibility are weak and driven by extreme outliers; no platform names speed as a factor. Scored pass/fail: the fetch must succeed without server errors and respond within a sane time. Cross-engine citation-driver study, deep vs shallow content and freshness gatekeeping (arXiv 2605.25517) · Vercel and MERJ AI crawler study (569M GPTBot requests: zero JavaScript execution, no image fetches) |
| Mobile Optimization | 20% | Deliberately down-weighted | No AI-engine evidence exists for mobile signals; this is classic search hygiene (Google page experience) retained at a modest weight. Google Search Central guidance on AI features (no special schema needed, no ideal page length) |
| Crawlability & Accessibility | 65% | Evidence-backed | The strongest technical factor: blocking a visibility crawler verifiably removes a site from that engine's answers, and OpenAI documents this directly. Scoring uses a bot-class matrix so only visibility-bot blocks count against the score; training-bot blocks are a policy choice reported neutrally. OpenAI crawler documentation (blocking OAI-SearchBot removes a site from ChatGPT search answers) · Vercel and MERJ AI crawler study (569M GPTBot requests: zero JavaScript execution, no image fetches) · Bing webmaster documentation: NOARCHIVE prevents content from being used in Copilot responses and grounding |
Authority Signals: 25% of the overall score
| Factor | Weight in category | Status | Why (with evidence) |
|---|---|---|---|
| E-E-A-T Signals | 28% | Rescoped to the evidence | No study tests author bios as a causal AI-citation lever; the real authority channel is off-site brand footprint. On-page bylines and dates are kept as best-practice hygiene at reduced weight. Cross-engine citation-driver study, deep vs shallow content and freshness gatekeeping (arXiv 2605.25517) · Google Search Central guidance on AI features (no special schema needed, no ideal page length) |
| External Citations & References | 36% | Deliberately down-weighted | Citing sources helps in controlled sandboxes but the effect shrinks toward zero in competitive replications, so claims are capped and the weight stays modest within the category. GEO: Generative Engine Optimization (arXiv 2311.09735) |
| Trust & Brand Signals | 36% | Deliberately down-weighted | On-page trust badges are untested as a class; the evidenced trust channel is off-site reviews. Retained for basic legitimacy signals (contact, policies, HTTPS) that remain sensible hygiene. Cross-engine citation-driver study, deep vs shallow content and freshness gatekeeping (arXiv 2605.25517) |
AI Readiness: 30% of the overall score
| Factor | Weight in category | Status | Why (with evidence) |
|---|---|---|---|
| Answer Box Potential | 24% | Evidence-backed | The real mechanism is passage self-containment: heading-scoped sections that stand alone match retrieval chunking, and fixing context-severed passages measurably cuts retrieval failures. Anthropic contextual retrieval (self-contained passages cut retrieval failures) · GEO: Generative Engine Optimization (arXiv 2311.09735) |
| Citation-Worthiness | 25% | Evidence-backed | Verifiable evidence units (statistics, quotes, sourced claims) causally improve citation in sandboxes, and confident wording is preferred over hedged wording. Effects shrink in competitive settings, so claims stay capped. GEO: Generative Engine Optimization (arXiv 2311.09735) · Cross-engine citation-driver study, deep vs shallow content and freshness gatekeeping (arXiv 2605.25517) |
| Content Freshness | 19% | Evidence-backed | Freshness is the only unanimous causal gatekeeper across engines in the 2026 evidence base, and AI-cited content skews measurably fresher at web scale. Google AI Overviews is the documented exception. Cross-engine citation-driver study, deep vs shallow content and freshness gatekeeping (arXiv 2605.25517) |
| Unique Value Proposition | 17% | Deliberately down-weighted | Unique-wording signals test near null; the real phenomenon is competitive redundancy, which a single-page crawl cannot measure deterministically. Weight halved accordingly. GEO: Generative Engine Optimization (arXiv 2311.09735) |
| AI Discoverability | 10% | Rescoped to the evidence | Rescoped to on-page machine-readability signals (semantic HTML, licensing clarity). The dead ai-plugin manifest check was removed and llms.txt stays informational only, because no major engine reads it. Vercel and MERJ AI crawler study (569M GPTBot requests: zero JavaScript execution, no image fetches) · Google Search Central guidance on AI features (no special schema needed, no ideal page length) |
| Raw HTML Availability | 5% | Evidence-backed | No major AI crawler executes JavaScript, so content that only appears after scripts run is invisible to AI search. This is among the best-evidenced technical findings in the audit. Vercel and MERJ AI crawler study (569M GPTBot requests: zero JavaScript execution, no image fetches) |
What we deliberately do not score
These checks appear in many AI-visibility audits. The best available evidence says they do not move AI citations, so scoring them would inflate work that does not pay.
| Not scored | Why not (with evidence) |
|---|---|
| llms.txt | Triple-null: the vast majority of llms.txt files receive zero crawler requests, Google states it ignores them, and no major engine reads them for citation. Detected and reported as informational only. Google Search Central guidance on AI features (no special schema needed, no ideal page length) · Vercel and MERJ AI crawler study (569M GPTBot requests: zero JavaScript execution, no image fetches) |
| Schema markup as an AI-citation lever | A large controlled test found no AI-citation lift from adding schema.org markup, and Google states no special structured data is needed for AI features. Schema is scored near zero, for classic rich results and Bing grounding only. Ahrefs controlled schema test (1,885 treated pages vs matched controls) · Google Search Central guidance on AI features (no special schema needed, no ideal page length) |
| Per-run AI rank tracking | Single-run AI answers are statistically unstable (same-day identical queries agree on citations only a few percent of the time), so a per-run rank is noise. Kuroma surfaces mention rates over repeated runs instead. Cross-engine citation-driver study, deep vs shallow content and freshness gatekeeping (arXiv 2605.25517) |
| Keyword density optimization | Keyword stuffing tests causally negative for generative engine citation and has been dead in classic search for years. Nothing in this audit rewards it. GEO: Generative Engine Optimization (arXiv 2311.09735) |
How should you read the score?
The score measures citation readiness, not predicted traffic. AI answers are probabilistic: identical questions produce different citation sets run to run, so no single number can promise placement. Readiness stacks the odds; Kuroma's visibility scans then measure the outcome across engines, and the two together tell you what to fix and whether it worked.