← Docs·Core Concepts

Understanding the AI Score

The 47-point rubric used to grade AI-readiness, weighted by empirical citation-lift impact.

4 min read

title: "Understanding the AI Score" slug: "understanding-ai-score" description: "The 47-point rubric used to grade AI-readiness, weighted by empirical citation-lift impact." group: "Core Concepts" order: 21

The AI Score is a single 0–100 number that tells you how likely your site is to be picked as a source by a language model. It's calculated from 47 weighted checks, grouped into four pillars. Every weight is derived from regression against real citation data — not opinion.

The four pillars

PillarWeightWhat it captures
Technical readiness20%Can a crawler reach, parse, and index your site.
Content extractability40%How well your content survives being chunked and cited.
Schema coverage20%Which structured data types you expose, and how cleanly.
Authority & freshness20%Whether external signals suggest you should be trusted.

The content pillar is the heaviest for a reason: in our 2026 cohort of 10,000+ sites, extractability improvements moved citation rate roughly 3× more than any other lever.

The 47 checks

Below is the full rubric with weights. Pass/warn/fail is determined per check; partial credit is allowed on spectrum checks (e.g. chunk density).

Technical readiness (9 checks · 20 points)

  1. robots.txt allows AI crawlers (2 pts)
  2. llms.txt present and well-formed (3 pts)
  3. llms-full.txt present or N/A (1 pt)
  4. Sitemap XML present and submitted (2 pts)
  5. Canonical tags set and consistent (2 pts)
  6. Mobile parity — same content on mobile as desktop (2 pts)
  7. Core Web Vitals in green (3 pts)
  8. Server response time under 400 ms TTFB (2 pts)
  9. No crawl-blocking JavaScript on key pages (3 pts)

Content extractability (19 checks · 40 points)

  1. Chunk density — at least one extractable 100–400 token span per 500 words (4 pts)
  2. Clear claim sentences — short, declarative, brand-included (4 pts)
  3. Unique claims — non-duplicate against a 1B-page reference corpus (4 pts)
  4. H1 is a specific answer, not a brand tagline (2 pts)
  5. H2/H3 hierarchy clean and semantic (2 pts)
  6. Dates on articles, prominently (2 pts)
  7. Author byline with verifiable bio (2 pts)
  8. FAQ-style blocks with direct Q/A (3 pts)
  9. Lists used for enumerations (1 pt)
  10. Tables for structured comparisons (2 pts)
  11. Blockquotes for citable claims (1 pt)
  12. Image alt text is descriptive, not keyword stuffing (1 pt)
  13. No content hidden behind tabs/accordions (2 pts)
  14. No content rendered only in iframes (2 pts)
  15. No content locked behind JavaScript tests (1 pt)
  16. Internal link graph — every key page reachable in ≤3 clicks (2 pts)
  17. Anchor text is descriptive (1 pt)
  18. Content freshness — updated or re-dated in the last 12 months (2 pts)
  19. Word-count appropriate — 400–1200 for answers, longer for pillar pages (2 pts)

Schema coverage (10 checks · 20 points)

  1. FAQPage on at least one high-intent page (4 pts)
  2. Product schema on commerce pages (3 pts)
  3. Review + AggregateRating on product pages (2 pts)
  4. HowTo schema on step-by-step guides (3 pts)
  5. Article schema on blog posts (2 pts)
  6. Organization with sameAs links (1 pt)
  7. BreadcrumbList on internal pages (1 pt)
  8. SoftwareApplication on app/product pages (1 pt)
  9. Schema validates against Schema.org + Google Rich Results (2 pts)
  10. No duplicated schema blocks (1 pt)

Authority & freshness (9 checks · 20 points)

  1. Referring domains — count and diversity (3 pts)
  2. Citations from Reddit / HN / SO / Wikipedia (4 pts)
  3. Recent changes to site — evidence of ongoing maintenance (2 pts)
  4. Named-entity linkage — your org on Wikipedia, Crunchbase, LinkedIn (2 pts)
  5. Social proof on-site (logos, testimonials, review counts) (2 pts)
  6. External G2 / Capterra / Product Hunt presence (2 pts)
  7. Press coverage — crawlable mentions in reputable publications (2 pts)
  8. Owner-verified Knowledge Panel presence (1 pt)
  9. Consistent NAP (name/address/phone) across the web (2 pts)

How weights are set

Every three months we re-fit the weights. The procedure:

  1. For each of ~3,000 tracked sites, capture current score components and 30-day citation rate.
  2. Run a gradient-boosted regression predicting citation rate from the 47 check scores.
  3. Derive per-check importance. Clip to the existing pillar-weight envelopes (so the product experience stays consistent).
  4. Publish the diff in the changelog when weights shift by more than 10%.

Reading your score

  • 85+: Top 10%. You're a default citation for your category.
  • 70–84: Solid. You show up reliably but probably miss some long-tail queries.
  • 50–69: Average. Two or three high-lift fixes usually move you twenty points.
  • Below 50: You're invisible to AI right now. Start with the top five fixes in your dashboard — they're ordered by expected delta.

Why a score, not a "rank"

There's no real ranking to measure against. Two answer engines asked the same question give different answers. What you can measure reliably is your own site's readiness to be cited, and the week-over-week delta. The score is a proxy — tuned against citation outcomes — rather than a leaderboard position. We think that's a more honest way to work.