How is a Citebook score computed?
A Citebook score is a weighted sum of six components, each scored 0–100 per page, rolled up across pages weighted by page priority. The weights below are the weights — they live in the codebase as constants¹, the scoring function is pure, and no language model ever sees or adjusts a score. Changing a weight is a product decision with a changelog entry, not a tuning knob.
Every component carries an evidence tier. Tier A (●) means published research supports a causal or strongly predictive link to AI citation. Tier B (○) means the evidence is correlational or mixed, and the weight is set accordingly low. We would rather under-claim than over-weight.
| Component | Weight | Tier | What it measures |
|---|---|---|---|
| Retrievability | 20 | ● A | Bot access, server-side rendering, status codes. A gate: nothing else matters if engines cannot read the site. |
| Claim density | 30 | ● A | Statistics, attributed quotes, and externally sourced claims per 500 words. Highest-weight component. |
| Structure | 20 | ● A | Question-phrased headings, answer-first passages, and self-contained extractable chunks. |
| Freshness | 15 | ● A | Visible published/updated dates, recency, and absence of stale in-body references. |
| Schema validity | 10 | ○ B | Presence and validity of JSON-LD. Correlational evidence only. |
| Entity consistency | 5 | ○ B | Brand-name consistency, Organization schema, and named author/founder entities. |
Why can a score never exceed 35 on a retrievability failure?
Because nothing else matters if the engines cannot read the site. When AI crawlers are blocked in robots.txt, or pages are client-side-rendered blanks that serve near-empty documents to non-JavaScript crawlers, the audit caps the total score at 35 and the finding leads the report — regardless of how well the content would otherwise score. Fixing retrievability first is the highest-return action such a site can take, and the score is built to say so bluntly.
What do we deliberately not score?
llms.txt. Reported when present or absent, never scored, never sold as a fix. The largest available correlation study found no relationship with citations², and Google’s guidance says you don’t need it⁴. We’ll add the file for free if you want it; charging for it is selling a placebo.
Anything we cannot verify. Fixes that inject statistics or quotes carry publisher, date, and URL, and each claim is confirmed against its source before it reaches you. Anything unverifiable is flagged and excluded — verification is blocking, not advisory.
How is engine data handled?
The current audit examines your site: crawler access per engine (GPTBot, OAI-SearchBot, PerplexityBot, Google-Extended, ClaudeBot), rendering, structure, claims, dates, and schema. Where we sample engine answers in future product tiers, that data will be API-sampled and labelled as such — API answers approximate what consumers see; they don’t replicate any individual user’s session, and we will say so in every report that contains them. Honest sampling beats confident fiction.