How CitedScore Works

How does CitedScore read your site?

CitedScore crawls your homepage first, then selects up to 15 additional pages worth checking, one representative per site section rather than every URL on the domain. The homepage is evaluated once, from the same context every other check reads, so nothing gets fetched twice or scored on stale content.

The crawler reads robots.txt before it fetches anything and honors every Disallow rule for User-agent: *. If a path is blocked, CitedScore never requests it. Every crawled page runs the same checks: schema validity, meta tags, heading hierarchy, canonical tags, HTTPS, sitemap presence, and image optimization. That sampled set is what produces the Technical SEO Score and the Sitewide Coverage Sample your Report shows.

How does CitedScore decide which questions to test?

Before a single platform is queried, CitedScore extracts a Company Profile from your site: pitch, category, offerings, price positioning, claimed differentiators, and the struggling moment your copy addresses. Nothing is invented. Every field is built only from signals actually present on the page.

From that profile it generates four buyer personas, one per stage of the buyer timeline, each carrying its own struggling moment, job to be done, and real search phrasing:

When a site's signals support only three distinct stages, CitedScore regenerates the full set once. If the second pass returns three again, the audit runs on three rather than inventing a fourth persona your site gave no evidence for. Three is the hard floor: below it the audit stops instead of guessing.

Each persona contributes three prompt runs. One question is asked twice, once in short conversational phrasing and once in long structured phrasing, as that persona's divergence pair. A second question is asked once. The pair exists because the two registers produce different AI answers, and a method that tested only one of them would miss half of what a buyer actually types.

A prompt is one question in one phrasing. An observation is what one platform returned for one prompt. The persona set is not the whole grid: eight further prompts run on every audit, four brand-neutral and four comparative, so the measurement is not built only from questions your own positioning suggested.

Which AI platforms does CitedScore test, and why those four?

ChatGPT, Google AI Overviews, Gemini, and Perplexity carry the overwhelming majority of buyer questions asked of AI today. Every audit checks all four: no platform add-ons, no per-platform pricing. They are not weighted equally, because your buyers are not evenly distributed across them.

Per-platform scores are always shown alongside the blend, because a brand can be strong on one platform and invisible on another, and an average alone hides that. Twenty prompts run on every platform, producing 80 result observations per audit. The full grid appears in your Report, not a summary of it.

One retrieval detail: Google publishes no public API for AI Overviews, so those results are retrieved through DataForSEO, the search results provider disclosed on the sub-processors page. The other three platforms are queried through their own APIs.

How are the three scores calculated and weighted?

GEO receives the largest weight because direct AI visibility is the outcome the Report exists to measure. The GEO Citation Score blends the four platform scores using the usage-share weights above, and each platform score is calculated as:

Mentions carry 70% because a named recommendation drives revenue. Citations carry 30% because a source link has value even without the brand name attached. When a brand is mentioned, its position in the answer sets the weight applied to that mention:

What is a ghost citation, and how does CitedScore detect one?

A mention names your brand in the visible answer text. A citation links your domain as a source. A ghost citation is a citation with no mention: the AI used your content, linked your URL, and never once said your name. Semrush and Kevin Indig's June 2026 research across 3,981 domain appearances found 61.7% of domain appearances were citations that never named the brand.

Every one of the 80 result observations is classified into exactly one of four states, using the same alias-matching logic on every platform:

The two failure modes need different fixes. A missing citation is an on-site problem: structure, schema, extractable content. A missing mention is a brand-recognition problem. CitedScore reports each lever separately, because treating them as one metric hides which one is actually broken.

How does CitedScore know a competitor is really a competitor?

Every AI answer names brands beyond yours, and a naive text extractor pulls in noise along with them: generic phrases, section headers, contact strings that only look like company names. Every deduplicated candidate runs through a dedicated validation pass before any of it reaches your competitor gap table, and that pass is a language model judging each candidate against the answer it came from.

The pass folds near-duplicate names into a single entity, drops candidates that are not real brands, and recomputes each result's mention position against the validated set, so the number beside a competitor's name reflects who was actually named rather than what a heuristic guessed. On failure, CitedScore falls back to the heuristic lists rather than risk publishing an unvalidated read.

Only validated competitors feed the gap analysis and the action plan. Owned family domains, channel partners, and reference or community sites are classified separately and excluded: they compete for a different kind of attention, and lumping them in would overstate your actual competitive gap.

How is share of voice measured?

Share of voice counts brands named, not words or sentiment. Your brand counts once per answer it appears in; every validated competitor named in that same answer counts once each. The count runs only across answers the platform actually returned, so a query that errored or never triggered a response is excluded from the denominator rather than counted as a zero.

Queries that name your brand directly (“who is X”, “is X legit”, “X reviews”) are excluded from this metric on every audit as a standing rule: they almost always name you and almost never name a competitor, so counting them would flatter the figure. Below a minimum mention threshold, share of voice is not reported at all rather than published as a misleading small-sample number.

The exclusion moves share of voice and nothing else. Those queries still count in full toward the GEO score, because a navigational query you fail to win is still a real visibility miss. Share of voice is diagnostic, never a scoring input.

What is the Maturity Ladder, and how is a rung earned?

A 0–100 score is precise, but it does not say what it means day to day. The CitedScore Maturity Ladder turns it into a five-rung read: your composite score proposes a rung, and your live AI results can pull it down from there.

Your composite score sets the candidate rung. From there, evidence caps can only lower it, never raise it: a strong Technical score cannot buy a rung your live AI results do not support. The ladder is a read of your evidence; nothing flows from it back into the score. Every Report shows the specific, evidence-backed gate to the next rung.

What evidence sits behind every finding?

Every score traces back to a specific, named piece of evidence: the live prompt that produced it, the platform that returned it, and the actual response text. Nothing in the GEO section is a projection or an industry average applied to your domain.

Findings carry a tier. What CitedScore measured on your own audit is stated as fact, because a named field on your run supports it. Where a recommendation is standard AI-visibility practice rather than something measured on your site, the Report says so, stated with conviction as practice and never dressed up as a number this audit did not produce.

The audit is engine-assisted and engineer-verified: a diagnostic engine runs the full breadth of every check, and every finding is verified by hand before it ships. The engine gives coverage a manual review cannot match; the review gives judgment a machine cannot provide.

What changes between runs

AI answers are not deterministic. The same prompt, on the same platform, on a different day, can name a different set of brands. Any measurement of these systems is a reading taken at a point in time, and a methodology that did not say so would be selling you a number wearing a costume.

This is why the audit is built on 80 observations rather than a handful. Every prompt runs across all four platforms, personas span the full buyer timeline, and the divergence pair asks the same question in both registers. The score is the pattern across that grid, not any single answer inside it. Results are aggregated across persona types and phrasing variants for the same reason: one query is an anecdote, eighty is a measurement.

What that does not do is make the number stable to the decimal. Re-run an audit a week later and individual rows will move. A rung change or a shift of several points is a real signal worth acting on. Two or three points in either direction is the system breathing. CitedScore does not yet publish a measured variance figure, and until it does, that is the honest ceiling on how finely you should read a single score. When we have run the same domains enough times to state a range, the range will be published here with the rest of the constants.

We ran this exact methodology on our own homepage and published the unedited result, including the fix it caught.