Skip to content
RookAtlas

How RookAtlas measures

What every number is built from, how its range is worked out, and what it can't tell you. Each stored metric carries its methodology version, so a change in method never passes for a change in visibility.

Methodology m3, October 2026. RookAtlas opens soon: this is the method it opens with, the provider pilot may refine it, and any change gets a new methodology version.

The eight engines

Eight AI engines are asked the same prompts on the same day. Their answers combine into one visibility number, and each engine also gets its own. ChatGPT, Gemini, Perplexity, Google AI Mode and Google AI Overviews are read logged out from their own interfaces. Claude, Grok and Muse are asked through their makers' official APIs with web search enabled, for the reasons below, and their answers can differ from the apps'. You choose which engines count toward the combined number. Here is how each one is asked.

EngineHow answers are collected
ChatGPTThe ChatGPT app's own interface, logged out, through a contracted data provider.
GeminiThe Gemini app's own interface, logged out, through a contracted data provider. Never through the Gemini API, whose terms forbid analysing grounded answers.
PerplexityPerplexity's own interface, logged out, through a contracted data provider.
AI OverviewsGoogle's results page with the AI Overview loaded. A search that shows no Overview is recorded as such.
AI ModeGoogle's AI Mode, logged out, through a contracted data provider.
ClaudeAnthropic's API with web search, at most one search per answer. We know of no provider that collects the Claude app.
GrokxAI's API with web search enabled and X search off, until a provider for the Grok app proves reliable.
MuseMeta's Model API with web search enabled, using the standard Muse Spark model (not the variant Meta may train on). The Muse app needs an account, so it is not collected.

Every answer records its engine, how it was collected, the provider, the model where the provider reports it, the market (a country and a language), when it was collected, and what it cost. Changing how an engine is collected starts a new series instead of joining the old one.

Pilot, 12 November 2026. The first live runs asked 100 Turkish and 100 English prompts on each of ChatGPT, Gemini, Google AI Mode and Google AI Overviews. All 800 collection tasks completed with no failures, and Google showed an AI Overview for 96% of the Turkish and 95% of the English searches. These runs test collection, not how accurately mentions are read. Perplexity, Claude, Grok and Muse are still to be tested.

What counts as a mention

An answer names a brand when it contains the brand's name or one of its aliases as whole words. Matching folds case and accents with Turkish-aware rules, so İ, ı, I and i are compared correctly. Position is the order of the tracked brands in the answer.

A citation counts when the answer shows the link, inline or in its source panel. Pages an engine retrieved without showing them are reported apart, as retrieved but not cited. Links are resolved and normalized before they are counted, and each domain counts once per answer.

Raw answers are kept. Each reading of an answer records the extractor and the brand context it used, so a better extractor reads old answers again as a new version rather than relabelling them. An AI-based extractor joins only after it is measured against hand-labelled answers for precision, recall and agreement.

The metrics

Visibility
Answers that name the brand, out of every answer on every engine. The headline number, also shown engine by engine.
Share of voice
Mentions of the brand, out of mentions of every tracked brand. It depends on which competitors are tracked, so reports list them.
Average position
The brand's mean place among tracked brands, only in answers that name it. Always shown beside visibility, never alone.
Cited
Answers that show a link to one of the brand's own domains, out of all answers.
AI Overviews shown
Searches where Google showed an AI Overview, out of all AI Overviews searches. Searches without one leave the AI Overviews visibility count.

Intervals and confidence

Visibility, the cited rate and the other shares of answers carry a 95% interval written as a range, such as 31–65%, because the interval is not symmetric around the estimate. It is the range of rates consistent with the answers collected, given how many there are and how they cluster by prompt. Share of voice and average position carry a 95% interval too, with prompts as clusters; an average position whose ranks don’t vary has none.

Answers to the same prompt tend to agree with each other, so they are not counted as independent. Prompts are treated as clusters: the effective number of answers comes from the cluster-robust variance, and a Wilson interval is drawn on it. Ten repeats of one prompt count for far less than ten different prompts.

Low confidence
Fewer than 30 effective answers, or fewer than 10 prompts however many answers they have. The rate is shown, marked low confidence.
Too little data
Fewer than 10 effective answers, or fewer than 3 prompts. The rate gives way to the words "too little data", with the estimate in small print.
Differences
A difference between two brands is highlighted only when their intervals don't overlap.
Changes over time
A change between periods is computed on the prompts asked on the same engines in both, each prompt on each engine counted once, so a prompt answered more often in one period can't move it. From 10 such prompts it is tested: when the 95% interval of the difference, clustered by prompt, excludes zero, it is more than run-to-run noise, and it counts as a real change when it holds for two periods in a row. Share of voice and average position take the same test on answers measured against the same tracked competitors, since adding or removing one moves them on unchanged answers, and a position change counts only prompts that named the brand in both periods. With fewer prompts, a change is shown as observed only.

An interval describes sampling within the prompts you chose. It says nothing about whether those prompts represent every buyer, or about mistakes in collection. Width comes from breadth: more distinct prompts and more days narrow it faster than asking the same prompt again. Telling a 10-point change apart from noise near 50% takes about 390 effective answers in each period.

Prompts, versions and trends

Prompts start from suggestions based on the brand, its category and its competitors, and a person approves each one. Each prompt belongs to one market and carries its intent. Branded prompts, which name the brand in the question itself, are marked as such, since they name it by design.

Prompt wording, the set of prompts and the list of competitors are versioned. An edit creates a new version, and every run records the versions it used, so a trend compares like with like or shows a break with its reason. Every report lists the prompts behind its numbers, in the wording they were asked.

Evidence grades

Every suggested change carries a grade for how strongly the evidence says it moves AI visibility. A grade describes the evidence, not a promise of results, and weak items never count toward a score.

Official
The engine's maker documents it.
Experiment
Shown in controlled tests, though results vary by engine.
Correlation
Seen alongside visibility in large studies; not proven to cause it.
Vendor data
Reported by a tool vendor from its own logs or tests.
Weak
Anecdotal; kept out of every score.
Prompt gap
A prompt with at least 6 answers where the brand's interval lies wholly below the most-named competitor's. Graded experiment: adding sources, statistics and quotations raised visibility in controlled tests, with results that vary by engine.
Source gap
A third-party site shown as a source in at least 3 answers that name a competitor but not the brand, where the brand appears in under half of the answers citing it. Graded correlation: brand mentions across the web track AI visibility in large studies, without proof of cause.
Site access
Robots rules for the crawlers each engine's maker documents, read as the standard defines them, and noindex or snippet limits. Graded official. A page whose served HTML has little text but loads scripts is graded vendor data, from crawler logs a vendor published.

What a credit buys

One credit runs one prompt, in one market, once, on every surface in the run: all eight unless fewer are chosen. A prompt is charged only when at least one surface was collected, and a search that showed no AI Overview counts as collected. Surfaces that fail are retried for free within the run's 24-hour window; those still failing are reported as missing and leave the counts. Every run shows its price before it starts.

What it can't tell you

  • What a signed-in user with a long chat history sees. Every answer is collected without a personal history, as a new visitor would get it.
  • Exactly what the Claude, Grok and Muse apps show. Their answers come from each maker's API with web search enabled, which can differ from the app.
  • Your whole market, from a set of prompts. The prompts you track are a sample of questions; choose them as you would a survey.
  • That a change will work. Grades say how good the evidence is; before-and-after comparisons are observations, not proof.

Answers come only from contracted data providers and official APIs, through each provider's normal interface. Suggested changes stay white-hat: real content, earned mentions and honest outreach.

RookAtlas opens soon, with live measurement.