Mention rate
The share of valid in-scope answers containing the target brand or a verified alias.
Every conclusion should trace back to its sampling conditions and its raw evidence.
This note covers one cycle of GEO monitoring, from freezing scope and designing prompts through controlled-browser collection, screenshot evidence and source review to re-measurement. Each platform uses 50 prompts every 72 hours; each prompt is executed once per platform by default.
Version 1.1 · Last updatedSix steps that leave a checkable record
- Freeze the scope: confirm the brand entity, website, products, markets, languages, competitors and the goal of this cycle.
- Build the prompt map: freeze 50 prompts per platform: 44 unbranded discovery/comparison questions and 6 branded fact checks.
- Collect on a fixed browser cycle: follow the normal consumer product-page path every 72 hours and execute each prompt once per platform by default, retaining platform, entry point, time, region, language and session conditions.
- Archive the raw evidence: keep the answer text, page screenshot, mention position, cited URLs, source records and anomalies — not just a summary.
- Compute on a fixed basis: use one consistent denominator, de-duplication rule and missing-value rule within a report, and keep observation separate from interpretation.
- Act and re-measure: give each issue an owner, an acceptance standard and a re-check date, then compare on exactly the same basis.
Metrics, not a mysterious overall score
Recommendation position
Whether the brand appears as a candidate or a suggestion, kept with its context. This is not a platform’s official ranking.
Factual accuracy
Verifiable claims in an answer compared item by item against confirmed facts, excluding statements that cannot be judged.
Citation coverage
Whether an answer gives a resolvable source, separated into official, independent third party, competitor, aggregator and wrong-entity.
Source diversity
Sources de-duplicated by resolvable domain, to show the shape of the source mix rather than treating citation count as quality.
Like-for-like change
A difference is read as a trend only when scope, prompts, execution conditions and metric version are all comparable.
Every claim states what it rests on
- Independently verified: an applicable regulatory, certification, inspection or other independent source matches on entity, product, time and scope.
- Officially published: observed on the brand’s own channels, and worded accordingly — “published on the company website” or “stated by the company”.
- Client-confirmed: confirmed explicitly by the client. Usable in the project’s fact base, never presented as independent third-party proof.
- Unverified claim: carried as a question to be checked, never written into a conclusion or a prompt as unqualified fact.
A sample observation is not the truth about a platform
Generative AI output can change with time, location, language, account, session history, model version and plain randomness. Citation panels can also change, collapse or become unavailable. So a report describes what was observed within a stated sampling scope. It is not extrapolated to whole-platform coverage, and it promises no fixed ranking, indexing, traffic or revenue.