Only questions with a commercial intent: the questions people ask when they’re looking for a product, a service or a supplier. For example, “best CRM for a small sales team” or “which running shoes for flat feet”. The panel covers both B2B and B2C. Purely informational questions, such as “what is a CRM?”, are excluded, because they rarely lead an AI to recommend brands, and brand recommendations are what this barometer measures.

A topic is a concept that matters in your industry: a product feature, a use case, a buying criterion, a technology, a standard. For a CRM, that could be “pipeline management”, “email integration” or “GDPR compliance”. Topics are detected automatically with Waikay’s language-analysis technology (from InLinks), which recognises named entities in text: concepts, products and organisations that refer to something specific, rather than ordinary words.

Each question we track belongs to a project, and each project has its own knowledge graph: a map of the entities that matter for that website and its market. When a topic is detected in an AI answer, it’s checked against that project’s knowledge graph. Topics unrelated to the project’s industry are set aside. This keeps irrelevant terms out of the counts, so the topic figures reflect the subjects that actually matter for the brands being monitored.

Not on its own. “Topics per answer” measures how broad an answer is: how many industry topics it covers. It doesn’t say which brand each topic is associated with. Knowing whether AI links your brand to your key strengths requires a brand-level analysis, which Waikay provides for monitored projects.

We compare matched questions collected from the provider's developer API and observed consumer-interface sessions. Comparisons use the same prompt on the same day. The two methods may use different backend models, settings or execution conditions. Neither is automatically definitive. Scraped responses reflect the observed sessions, not necessarily what every user sees, and brand mentions are not necessarily recommendations.

For each matched API and scraped answer pair, overlap is the mean of two proportions: (shared brands divided by API brands) and (shared brands divided by scraped brands). The weekly indicator averages these values across pairs. It does not measure the share of identical answers or the proportion of unique competitors in the entire union of all answers.

The weekly directional calculation is shown when the required matched brand averages are available.

The estimate is 100 × (1 − average shared brands per matched answer pair ÷ average API brands per answer). Reversing the denominator gives the opposite direction. These averages must use the same matched pairs. Because exported averages are rounded, the result is approximate. It describes appearances absent from corresponding answers, not the percentage of unique competitors missing from every result.

The illustration uses the selected week's average brand counts.

Results reflect the prompts, websites and responses included in each monitoring dataset. The public JSON summary does not provide a verified market-by-market, language-by-language or industry-by-industry breakdown. These results should not be assumed to represent every market, language, industry or ChatGPT/Gemini user.

Weekly movements and their meanings appear when data loads.

Each icon marks a documented browser-model default or routing change. Hover, tap, click or focus the icon for model, rollout date, audience and source. Rollouts can vary by account, and markers do not prove that a particular response was produced by that model or that a chart movement was caused by it.

The shaded period is the current incomplete week. It is preliminary and excluded from closed-week headline figures. Recently closed weeks can remain provisional while data settles.