How BrandWater AI measures AI visibility.

• Evidence over decoration: every finding links to an observable answer, prompt, source or measurement.
• Defined denominators: a rate always says what it is a rate of.
• Sample sizes are always shown.
• Measured results and interpreted labels are kept visibly apart.
• Nothing implies access to private model data or hidden reasoning.

An observation is one stored answer to one prompt on one AI platform at one point in time, saved in full together with any visible sources. Every rate BrandWater AI reports is built from observations, so any figure can be traced back to the exact answers behind it.

Measurement runs on a defined prompt set built from your topics and customer intents, written in the way customers ask, including regional languages. The set is kept stable and extended by adding prompts, so a change in results reflects the market, not a change in questions.

AI answers vary between runs, models, locations and accounts. Each prompt is therefore run repeatedly on each platform, and results are reported as rates with the number of observations shown. A single answer is treated as an anecdote, never as a measurement.

Rates use eligible observations as the denominator: answers relevant to the brand and topic being measured. Dividing by an arbitrary total would make results look better or worse than they are, so eligibility rules are fixed before a run and applied consistently.

Counts and rates come from stored observations. Labels such as sentiment or framing, and any AI generated explanation, are interpretations. They are shown as interpretations and always link to the answer behind them.

Brand Intelligence records each fact about a brand with its source, a confidence level and a review status of confirmed, discovered or conflicting. Facts are reviewed by the customer before they are used to identify omissions or to shape recommendations.

At launch the Agent only recommends, and you carry out any change yourself. Coming in V3: every action requires human approval, and after a change the affected prompts are rerun and compared with the earlier baseline. Results will be labelled verified, inconclusive or pending, and describe association, not proof of cause.

• Platforms change their behaviour without notice.
• Some platforms do not expose sources, so citation analysis is not possible on them.
• Traffic share figures used in market data are website visits only.
• Results describe a sample of prompts at a point in time, not everything people ask.

Published statistics on platform usage come from company statements and named third parties such as Similarweb, Sensor Tower and Anthropic. Each figure is dated and linked on the statistics page. Where no reliable figure exists we say so instead of estimating.