How the numbers on a Rave report are produced, in plain words. Everything here is checkable against the answers stored with the report.
What we measure
Three rates, per brand, per week. Visibility Rate is how often an answer names the brand. Top 3 Rate is how often it names the brand among the first three options. #1 Pick Rate is how often the brand is the first one named. All three are plain counts over the same denominator: the answers the week planned to ask. There is no blended score, no index, and no weighting.
What we ask
Each brand has a set of questions built from its own market: what a buyer would type at each stage. The set is reviewed by the agency and then locked, so week to week the same thing is being measured. A locked set stays locked. If it changes, the change is visible and the trend is re-baselined rather than joined up across it.
How answers are captured
Answers come from the public web interface of each answer engine, logged out, with a fresh session for every question, so nothing is personalised and nothing carries over between questions. Claude is measured through its API, because it has no logged-out public interface. Google AI Overviews are read from Google search results: an overview only appears for some searches, so a search with no overview is recorded as a real reading, not as a failure.
We send no instructions with a question: no system prompt, nothing that asks an engine to list or rank businesses. A real person does not send one either, and asking for a list changes the answer.
Every answer is stored as it came back, with its sources and the date it was captured. The counts on a report can be re-derived from those answers by anyone holding the report.
The noise band
Every question is asked once on every engine, every week. Asked again, an engine does not always give the same answer, so every trend line carries a band, and a move inside the band is not claimed as a change. It is reported as flat, because that is what it is.
The band is real arithmetic, not a cushion: it is the 95% interval for the question and engine cells in the wave, using how often a repeated answer flips, from our own measured replicates. More questions and more engines make it narrower. A single engine on its own carries a much wider band than the combined line, which is why movement is claimed on the combined line and single engines are read as detail.
Models change, and we say so
Each trend line is pinned to one engine and one provider. When a vendor changes the model behind it, the readings either side are not comparable, so the trend is re-baselined and the brand page says where and why. Whatever model label a provider reports is stored as given. We never invent one.
Visits and leads, never added together
The optional site snippet records two separate things. AI-sourced visits are visits whose referrer or campaign tag names an AI assistant: measured, but only a floor, because assistants often send no referrer at all. AI-influenced leads are what a person typed when asked how they heard about the business: self-reported, and a different thing entirely. The two are reported side by side and never summed, because summing them counts the same person twice.
What this does not measure
- Training data. We measure what answers say, not what a model learned.
- Personalised answers. Everything is captured logged out. A logged-in person with history and memory may see something different, and we make no claim about that.
- Volume. We do not estimate how many people asked a question. The rate is how the engines answer, not how many asked.
- Paid placement, and any engine surface that is behind a login or not publicly reachable.
- Cause. A rate that moves outside the band is a real change. What caused it is a judgement, not a measurement.
Questions about a number
Every report is backed by stored answers with their dates and sources. If a number looks wrong, ask the agency that sent it for the answers behind it. If something here is unclear, write to [CONTACT EMAIL].