Share of AI Voice Benchmark With Confidence Intervals

Most share-of-AI-voice numbers are noise reported to two decimal places. Run the same fifty prompts against the same model an hour apart and a brand's mention share moves eight to fourteen points at one sample per prompt. Every figure in this benchmark therefore carries its sample size and an interval, and when the change between two measurements sits inside that interval the report says so instead of drawing an arrow. How the prompt set is built: From demand, not from a keyword tool. Search Console queries, on-site search logs, the questions your sales team fields in a first call, and the phrasing distributors use when they email you. The Spanish and English sets are built separately and are not translations of each other: Mexican buyers raise warranty and availability far earlier in the question than US buyers do, and a translated prompt set measures a market that does not exist. Sampling: Thirty runs per prompt per surface, timestamped, with the model version recorded, because a provider quietly shipping a new version is the most common cause of a movement that looks like a competitor doing something. Surfaces are the ones your buyers use, sampled from where they are. AI Overviews for Mexico is measured in Spanish from Mexican exit nodes, not from a US datacentre with a language parameter, and the gap between those two readings is routinely wider than the gap between you and the brand you asked me to benchmark against. What is counted: Whether the brand is mentioned. Whether it is cited with a link, which is a different and much rarer event. Where in the answer it lands, since the third brand named in a list is not the recommendation. Whether the claim attributed to you is actually yours — misattribution is common in pet nutrition and in toy safety, and a mention carrying a competitor's specification is a problem, not a win. And the domains the model does cite: in these categories that is usually a retailer, an association, a standards body or a review site, and knowing which four hold the answer is the finding that changes what you spend next quarter. What the report does not contain: A dashboard. A single composite visibility score. A monthly cadence — these numbers do not move fast enough to justify measuring them monthly, and a monthly report is an invitation to explain noise as progress. I measure on a schedule slow enough that a change means something, which in these categories is quarterly. What I will not do: Report a movement without attempting to attribute it; where the attribution is unknown, the report says unknown. Include prompts nobody asks because they make the numbers look better. Or benchmark against a competitor set chosen for reasons of pride rather than overlap — I propose the set from who actually holds the answers, and that list has surprised every client so far. Who should not buy this: Anyone who needs the number to be higher before a board meeting. And brands with nothing yet to measure: if you are absent from every surface the benchmark returns zeros, and you will have paid for what an afternoon of manual checking would have told you. Do the work first and measure after it has had a quarter to land. What you keep: Prompt text, raw per-run outputs, timestamps and model versions in a file you own, plus the script that produced them. The point of handing over the instrument is that you can re-run the benchmark without me, check my arithmetic, and hold the next measurement to the same method rather than to my summary of it.