How BotSay measures your numbers
BotSay puts your questions to six AI search engines through their official services, reads each answer and reports five measures by week. Every figure states what it was counted out of. Mention rate also carries a likely range.
How does BotSay measure AI mentions of my brand? #
BotSay measures in four steps: ask, read, count and compare.
- Ask. Your set of questions is put to six engines, listed in Which AI engines BotSay asks.
- Read. Each answer is read for the brands it names, the order they come in, how each is described, and the links it cites.
- Count. The answers are counted into five measures: Mention rate, Share of voice, Average rank, Your pages cited and Sentiment. What the five measures mean defines each one.
- Compare. Each week is compared with the one before, and a change is called real only when it is bigger than chance would produce.
Each reading asks your set of questions once, in a random order, and moves on. Every reading asks your whole set, so any two readings compare like with like.
Where do the answers come from? #
Every answer comes from an AI engine's own service. Nothing is taken from signing in to a consumer app and copying what a screen shows an ordinary user.
Most engines answer you directly, the way a person would ask them. The two Google surfaces, AI Overviews and AI Mode, are read through a licensed search-data connection rather than a browser. They are two different surfaces and are never combined into one engine: a brand can appear in one and not the other.
An answer's links come in two kinds. A link is cited when the answer points the reader at it, and retrieved when the engine read the page without pointing at it.
When is a week's reading complete? #
A week is complete the day after your reading day, because the two Google answers come back a day later than the rest. The edition line on every screen names the next reading, and Setup, How we measure states your own schedule, for example "You read your questions once a week." When readings happen has the days.
An answer that failed and is retried can still move a figure a little after that. While a week is still being read, the edition line says so, and the figures that reading could still move are marked Provisional.
Why does BotSay report by week? #
How you show up moves from week to week, not from day to day, and a weekly reading keeps it affordable to read every question on every engine.
Each week stands on its own reading and its own likely range. BotSay never averages across weeks, so your figures over time show what was actually measured, not a blended trend. A week is named by its dates, for example Week of 14 to 20 September, and is cut in UTC.
If a reading is resumed or corrected, the whole week is worked out again from the answers on file, so a retry, a fix or a restart all land on the same result.
What is each figure counted out of? #
A rate is worked out only over the answers that came back. An answer that never arrived, was declined, errored, or has not been read yet is left out of the count. It is not counted as though you went unmentioned.
An answer here means one question answered by one engine. BotSay averages any repeat answers within a question first, so a question answered more than once never counts for more than one answered once.
The table shows what each measure is counted out of.
| Measure | Counted out of |
|---|---|
| Mention rate | Questions answered |
| Share of voice | Every mention of the brands you track |
| Average rank | The answers that name you |
| Your pages cited | Every link cited |
| Sentiment | Mentions that carry a score |
Which questions are counted? #
The five measures count the questions that do not name you. A question that names your brand would be answered with your name almost by definition, so counting it would lift your figures and leave your rivals' where they were.
Questions are classed as Brand (names your brand), Rival (names a rival you track) or Open (names no brand). Brand questions are left out. The rest count together.
Google decides for itself whether to show an AI Overview for a query. When it shows fewer, only the questions it did answer count towards Mention rate, so the rate can rise with nothing about your brand having changed. BotSay records how often each Google surface answered.
How does BotSay tell a real change from chance? #
BotSay calls a week-to-week change real only when it is bigger than the uncertainty of the two weeks put together. A smaller move is greyed and called chance, even though it might turn out to be real.
Mention rate carries a likely range, for example Likely 19.8 to 31.0%, with the number of answers behind it. The range is computed at the standard 95% level from how many questions the figure covers, so fewer answers widen it.
The change check is a standard two-sample test at the same level. A thin week on either side makes the bar wider, and where a family of themes is compared at once, the bar widens to allow for it. Why a number is blank or shows a range explains each message you may see.
Which country and which language are questions asked in? #
Every question is asked in English, everywhere. A market sets where a search is made from, not the language.
Setup, Readings, section Country, shows the country your questions are asked as. Five of the six engines can be pointed at it: ChatGPT, Perplexity, Claude and both Google surfaces. Gemini cannot, so its numbers are not specific to any market.
If you change the country, the week of the change starts a new baseline, and BotSay compares nothing across it.
What does "estimated" mean for Sentiment? #
Sentiment is read, not counted. BotSay's own model reads how an answer describes you and scores it from 0, described very negatively, to 100, recommended outright, the same way every time.
The bands are fixed: 0 to 20 is explicitly negative, 40 to 60 is a neutral mention, 80 to 100 is an explicit recommendation. Every mention with a score carries the quote it was scored from, so you can check the reading against the words that earned it.
Another product's 0 to 100 score reaches its number with a different model and rubric, so the two do not compare. Treat Sentiment as a hint, not as a verdict.
Where should I read these numbers with care? #
Four things shape how to read the figures.
- They are a leading indicator. They show how you appear in AI-mediated search through each engine's own service, which can behave slightly differently from its everyday consumer app.
- Average rank looks only at answers that already name you. It can get worse for a good reason: get named in more answers, including harder ones, and you appear further down in more of them. Read it beside Mention rate.
- English. Brand matching and the Sentiment reading are both tuned for English. If you ask questions in another language, BotSay has not yet reviewed those figures for it.
- Your pages cited counts the sites listed as yours today, so changing the list changes past weeks.
Have any figures been corrected? #
Yes, once, in September 2026. Every figure is now rounded half up at the precision it is shown at, on every screen, so a true 5.05 shows as 5.1. That holds for a figure, for the likely range beside it, and for a change from one week to the next.
Before the correction, a figure sitting exactly on a half could show one step too low in its last decimal. Some earlier figures changed by that one step, 0.1 for most, each to its correct value, and none moved further.
Where is the full method? #
The full method is in the app, at Setup, How we measure. That page is the authority if it and the docs ever differ.
The page holds the detail behind some answers in a section called For the technically minded, closed until you open it. A final section, Trademarks, says whose marks the engine names are.