What we count
VozRadar asks AI answer surfaces the questions buyers in a market ask, many times, and reads every answer. From those answers it reports three rates, each a share of answers:
- Mention rate: answers that name your brand.
- Recommendation rate: answers that present your brand as an option to choose, not just a name in passing.
- Citation coverage: answers that link to a page on your own domain as a source.
Every rate belongs to one market. The headline rate counts the answers from the surfaces you select, and the same rate is always shown per surface too, so you can see which assistant moves it. Markets are never pooled or averaged: Mexico and Colombia are measured as two series, even in the same language.
How answers are collected
Answers come from the providers' official APIs: OpenAI's web search API, Gemini with Google Search grounding, Perplexity's Agent API with web search and Claude's web search API. Google AI Mode and AI Overviews come from Google result pages, fetched through a search data provider's API.
VozRadar does not log into consumer chat apps or automate their screens. An API call can be repeated and its settings recorded, so every answer can be traced back to the request that produced it.
Each answer carries the name of the surface that produced it, for example "OpenAI · Web-grounded API". An API answer is not what a person sees in the ChatGPT app, so VozRadar never calls it a ChatGPT ranking.
Why an answer without your brand counts
The denominator of every rate is every answer to the tracked questions. An answer that does not name your brand counts as a zero. If those answers were skipped, a rate could only go up, and a brand named once in a hundred answers would look as visible as one named every time.
The tracked questions are brand-neutral buyer questions, such as "which payroll software complies with CFDI in Mexico", not your product name. So an absence is the answer a buyer would have read: one where you were not an option.
Missing data is different. When a provider call fails or an answer cannot be read, that answer is left out of the rate. It is a missing observation, not an absence.
An absence is also not a certainty. A brand seen in 0 of 10 answers is not "never mentioned": its 95% interval still reaches 28%.
Rates with 95% intervals
A rate from a handful of answers is uncertain, so VozRadar treats each rate as an estimate. It uses a Beta-Binomial model: before any answers, every rate between 0% and 100% is equally plausible (a Beta(1, 1) prior). After seeing answers with and without the brand, the estimate is Beta(answers with + 1, answers without + 1).
Each rate is shown with three numbers: the estimate (the posterior mean), the 95% interval (the range that holds the true rate with 95% probability under the model) and the number of answers.
For example, a brand recommended in 12 of 40 answers has an estimated rate of 31%, with a 95% interval from 18% to 46%. More answers narrow the interval:
| Answers with the brand | Answers | Estimate | 95% interval |
|---|---|---|---|
| 3 | 10 | 33% | 11% to 61% |
| 12 | 40 | 31% | 18% to 46% |
| 36 | 120 | 30% | 23% to 39% |
| 120 | 400 | 30% | 26% to 35% |
Two rates whose intervals overlap widely are not clearly different. That is why VozRadar never shows a rate without its sample size, and never shows a position like "#2".
When the instrument changes
Providers change their models and their search behaviour. When a provider's model changes, when its search grounding behaves materially differently, or when VozRadar changes how it reads a surface, a new surface epoch starts.
A trend that crosses an epoch is a comparison between two instruments, so it is flagged rather than read as a change in your visibility. An experiment whose window crosses an epoch is classified as inconclusive.
The synthetic test surface
VozRadar has one simulated answer surface. It produces made-up answers so the whole pipeline (collection, reading, rates, reports) can be tested without paid provider calls.
Its answers never mix with real ones. They are shown only while a project has no real answers yet, under a striped "Synthetic test data" banner, and they never trigger alerts. They do not measure any AI product.
How results of a fix are worded
When you ship a fix, VozRadar opens an experiment. It compares the rate in a baseline window before the change with the rate in a window after it, each with a minimum number of answers, and words the result in one of four ways:
- Likely worked: the rate rose by at least 10 points, with at least 85% probability that it increased.
- Likely worse: the rate fell by at least 10 points, with at least 85% probability that it decreased.
- No clear change: the difference is too small or too uncertain to call.
- Inconclusive: a surface changed during the window, or control questions the fix did not target moved the same way, which points to model or search drift.
These are observations about the answers, not proof of why they changed. Nobody controls what an AI assistant says, so an improvement after a fix is evidence that it helped. The guide on how to measure AI search visibility covers the same discipline for any team.
See how AI sees your brand.
Start with one market. Expand when you're ready.