Key takeaways
- Use separate figures, not one score: whether you are named, whether you are recommended, whether your own pages are cited and how settled the answer is.
- Every figure needs its count, its base and an interval, or it is only a direction.
- Give each audience a short set. The team doing the work needs the diagnosis. Leadership needs the movement in recommendations, with the uncertainty beside it.
The honest answer to this question is a short list, not a dashboard. A B2B marketing team that wants to know how its brand shows up in AI search needs a handful of figures that each answer one question, a way to read them, and the discipline not to add them together.
This guide sets out the figures we use and how we define them. The definitions are the ones in our published methodology, written here in plain words. Throughout, rates are written as decimals between zero and one, or as a count over a base. Read 0.75 as three in every four.
Start with three outcomes, not one score
A single AI answer can name your brand, cite your pages, recommend you, or any mix of the three. Those are different outcomes with different causes, and mention, citation and qualified recommendation each need their own count. Every figure below is built on one of them, or on how they hold up when the same question is asked again.
Two rules apply to all of them.
- One assistant at a time. Each assistant is its own series. A figure for Gemini and a figure for Perplexity are never added or averaged, because there is no defensible weight for one assistant against another.
- A fixed set of questions. A figure is only comparable over time if the buyer questions behind it stay the same. Building a buyer-question set covers how to freeze one.
The eight figures, in plain words
| Figure | The question it asks | Read it as |
|---|---|---|
| Mention rate | In how many answers is your product named? | A rate over the answers that track you |
| Qualified recommendation rate | In how many answers are you put forward as the choice or as a named alternative, weighted by how much each question matters commercially? | A weighted rate from zero to one |
| Citation rate | When an answer cites any source, how often is one of them on your own domain? | A rate over the answers that carry a citation |
| Share of recommendation | How much of the recommendation credit in these answers goes to you rather than to the other brands you track? | A share from minus one to one |
| Answer position | When an answer presents an ordered list, at what place does it put you? | The median place, where a lower place is nearer the top |
| Evidence coverage | For how many of your question families do answers back you with a page on your own domain and a source elsewhere? | A weighted rate from zero to one |
| Citation diversity | How many different websites do the sources that support you come from? | An effective number of hosts, from one upward |
| Visibility stability | When one question is asked more than once, how often does the answer treat you the same way each time? | A rate over repeated question groups |
A few details decide whether these figures mean anything.
A mention is a name in the answer. It counts your product by its name, a listed alias, or an edition, tier or plan of it. A company named only as an owner or parent does not count. Being named tells you the assistant connects you with the topic. It does not tell you whether you were put forward.
A qualified recommendation is a role, not a name. The answer puts you forward as the choice, or as a recommended alternative, for the buyer's stated need. "Popular tools include" followed by a list is a mention, not a recommendation. This is the figure closest to the shortlist, and the one most worth watching.
A citation is a source the answer actually points to. Citation rate asks whether your own domain is among them. Evidence coverage asks a harder question: whether a question family is backed by your own page and by a source on another host. A brand that is only ever supported by its own site is easier to dismiss than one that others also describe.
Share of recommendation rests on an assumption. It gives each role a fixed number of points, for example 1 for a primary recommendation and 0.65 for a recommended alternative, and takes your points as a share of everything tracked in the same answers. Those points are a product assumption, not a measurement, and the figure should say so wherever it is shown.
Answer position is a place in an ordered list, not a search ranking. An AI answer has no results page. A place exists only where the answer presents an ordered recommendation, so the figure always needs the number of answers it rests on beside it. Place 1 is the first brand named.
Citation diversity counts hosts by their share. Ten pages from one site count as about one host. The figure rewards support from many different places and gives no credit for volume from one.
Visibility stability is about repeat runs. AI answers vary from one run to the next. Stability is the share of repeated question groups where every answer gives you the same role. A brand that is recommended in one run and absent in the next has a different problem from one that is steadily absent.
Read every figure with its interval
A rate without its count is not ready to be acted on. "0.75" might be three of four question families or three hundred of four hundred, and those are not the same claim.
An interval says how wide the truth could be, given the evidence. Three of four families is a rate of 0.75, with an interval running from 0.30 to 0.95. Thirty of forty families gives the same rate with an interval from 0.60 to 0.86. The first is a direction. The second is starting to be a measurement. Reading an AI visibility figure honestly sets out how the intervals are built and why the unit of independence is the question family, not the single answer.
Before you report a change from one period to the next, check whether the two intervals overlap substantially. If they do, the honest report is that the change has not been shown.
Which figures to report to whom
Different people make different decisions, so they need different figures.
- The head of marketing or demand generation needs mention rate and qualified recommendation rate by question family. The gap between the two shows where you are named but not put forward.
- The content and web team needs citation rate, evidence coverage and citation diversity. These point at what to publish and where the supporting evidence is missing. Evidence gaps turns a lost answer into a piece of work.
- Leadership needs qualified recommendation rate and its movement, with the interval and the number of answers beside it, plus one line on stability so a single good run is not mistaken for a trend.
Answer position and share of recommendation are useful once the first three are steady. They add detail. They should not lead a report.
What not to count
- A single answer or a single screenshot. Answers change between runs and between assistants. One answer is an anecdote.
- A blended visibility score. Folding mention, citation and recommendation into one number hides exactly what you need to know in order to decide what to do next.
- Figures added across assistants. Report each assistant as its own series.
- A rate without its count and its base. If a report cannot say out of what, it cannot be compared with anything.
- Answers that could not be read. Report them as missing. Do not drop them silently or count them as zero.
- Traffic from AI assistants as a stand-in for visibility. It is worth watching in your analytics, but it only shows the people who clicked. It cannot tell you whether you were named or recommended at all.
Where to start
A small team can begin with three figures: mention rate, qualified recommendation rate and citation rate. Measure each assistant separately, on a fixed set of buyer questions, each asked more than once, and report the interval with every figure. Add the other five once the baseline is steady.
The first ninety days of AI visibility work sets out the order of the work around those figures: agree the questions, set the baseline, fix the evidence, then measure again on the same questions.
Tools that report these figures
You can compute these figures by hand from a spreadsheet of recorded answers, and a small team often starts there. Past a few dozen questions, most teams use a tool for the repeated asking and counting. Tools measure; the work that changes the figures is done by people, in your team or at an agency. Some teams pair the two. Check that any tool reports a count and a base with each figure, not a score alone.
The lines below say only what each company's own page says it reports and who it says it serves. They are not a ranking, and each carries the date its page was last checked.
Peec AIMeasurement tool
Its page says it reports visibility (the share of chats in which a brand is mentioned), position and sentiment across AI search platforms, with competitor benchmarking. Its site presents the product as AI search analytics for marketing teams.
Their pageLast checked
ProfoundMeasurement tool
Its page says it reports visibility scores and share of voice, sentiment and citation analysis, with benchmarking against competitors. It lists solutions for AEO, content, and PR and brand teams, and for agencies.
Their pageLast checked
Ahrefs Brand RadarMeasurement tool
Its page says it reports mentions, citations, AI share of voice and estimated impressions across AI platforms. Its page speaks to anyone who wants to track and grow a brand's visibility in AI answers, and presents the product as part of the wider Ahrefs marketing platform.
Their pageLast checked
ScrunchMeasurement tool
Its page says it reports brand presence, citations and sentiment over time, by prompt and by model, with a comparison against competitors. It lists use cases for enterprise brands and for agencies.
Their pageLast checked
Answer Position is one option too. It is a UK agency for B2B technology brands, serving UK and EMEA vendors. Its Monitor service tracks 50 buyer questions every month and gives a monthly report with the five actions that matter most. It is described on our services page.
