Short answer: "brand share in AI answers" is used for at least four different numbers: mention rate, share of voice, recommendation rate and citation share. Each answers a different question, and each can be inflated in its own way without anything looking wrong. Before you trust a brand-share figure, find out which of the four it is, which questions and engines it was computed on, and what was left out.
This guide is about reading a number, not producing one. If you are building the measurement yourself, how to measure share of voice in AI answers walks through the calculation step by step.
Four numbers that all get called brand share in AI answers
| Metric | What it counts | The question it answers | How it most often misleads |
|---|---|---|---|
| Mention rate | Answers that name you, divided by completed answers | How often am I in the conversation at all? | Questions that contain your brand name, which the assistant will of course answer by naming you |
| Share of voice | Your mentions, divided by mentions of every brand named | How much of the conversation is about me rather than others? | Competitors detected against a fixed list while your own brand is matched loosely |
| Recommendation rate | Answers that recommend you, as opposed to merely naming you | Would a buyer reading this choose me? | Merged with mentions, so a critical mention counts as a win |
| Citation share and breadth | Sources marked in the answer text that point at your domain, and how many distinct questions cited it | Is my content being used as evidence? | Counting every page the engine retrieved rather than the sources it actually marked |
None of these is the right one in general. A report that shows only one of them, without saying which, is the first thing to question. The mention rate and share-of-voice glossary entries give the short definitions.
How each number gets inflated
Mention rate: the echo of your own name
Ask an assistant whether a named company is any good and it will name that company. If a question set contains such questions, mention rate rises without saying anything about whether you surface on your own. This is not hypothetical: an LLM-generated question set once contained the brand name in 13 of 15 questions, and the resulting score read 13% across five engines when the honest value was zero. Brand-named questions belong in a separate reputation view, out of the visibility numbers entirely.
Share of voice: the shrinking denominator
Share of voice is a fraction, so it can be raised by shrinking the bottom as well as by growing the top. If every brand an answer names is not counted, the brands nobody put on the list fall out of the denominator and your share rises by construction. Asymmetric detection once showed a brand at 31% share of voice and rank two; measured symmetrically it was 0% and last (Figure 2). Always ask how many distinct brands the share was computed against: being 1% of 117 brands is a completely different situation from 1% of five.
Recommendation: a mention is not an endorsement
An answer can name you in order to steer the buyer elsewhere. Your page can also be cited as a source while the answer recommends a competitor. Merging being named, being cited and being recommended flatters the report and hides the actual gap. They need to be separate columns.
Citations: the retrieval pool
An assistant typically retrieves many more pages than it cites. Only sources marked inside the answer text count as citations. Counting the retrieval candidate pool instead overstated our own citation figures by two to four times before we fixed it. Breadth is the harder number to inflate: one answer citing twelve of your pages is not twelve times the reach, so ask how many distinct questions cited your domain, not only how many citations were recorded. Engines differ here too. Perplexity exposes sources on nearly every answer, which makes it the engine where citation share is easiest to observe; its bot documentation explains which agent fetches pages for those answers. Citation analysis goes further into reading citation data.
Why the denominator decides how to read brand share
Every number in the table above is a fraction, and most disputes about brand share are really disputes about the bottom of that fraction. Three properties of the denominator change what a share means, and a report should state all three next to the percentage.
How many brands are in it. Being 1% of 117 brands and 1% of five are different situations, and the percentage alone cannot tell them apart. In a crowded category a small share can still be a respectable position; in a category with five names it means you are barely there. The total number of distinct brands belongs next to every share.
Whether one brand is counted as one. Casing, punctuation, legal suffixes and a product line named instead of the company each split one brand into several rows. Splitting a competitor understates it and flatters you; splitting your own brand understates you. Spelling variants should be resolved to one entity before anything is counted, and a report should say that they were.
How many answers it rests on. A percentage computed from a handful of answers is noise wearing the costume of a signal, and it will swing on the next run. When the sample is too thin, the honest move is to withhold the share rather than print a number that will reverse itself.
Engines also differ in how many names they put in an answer, which changes the field each share is computed over:
That is why the per-engine question in the list below is not a formality. A share on an engine that names fewer brands per answer is computed against a smaller field, so the same position can produce a different percentage on each engine.
Seven questions to ask about any brand-share report
- Can I see the question set, and does any question contain a brand name? If the questions are hidden, the number cannot be checked.
- Is each engine reported separately? In our own benchmark one model named 3.5 providers per answer and another 2.1. A single cross-engine figure describes neither.
- Was the methodology the same across the whole trend? Enabling web search once moved a site from 3% to 13% while the site did nothing. Every run should carry a methodology version, and runs on different versions should never be compared.
- How are failed or missing runs shown? A gap is honest. A zero creates a fake decline. Quiet omission can be just as misleading: in one of our own incidents, runs that took longer than 10 minutes were never snapshotted, so the trend silently dropped them.
- Is every named brand counted, and is the total shown? A share without its denominator is a decoration.
- Are citations counted from the answer text? Ask directly whether the retrieval pool is included.
- Can I open the original answer behind any number? Without the stored text, engine, date and cited URLs, nobody can tell a counting bug from a real change.
A report that answers all seven precisely is probably sound, whatever it looks like. A vague answer to any one of them marks the place where the number was made rather than measured.
Which number to use for what
| Your situation | Lead with | Why |
|---|---|---|
| A brand or store competing for recommendations | Share of voice and mention rate, split by intent | Being named in research questions and absent from purchase questions is a different problem from the reverse |
| A publisher or content site | Citation breadth | You are rarely the brand recommended; you are the source cited, so distinct citing questions are the metric |
| An agency reporting to clients | The list of questions the client lost | That list is the content brief, and it only means something on a fixed, versioned question set |
| Reputation monitoring | Answers to brand-named questions, read for wrong details | These questions say nothing about visibility, but they show what assistants say about you when asked directly |
Whichever number you lead with, split it by intent before drawing conclusions. A brand can be named often in research questions, where buyers are still learning the category, and rarely in purchase questions, where they are choosing. The reverse pattern is a different problem with a different fix, and one overall percentage hides which of the two you have. Record the intent on every question when the set is built, so the split costs nothing later.
One caveat applies to all of them. Being cited does not send traffic: most answers are read without a click. A rising citation share is a brand outcome, measured on the answer side, and it will not necessarily show up in your analytics. Our guide to GEO reporting for clients covers how to present these numbers without overstating them.
Sources
- Perplexity: crawlers, on which agent fetches pages for Perplexity answers
- OpenAI: overview of OpenAI crawlers
- First-party figures in this guide come from InsightWonder's own incident records (July to September 2026) and benchmark, described in each figure caption.