Share of voice in AI search is the proportion of a category's brand mentions that answer engines give to you rather than your competitors. The core formula divides your brand's mentions by all category mentions, read as a percentage. A single number means something only once you fix the prompt set, the run count and how null answers count.
What share of voice in AI search actually measures
Share of voice in AI search is the percentage of brand mentions or citations in AI-generated answers that go to one brand rather than its competitors inside a defined category. It is a relative measure. If your category's answers name brands and only a slice of them name you, your share is that slice, whatever the absolute traffic behind them.
Read it across several answer engines, not one. ChatGPT, Gemini, Perplexity and comparable systems each build answers from different data and different retrieval steps, so a share measured on one of them describes that engine alone. A category share is the reading you get after running the same prompts through each engine you care about and combining the counts.
Treat the result as a tracked trend, not a snapshot. A share of voice is worth something when you benchmark it against the same competitor set over weeks and watch the direction, because one reading on one day carries the noise of that day.
The unit you pick shapes everything downstream. A mention count rewards being named; a citation count rewards being linked as evidence. Neither is more correct, but they answer different questions, and a brand that is often named yet rarely cited will read very differently depending on which one you fixed. Write the choice down.
Before counting anything, separate three events people fold together.
| Event | What it is | Whether the formula counts it |
|---|---|---|
| Mention | Your brand named in the answer text, with or without a link | Yes, the default unit |
| Citation | Your brand's page linked or footnoted as a source for a claim | Yes, if you count citations instead of mentions |
| Source | A page the engine retrieved to build the answer, whether or not it surfaces | Not directly; it sits behind the answer |
Decide which event your formula counts and hold it fixed. Counting mentions and citations in the same reading, then comparing it to a reading built on mentions alone, produces a difference that is about method, not visibility.
One more distinction: share of voice is not share of search. Share of search is the branded-query volume your brand pulls inside a category on a conventional search engine — a demand signal. Share of voice in AI search is a supply signal about what the engines say. The two move for different reasons; you can see why competitors get cited instead of you without any change in branded search.
The formula, and how to calculate it by hand
The formula is one line: your brand's mentions ÷ all category mentions, expressed as a percentage. Substitute citations for mentions if that is the event you fixed above, and do not mix the two.
By hand, the method is repeatable without a paid tool:
- Define a fixed prompt set — the questions a buyer in your category would ask.
- Run that set through each engine you chose.
- Tally, per brand, how many answers name it.
- Divide your brand's count by the total across all brands, and read the percentage.
The same arithmetic sits inside vendor dashboards. Semrush documents that its Brand Performance report, in its AI Visibility Toolkit, calculates AI share of voice from both how many times a brand is mentioned and how high it appears — verified on 27 August 2026 in Semrush's documentation. HubSpot's AI Search Grader is a free brand-visibility tool that measures how a brand shows up in AI answers, with no account required — verified on 27 August 2026 on HubSpot's product page. A dashboard saves the tally; it does not change the formula.
The denominator is where readings quietly diverge. Many prompts return an answer that names no brand at all, or no citation. Decide once: either exclude null answers from the denominator, so the share describes only answers that mention someone, or count a null as a no-mention that inflates the denominator and lowers every brand's share. Whichever rule you pick, note it beside the figure: a share built by excluding nulls is always higher than the same data with nulls counted, and the two are not interchangeable. Both are defensible; a share is only comparable to another share built the same way.
Sample size decides whether the number holds. A handful of prompts run once will swing on noise. Use a prompt set large enough to cover the real questions in your category — a couple of dozen at least — and repeat the whole set several times per engine rather than trusting a single pass. In the reports we generate, the smaller and the less repeated the set, the wider the reading moves between runs.
Why the number moves: run-to-run and platform variance
Two forces move a share of voice that has nothing to do with your brand.
The first is the platform. Because each engine trains on different data and retrieves differently, the same prompt set yields different shares on each one. A single blended number hides this, so keep the per-engine shares next to the blend and you can see which engine is driving a move.
The second is run-to-run drift. The identical prompt, on the same engine, on the same day, does not always return the identical answer. We observe this drift in the reports we generate: a brand can appear in one run and not the next without anything changing outside the model. So a share of voice is properly a range, not a point. Report the low and high across your runs, or the average with the spread, and treat any move smaller than that spread as noise.
To turn drift into a usable margin, run your set several times in a single sitting and record each run's share separately. The gap between the lowest and highest run is your error band for that day. A later reading counts as real movement only when it clears that band; inside it, the two readings say the same thing.
Platform-level changes matter too. When an engine adjusts how freely it links out or names brands, every tracked share shifts at once, in the same direction, with no action by any brand. We observe that a reading which drops across your whole competitor set on the same date is more likely a platform change than a loss on your side — a reason to log the date of every reading and to automate these AI visibility checks so a shift is easy to place in time. Track the metric over time; do not rewrite strategy on one reading.
What counts as a good share, and comparing the measurement approaches
There is no universal healthy number. A good share depends on the category and the competitor set: a share that is strong against a few rivals is weak against dozens, and a niche with a couple of serious names sets a different bar than a crowded one. The same percentage can be a strong result in one category and a weak one next door, so a number without its competitor set attached says almost nothing. Read your figure against the named competitors you chose, not against an absolute target, and judge it by direction over time.
A complete share — every category mention in every answer going to one brand — is the ceiling, and no brand realistically approaches it. Engines pull from many sources, name alternatives, and hedge, so a category with real competitors will always spread its mentions. Treat a rising share against your set as the goal, not a full sweep.
The measurement approach itself splits into two things a tool may report under one name.
| Approach | What it counts | Reads as |
|---|---|---|
| Prompt-response share | Brand mentions or citations in live answers to your prompt set | How often engines name you right now |
| Domain-level share | Brand presence on authoritative sources the models learn from | How much the training and retrieval layer favours you |
Evertune calculates a domain-level Brand Share of Voice for the top 50 URLs ranked by its AI Education Score, treating share as a training-data signal rather than a live answer count — verified on 27 August 2026 on Evertune's resource page. Both are valid; they answer different questions, and a figure is meaningful only once you know which one produced it. Pick the one that matches your question: a prompt-response share tells you what a buyer sees in an answer today, while a domain-level share tells you how well positioned you are in the material the models draw on.
Where to start this week
A defensible first reading takes an afternoon.
- Pick one category and the competitor set you will measure against.
- Write a fixed prompt set of real buyer questions and save it verbatim.
- Choose the engines — ChatGPT, Gemini, Perplexity or the ones your buyers use.
- Decide your unit, mentions or citations, and your null rule before you count.
- Run the whole set several times through each engine and tally per brand.
- Record the date, the prompt count and the run count next to the figure.
Report the result as a dated range, not a single percentage: the low and high across your runs, with the date. That first report then carries its own confidence context, and the next one is comparable because the prompt set, the unit and the null rule are written down. Change one of them and you have started a new series.
Run a free Namedrop scan to see how often the engines mention your brand against your competitors today.