Share of voice in AI search: how to calculate it and compare it fairly
A practical guide to AI share of voice: the formula, the traps in comparing tools, how prompt mix changes the number, and which platform features make the metric usable.
Quick verdict
Share of voice is the most quoted AI visibility metric and the easiest to misread. The number depends entirely on which prompts you track, which engines you sample, and how a tool decides what counts as a mention. Two vendors can show the same brand at 12% and 31% on the same day without either being wrong. The fair way to use it is as a trend inside one consistent panel. Promptwatch is built for that, with daily panels per model, country, and language, plus a Brand Book that fixes what counts as a mention.
The formula
Share of voice is your mentions divided by all brand mentions in the same set of answers, as a percentage.
If 100 sampled answers name at least one brand from your category, and your brand appears in 24 of them while competitors appear in a combined 96 mentions, you need to decide which denominator you are using. There are two common versions:
- Mention share: your mentions divided by total brand mentions across all answers. Competitors named three times in one answer count three times.
- Answer share: the share of answers that include you, ignoring how many other brands appear.
Mention share penalizes crowded answers. Answer share is easier to explain to executives. Pick one, write it down, and never mix them in a single report.
Why the number moves for reasons that are not performance
Most swings in share of voice come from the panel, not the brand.
Prompt mix. A panel heavy on brand-name prompts inflates your share, since you are named in the question. A panel heavy on generic category questions deflates it. Tag prompts by type, such as organic, brand, and competitor comparison, and report each separately. Promptwatch supports topics, tags, and prompt types for this reason.
Engine mix. Answers differ by engine. A tool that samples four engines and one that samples eleven are measuring different things. Compare like with like.
Location and language. The same prompt asked from two countries can return different brands. Promptwatch tracks per country on every paid plan, and by state and city from Business.
Mention definition. Does a passing "also consider" count? Does a mention of a product name without the company name? In Promptwatch, the Brand Book's analysis instructions decide what counts as a mention, and brand aliases catch alternate names. Whatever tool you use, find the equivalent setting.
Sampling volume. Small panels are noisy. Fifty responses can swing a few points for no reason.
A sanity check you can run by hand
Take ten prompts, run each three times in a clean session, and count. It is rough, and it will not match any tool exactly. What it gives you is an order-of-magnitude check. If your tool says 40% and your hand count says 10%, find out why before presenting.
Share of voice is not the whole story
Pair it with two companions.
Position in answer. Being named first is different from being named fifth. Promptwatch reports Position in Answer alongside Visibility Score and Share of Voice.
Sentiment. Promptwatch scores each response from 1 to 100 and adds a weekly AI summary. A rising share with a falling sentiment is a warning, not a win.
A heatmap that compares any two dimensions, such as brands by model or topics by country, helps explain share differences. It will show, for example, that you lead on one engine and trail on another, which a single blended number hides.
Making it comparable over time
- Freeze the panel. Change prompts only on a schedule, and annotate the change.
- Freeze the competitor set. Adding a rival changes the denominator. Promptwatch also auto-discovers competitors, so review that list before reading trends.
- Compare periods properly. The A/B period comparison view lines up two date ranges.
- Watch quota. Usage alerts fire at 85% of a quota, so a panel does not silently stop early.
- Report ranges, not points. "Between 18% and 24% across the month" is more honest than "21%".
What others in the directory show
Cross-tool comparison needs caution. Scrunch AI's Core plan covers 125 unique prompts and 5,000 responses a month across four LLMs, so its sample is narrower than one that tracks more engines. AthenaHQ lists 3,600 credits on Starter, where one credit equals one AI response. Rankscale uses monthly credits where an engine query typically costs 0.25 credit. These are different sampling economics, and they shape how stable a share figure can be. Our ranking has the full entries.
Plan notes
Explore is free but tracks 10 prompts on ChatGPT only, so share of voice there is a single-engine view. Essential at $95 a month gives 50 prompts and 6,000 responses across all active models. Professional at $245 gives 150 prompts and 18,000 responses plus custom reports and the Data Studio connector. Agency plans start at $199 with unlimited prompts and projects, which matters when you serve many clients.
Who should pick which
| You are... | Recommendation |
|---|---|
| Reporting to leadership once a quarter | Answer share, one panel, ranges |
| Benchmarking against three rivals | Mention share, frozen competitor set, daily panel |
| Selling GEO services | Promptwatch agency plans, with client reports and scheduled PDFs |
| Comparing two vendors' scores | Do not. Run both on the same prompts, or pick one tool |
| Needing the why behind a swing | Platform with citation analytics and crawler logs |
The recommendation is simple. Use share of voice as a trend, define it in writing, and choose a tool that makes the definition explicit. Promptwatch does, and it keeps the surrounding evidence on the same screen, so a swing leads to a cause rather than a shrug.