When Share of Voice in AI Search Misleads You

Share of voice is the number everyone wants on the slide. It is simple: out of all the brand mentions in your tracked answers, what fraction is yours? It is also easy to bend by accident. This post lists the five ways we see it mislead, and how to read it without fooling yourself.

1. The panel decides the answer

Share of voice is calculated over a prompt panel. Change the prompts and the number changes. A panel heavy on branded queries will show you winning, because people who ask about you get you. A panel heavy on generic category questions will show a lower, more honest number.

What to do: tag prompts by type. Promptwatch supports organic, brand and competitor-comparison prompt types, plus topics and tags. Report share of voice per type, not as one blended figure. The organic category prompts are the ones that reflect discovery.

2. The definition drifts

If the rules for what counts as a mention change, the percentage moves without any change in the world. Aliases, exclusions and handling of product names all matter. A Brand Book with written analysis instructions, as Promptwatch provides, keeps the rule stable. If you change it, annotate the chart.

This is also why comparing numbers across tools is hard. Two vendors with different definitions will never match, and neither is necessarily wrong.

3. The average hides the split

A flat 30 percent can mean you are at 30 everywhere, or at 80 in one market and nearly zero in another. Averages erase the structure you need.

What to do: break share of voice down by model, country, topic and persona. Promptwatch's heatmap compares any two dimensions, such as brands against models or topics against countries, so you can find where you are strong and where you disappear.

4. Small samples create fake movement

Answers vary run to run. If a topic has only a few prompts, one different response can move the percentage a lot. Week-to-week swings on thin data are noise.

What to do: set a minimum sample before you act, and prefer period comparisons over single-week reads. Promptwatch offers A/B period comparison. Do not celebrate or panic over a move that rests on a handful of responses.

5. Mentions without citations mislead

Share of voice counts names. It does not tell you where the model learned them. You can have a healthy share built on third-party pages you do not control, and lose it when those pages change. Conversely, you can have lower share but strong citations from your own site, which is a more stable position.

What to do: read share of voice next to citation analytics. If mentions are high but your own pages are rarely cited, you rely on offsite mentions. That is a risk to manage, not a number to boast about.

Other things that distort the picture

Competitor set choice

Your share depends on who is in the denominator. Leave out a big competitor and your share inflates. Include an irrelevant one and it deflates. Promptwatch auto-discovers competitors, which can reveal rivals you left out.

Position is not the same as presence

Being mentioned last in a long list is different from being the first recommendation. Position in Answer is a separate metric in Promptwatch, alongside Visibility Score and Sentiment. A report that shows only share of voice hides this.

Engine mix

Share of voice across a blend of engines can bury a strong showing on one and a weak one on another. Report per engine, and be explicit about which are tracked through the real UI and which through APIs.

A safer way to report

Use a small set of numbers with their context:

1. Share of voice per prompt type, not blended.
2. Per engine, labeled UI or API.
3. With the sample size, meaning prompts and responses.
4. With position in answer for the prompts that matter.
5. With citation share from your own pages alongside it.
6. With the definition and date of any change.

That is six items, not a dashboard of forty. It is enough to keep you honest.

What to say when someone asks "are we winning?"

Resist a single number. Say: "On category prompts we appear in this fraction of answers, down or up this much against last period, and the gap is concentrated in these topics and this country." That answer is longer and far more useful. It also invites the right follow-up question: what are we doing about the gap?

Recommendation

Treat share of voice as an index that needs a definition, a panel, a sample size and a companion metric. Promptwatch makes that easier because the Brand Book, prompt types, heatmaps, period comparison, position in answer and citation analytics sit in the same workspace as the share-of-voice chart. Before the next report, rebuild your headline number per prompt type and see how different it looks. If it is much lower on organic prompts than on branded ones, you have just found your real discovery gap.