Share of voice in AI answers: what it is and how to measure it

Share of voice in AI answers is your brand's share of all the recommendations made across a fixed set of questions. To measure it, ask the same buying questions of the assistants, count every brand each answer recommends, and divide your count by the total. It is a relative measure: it can fall when rivals gain even if you stay still. For that reason it is best read alongside your own recommendation rate.

Where does the idea come from?

In marketing, share of voice is a brand's portion of the conversation in its category, often expressed against competitors. The AI version applies the same logic to the answers assistants give. Instead of counting ad impressions or media mentions, you count how often each brand is put forward.

It is appealing because it fits the way AI answers work. An answer names a small number of brands, so the "slots" are scarce, and your share of them is a clear way to see who is winning.

How do you calculate it?

Define the pieces first.

1. A fixed question set, written as buyers write, with no brand names.
2. The assistants, and a fixed way of asking, including repeats.
3. A definition of a "positive recommendation", applied consistently.

Then:

• Count the recommendations of each brand across all valid answers.
• Add them up to get the total recommendations.
• Your share is your count divided by the total.

For example, imagine your brand is recommended 12 times, and the total across all brands is 80. Your share of voice is 15 percent. The numbers here are purely illustrative.

What should count as a recommendation?

This is the crux. If you count every name in the answer, a share of voice will reward brands that are merely mentioned, including as bad examples. Count only answers that put the brand forward as a choice, and keep neutral and negative mentions separate. Mentioned, recommended or cited draws these lines.

Be wary of order, too. A first position in a list is not proof of preference. Treat it as preference only where the answer says so in words.

Why is it relative, and why does that matter?

Because the denominator is everyone's recommendations. If a rival is recommended more, the total grows and your share shrinks, even if your own count has not moved. Conversely, if a rival drops out, your share rises with no change from you.

This is not a flaw, but it can mislead if it is the only number you watch. Pair it with an absolute rate: the proportion of answers in which you are recommended. If the rate holds and the share falls, competitors are gaining. If both fall, you are losing ground in absolute terms.

Citewake's method makes the same distinction: it reports a recommendation rate and a competitive share separately, and notes wherever the share is shown that it depends on rivals.

How do you handle the awkward cases?

Brand variants. If a brand appears under several names or product names, decide in advance how to group them, and apply it everywhere. When is a brand the same brand? explores why this is harder than it looks.

Answers with no brand. Some answers name nobody. They still count as valid answers, and they matter for your rate, though they add nothing to the share total.

Failed or refused answers. Exclude them from counts and report how many there were. Do not quietly fill gaps.

Repeated runs. Keep the answers to one question together, rather than treating each run as independent evidence of the whole category.

What are the common mistakes?

A few come up often. The first is counting mentions as recommendations, which rewards brands that are named as warnings or as background. The second is changing the competitor set between periods, which changes the denominator and makes shares incomparable. The third is reading a share from a small number of answers as precise.

A fourth is forgetting that the share depends on the questions. A set weighted toward broad questions, where large brands dominate, will give a different picture from one weighted toward narrow ones. Report the mix of questions alongside the share.

A fifth is quoting one blended share across all assistants without showing the split. The assistants differ, and your position may be strong on one and weak on another. The blend hides precisely the information you most need.

How should you read it?

Read it by question type and by assistant, not only overall. A modest overall share might hide a strong position on narrow, specific questions. Look at where you are strong and where you are not.

Also report uncertainty. A share calculated from a small number of answers can swing a lot. Say how many answers and questions sit behind it, and be cautious about reading small changes as real. Margins of error in AI measurement explains why.

What is a sensible way to start?

Choose a competitor set, perhaps five to ten brands you meet in deals. Use a fixed question list and several runs on each assistant. Tally recommendations by brand. Calculate your share and your recommendation rate. Save every answer so you can recount.

Repeat on a schedule and compare. For a first look at where you appear, try the Citewake checker.

What is the short version?

Share of voice is your slice of all recommendations across a fixed question set. Count only real recommendations, keep the competitor set and questions steady, report your own rate beside it and state how many answers stand behind it. Read it by assistant and by question type, not only as one overall figure.