Traditional share of voice measures what percentage of clicks, impressions, or ad auctions your brand captures relative to competitors in the same space. It is a useful concept precisely because it is relative rather than absolute — being cited on 30% of your queries means something very different if competitors average 10% versus if they average 60%. I think this concept translates well to AI answers, and it is a better lens than the simple binary cited/not-cited tracking I described in measuring AI visibility.
Why Binary Citation Tracking Undersells the Real Picture
Imagine two scenarios. In the first, your site is cited on 40% of your tracked queries, and no single competitor exceeds 15% on the same set — you are clearly the dominant cited source even though you are not appearing on every query. In the second, your site is also cited on 40% of queries, but one competitor is cited on 85% of the same set. Both scenarios produce an identical citation rate for you, but they represent completely different competitive positions. A measurement program that stops at “are we cited” cannot tell these apart. Share of voice fixes this by forcing you to look at the whole field of citations for a query set, not just your own row.
How to Actually Calculate It
Run your query set across your chosen AI platforms, but instead of only recording whether your site was cited, record every source cited in each answer. For a set of 30 queries checked across two platforms, you end up with a list of every domain that showed up as a source, and how often. Your share of voice is your citation count divided by the total citation count across all sources in the set (not divided by the number of queries, which is the binary metric). This distinction matters: if the average AI answer in your space cites four sources, the maximum theoretical share for any one player, including you, is meaningfully less than 100%, and you need to know what a realistic ceiling looks like before you can judge your own number.
Winner-Take-Most vs. Fragmented Query Types
One pattern I see consistently: some query types produce answers that cite one dominant source almost every time, while other query types produce answers pulling from four or five sources with no clear leader. Highly specific, well-defined questions (a precise how-to, a factual lookup) tend toward winner-take-most, where whichever site has written the clearest, most directly structured answer to that exact question gets cited disproportionately. Broader, more exploratory queries tend to fragment across more sources because there is no single definitive answer to point to. Knowing which pattern applies to a given query cluster changes your strategy — for winner-take-most queries, the goal is genuinely to become the one answer; for fragmented queries, a respectable share alongside several competitors may be the realistic ceiling.
Segmenting Share of Voice by Competitor Type
When I build this out for clients, I split competitors into two groups: direct competitors (same market, similar offering) and reference sources (general publications, Wikipedia-style resources, government or research sites that get cited regardless of who the commercial players are). Comparing your share against direct competitors tells you about competitive standing. Comparing your share against reference sources tells you something different — whether AI systems are treating you as a peer to established authoritative references at all, or whether you are still competing mainly for the leftover slots after the reference sources get picked. Both numbers are useful; conflating them is not.
Using Share of Voice to Prioritize Work
The most practical use of this metric is deciding where to spend content and technical effort. A query cluster where you hold 5% share against a competitor holding 70% is a much harder climb than one where the field is fragmented and no one holds more than 25%. I generally advise clients to prioritize the fragmented, no-clear-leader clusters first — the marginal effort to move from 15% to 30% share tends to be smaller than the effort required to dislodge an entrenched leader, and you get measurable wins faster, which keeps the whole measurement program credible internally. For the harder, leader-dominated clusters, see benchmarking AI visibility against competitors for how to figure out specifically what the leader is doing that you are not.
I also weight this by query value rather than treating every point of share equally. A three-point share gain on a high-intent, bottom-of-funnel query cluster is worth more in practice than a ten-point gain on a broad, top-of-funnel cluster that rarely converts to anything measurable. Before you commit effort to closing a share gap, check whether the queries in that cluster are ones you actually want to win, not just ones where winning looks achievable.
A Worked Example
Say you track 30 queries in a niche home services market, checked monthly across Google AI Overviews and Perplexity. Across those 30 queries, the answers cite a combined total of 90 sources (an average of 3 per answer). Your site appears 18 times — a 20% share of voice. Your leading competitor appears 40 times, a 44% share. A second competitor appears 12 times, a 13% share, with the remaining sources spread across smaller or reference sites. That single calculation tells you more than any of the individual numbers alone: you are the clear second position, meaningfully behind the leader but ahead of everyone else, and the gap to close is roughly double your current share rather than some vague sense of “behind.” Recalculate the same breakdown three months later, and the direction of movement — is the 20% climbing toward the leader’s 44%, or is the leader pulling further ahead — is the actual signal worth acting on.
The Caveat Worth Repeating
Share of voice in AI answers is noisier than traditional share of voice metrics, because the underlying systems are less deterministic and change more often. Do not treat a single month’s calculation as a fixed competitive map. Treat it as one data point in a trend, recalculated on the same cadence as the rest of your AI visibility tracking, and trust the direction of movement over two or three cycles more than any single snapshot.