Share of Voice

本文目前只有英文版。站点其余部分已汉化，这篇内容尚未翻译。

Share of voice is a borrowed metric. In advertising it meant your spend as a fraction of category spend; in traditional SEO it came to mean your share of the clicks or impressions available for a keyword set. Both versions work because the denominator is knowable: total category spend, total impressions for those keywords. Carried into AI search, the metric keeps its usefulness only if that property survives the move, and this is where most published figures quietly fail. Share of voice in AI search is the proportion of a defined, collected set of AI answers in which your brand is named or cited, measured against the other brands that appear in those same answers. The definition is unremarkable. The discipline is entirely in the words "defined" and "collected".

The reason the denominator matters more here than anywhere else is that AI answers are generated, not indexed. A search engine's results page for a keyword exists as an artifact you can fetch and count. An AI answer exists only because someone asked, and it may differ the next time the same question is asked. There is no complete list of the answers an engine has produced, and no way to sample them representatively without knowing the distribution of what people ask, which no vendor has. So a claim that your brand holds a given share "of AI search" is a ratio whose bottom half nobody possesses. It reads like a measurement and behaves like an estimate with no stated error. The honest form of the metric is narrower and far more useful: within these prompts, run on these dates, against these engines, your brand appeared in this fraction of the answers, and here is who else appeared.

Once the denominator is fixed, share of voice starts telling you something that plain visibility cannot. Visibility is a question about you alone: were you named, were you cited, yes or no. Share of voice is a question about the room. It is entirely possible for a brand to hold its visibility flat for a quarter while its share of voice erodes, because the same answers that still mention you have started mentioning two competitors who were not there before. That erosion is invisible on any dashboard that only counts your own appearances. It is the earliest signal that a category is getting crowded, and it usually arrives before any change in rankings or traffic, because an engine can start naming a new entrant long before that entrant earns links or positions.

There is a second distinction worth holding onto, because collapsing it is how a share-of-voice number becomes dishonest. Asking "who gets cited in this space" and asking "who appeared alongside me" are different questions with different denominators. The first has no bound: the set of sources an engine might cite across all its answers is open-ended, and a ranking built from whatever you happened to collect is partly a ranking of who you have been collecting longest. The second is bounded by construction, because the answers were the ones you ran, and every brand in them either did or did not appear. Only the second supports a percentage. This is why a well-built AI visibility product will name the two panels differently rather than calling both share of voice, and why a figure presented without its prompt set and date window should be read as a claim rather than a measurement.

Choosing the prompts is therefore the analytical work, not a setup step. A prompt set assembled from your own brand name will return a flattering and largely meaningless share, because a question containing your name is a question you will tend to win. The prompts that carry information are the ones a buyer would actually type without knowing you exist: category questions, "best tool for" questions, and above all alternatives and comparison questions, which is where an engine is most likely to volunteer a list of rivals. Our own measurement of this pattern found that a large share of brands an engine can describe accurately by name vanish entirely when the same engine is asked for alternatives in their category, which is exactly the gap a self-named prompt set will never reveal.

What share of voice cannot tell you is worth stating as plainly as what it can. It is not market share; an engine's willingness to name you reflects what it has read, not what customers bought. It is not traffic; appearing in an answer and receiving a visit are separate events, and most AI surfaces report neither reliably. It is not sentiment; a brand named as the cautionary example scores identically to one named as the recommendation, which is why we do not publish a sentiment score at all rather than infer one from wording. And a share computed over a small prompt set moves violently for reasons that have nothing to do with your standing, so the figure needs both a denominator and enough of it before a trend in it means anything.

TriRank measures share of voice only in the bounded form. The free audit runs a category prompt alongside the brand prompt, without a login, and returns the other domains that came back in the same answers — the anonymous version of the question, on one domain, on one day. Inside the product, brand watchlist tracking accumulates the same comparison across a scheduled prompt set so the ratio has a series behind it rather than a single reading; that competitor comparison is part of the Pro and Agency plans. Both surfaces carry their prompt set and their dates with them, because a share of voice detached from those two things is not a smaller measurement, it is a different kind of statement. If you want to see who is currently appearing next to you, the free audit will show you on your own domain in a couple of minutes.