Share of voice (AI search)

The computation is simple once the inputs are disciplined. Take a fixed set of unbranded buying questions for the category. Run each question against each tracked engine many times over a period, storing every raw answer. For each brand, count the answers whose text literally contains the brand name, divide by total answers, and report the percentage per engine. Because the same answers also name competitors, the identical pipeline produces every rival's percentage at once, which is what turns a visibility number into a share.

A worked example makes the shape clear. Suppose 20 questions run 10 times each against one engine, producing 200 stored answers. Your brand appears in 46 of them, so your share of voice on that engine is 23 percent. A competitor appearing in 92 answers holds 46 percent. Those two numbers together say more than either alone: the category conversation exists, the engine recommends actively, and it recommends the rival twice as often. That gap, question by question, is the work list.

Each input exists to kill a specific distortion. Unbranded questions prevent inflation by construction, since a question containing your name yields a mention automatically. Repeated runs handle answer variance: SparkToro measured under a 1 percent chance that two identical ChatGPT runs return the same brand list, so single checks measure luck. Per-engine scores prevent averaging away a collapse on one engine with strength on another. Stored raw answers make every reported number auditable rather than taken on faith.

Share of voice matters because AI answers now function as shortlists. G2 found in 2026 that 51 percent of software buyers start research with an AI chatbot more often than with Google, that 69 percent chose a different vendor than they originally expected because of chatbot output, and that 33 percent bought from a brand they had never heard of before an AI named it. Forrester's 2026 survey of 18,000 buyers found 94 percent used AI during their most recent purchase. When the shortlist forms inside the answer, the brand with the larger share of that answer wins more of the pipeline before any website comparison happens.

The metric also converts a vague anxiety into an allocation tool. A raw visibility score says how often you appear; share of voice says who appears instead of you, on which questions, on which engines. That resolution is what makes the number actionable: the questions where a competitor dominates and you are absent are ranked, concrete content targets, and the engines where your share lags are the ones whose cited sources deserve attention first.

Share of voice is also the version of AI visibility that survives executive scrutiny. An absolute score invites the question of what a good number looks like, which has no universal answer because categories differ in how list-like their answers are. A share reframes the question competitively: whatever the category's mention dynamics, the brand and its rivals face the same ones, so the gap between them is a fair reading of relative position, and closing it is a goal a team can be held to quarter over quarter.

The most common inflation is mixing branded questions into the denominator. Ask an assistant whether Acme is good for invoicing and Acme gets mentioned by construction; fold enough of those into the set and a brand nobody discovers unprompted still reports a healthy share. Vendors and internal teams both fall into this, sometimes silently. The tell is a score that looks strong while unbranded spot checks keep returning competitor lists.

The second trap is sampling too thinly. Because identical runs disagree, a share computed from one run per question is dominated by variance, and week-over-week movement in such a number is mostly noise being narrated. The third trap is loose matching, where parsers credit fuzzy or implied references as mentions; the honest rule is that a mention counts only when the brand name literally appears in the answer text. Reachroller enforces all three disciplines by design, excluding branded questions from the headline score, sampling repeatedly through official APIs, and linking every counted mention to the stored raw answer, which is the standard any share of voice number should be held to before it reaches a dashboard.