Every marketing function in pharma can tell you its share of voice. It is on a slide in every quarterly review: impressions, reach, the percentage of category spend the brand owns, the share of the conversation versus the competitive set. The number is comforting because it is countable, and because everyone agrees on how to count it. The problem is that it measures a world that is quietly closing. When a clinician asks an AI engine about a condition, share of voice predicts nothing about what comes back. The metric that does is Share of Answer, and almost no brand team can put a real number on it yet.
This is not a vocabulary upgrade. Share of voice and Share of Answer measure two different physics. Share of voice counts how loud you were. Share of Answer counts whether the machine repeated you, and whether what it repeated was true. As Google now resolves a large share of health queries inside an AI Overview before the user reaches a result, in one 2025 analysis of more than 130,000 health queries AI Overviews appeared on 51 percent of them, the gap between the two metrics stops being academic. You can win share of voice and lose Share of Answer in the same quarter, and never see it on a dashboard.
What Share of Answer actually is, defined so you can measure it
Share of Answer is the percentage of relevant AI answers, across the engines your audience uses, in which your brand is mentioned. That is the headline number, and on its own it is incomplete, because a mention can be wrong. So the honest definition has three layers, and a real instrument captures all three.
The first layer is presence. For a defined set of questions, in what fraction of the answers does the brand appear at all. This is the closest analog to share of voice, and the one most "AI visibility" tools stop at. Presence alone is a vanity number in a regulated category, because a confident, frequent, off-label mention is worse than silence.
The second layer is accuracy. Of the answers that mention the brand, what fraction describe it in a way that matches the approved label: the right indication, the right population, the safety language intact. This is the layer that makes Share of Answer a pharma metric rather than a generic SEO metric. Presence without accuracy is exposure measured as if it were reach.
The third layer is drift. Of the mentions, what fraction read off-label: an unsupported indication, a dropped contraindication, a benefit generalized past what the data supports. Drift is the number that should travel from marketing to medical and regulatory, because it is not a performance metric. It is a compliance signal wearing a marketing metric's clothes.
A Share of Answer report that gives you only the first layer is selling you a louder share of voice. The metric only earns its name when it carries all three.
Why share of voice cannot be patched into Share of Answer
The temptation is to treat Share of Answer as share of voice with a new data source: point the old measurement habit at ChatGPT and call it done. It does not work, for three structural reasons.
First, the artifact is not fixed. Share of voice counts stable artifacts: an ad ran or it did not, an email sent or it did not. An AI answer is generated fresh, slightly differently, for every user, on every engine, and it is re-generated when the model updates. You cannot count it once. You have to sample it continuously, or you are reporting a screenshot of a river as if it were the river.
Second, the denominator is contested. Share of voice has a clean denominator: total category impressions, total spend. Share of Answer's denominator is "the questions your audience actually asks," and if you pick your campaign headlines instead of real prompts, you measure a flattering fiction. The denominator has to be the real questions, in the audience's own words, or the percentage is meaningless.
Third, the unit is a paraphrase, not a placement. Share of voice never had to ask whether the message was correct, only whether it appeared. Share of Answer cannot skip that question, because the machine rewrote the message on the way out. A metric that counts mentions but not fidelity is, in pharma, actively dangerous: it rewards a brand for being frequently, fluently, wrongly described.
A worked example: Varigel's two scoreboards diverge
Take a fictional brand, Varigel, approved for a single narrow indication with a known contraindication. Through a strong campaign quarter, Varigel's share of voice climbs: more impressions, more reach, more of the category conversation than the competitive set. The traditional scoreboard is green.
Now point an instrument at the engines its HCPs actually use and ask the twenty questions those clinicians actually type. Varigel appears in a healthy share of the answers, so presence looks fine and confirms the share of voice story. Then the other two layers come back. In a meaningful fraction of the mentions, an engine attaches a "commonly discussed" second use that traces to an old abstract the brand never promoted, so accuracy is lower than presence. In another slice, the contraindication is missing entirely, because the source the model leaned on summarized efficacy and skipped safety, so drift is non-zero and rising.
Two scoreboards, same quarter, opposite stories. Share of voice says the brand won. Share of Answer says the brand is being described more often, and more often incorrectly, by the machines that increasingly stand between the brand and the prescriber. The brand that only looks at the first scoreboard will celebrate exactly the quarter it should have escalated.
Three moves to start measuring Share of Answer
You do not need a perfect instrument to start. You need an honest one, built on the three layers.
1. Fix the question set before you fix anything else. Write down the twenty questions your HCPs and patients actually ask, in their words. Validate them against field and medical-information logs, not the brand plan. This set is your denominator, and it is the single most abused part of any AI-visibility report. A flattering question set produces a flattering number that predicts nothing.
2. Score all three layers, every time, not just presence. For each question on each engine your audience uses, record presence, accuracy against the label, and drift. Presence alone is share of voice with a fresh coat of paint. The accuracy and drift layers are what make the number a Share of Answer, and what make it safe to show to medical and regulatory rather than only to marketing.
3. Sample continuously and trend it, do not snapshot it. Run the set on a cadence, because the answer moves when models update, when a competitor publishes, when a new abstract surfaces. A single reading is noise. The signal is the trend: is accuracy rising or falling, did drift spike after the last data readout, did a model update help or hurt. The point of the metric is the derivative, not the snapshot.
Where this leaves you
The reason Share of Answer is hard to measure is not that the data is unreachable. It is that the two halves of the number live in two different worlds that have never been joined. Presence is a marketing measurement. Accuracy and drift are judgments against the approved label, which lives inside MLR. So the marketing tools count mentions they cannot grade, and the regulatory systems hold a label that never gets pointed at the machine. The metric falls into the seam.
Juncture is built for that seam. It already holds the approved sentence you cleared on the inside, pre-checked against the label before MLR and backed by a 21 CFR Part 11 trail and e-signature, so when its Answer Monitor measures Share of Answer on the outside it is not just counting mentions. It grades each mention against the exact approved clause it should have matched, which turns presence into accuracy and surfaces drift as a deviation from a known-good source rather than a free-floating opinion. Content reuse is the tangible payoff that funds the rest: a small approved core, recombined and shipped, is the same core the machine learns to repeat, so the number you optimize inside is the number you measure outside.
Share of voice told you how loud you were in a world where the audience read what you wrote. Share of Answer tells you how often, and how faithfully, a machine repeats you in a world where the audience reads what it writes. The brands that learn to measure the second number will manage the shift. The ones still reporting only the first will keep winning a contest the audience stopped watching.
Bring one brand and the twenty questions your audience actually asks. We will compute its Share of Answer today, presence, accuracy, and drift, across the engines that matter, and show you where the two scoreboards disagree. See it on your brand, then decide.
Sources
- WebFX, "AI Overviews in Healthcare: What Our Study of 130K+ Health Queries Reveals," 2025. webfx.com