What a healthy AI share of voice looks like
Guide · Continuous Intelligence · 4 min read · last verified 2026-07-25
Ask a marketing team what a healthy share of voice looks like and most will reach for a percentage — a third of the conversation, half the category, some round number that feels like winning. That instinct comes from the old media world, where share of voice was a spend ratio: your advertising weight divided by the category's total. AI answers do not work that way, and importing the old number does more harm than good.
In an answer engine, share of voice is not how loud you are. It is how often your brand appears in the model's response across the specific questions your buyers actually ask — and, just as importantly, whether that presence holds up when the same questions are asked again. A brand can look dominant on one flattering prompt and vanish on the ten questions that decide a purchase. So "healthy" is less a single figure than a shape: broad enough to cover the buying journey, stable enough to trust, and not propped up by a single lucky source.
Share of voice in AI answers is not one number
The first mistake is aggregating everything into one percentage. AI answers are fragmented — different assistants, different phrasings of the same intent, and different buyer questions all produce different rosters of named brands. Averaging them into a single tidy figure feels rigorous and is quietly useless, because it treats a roster you dominate and a roster you are missing from as the same event.
A more honest unit is presence per question cluster. Group the questions buyers ask by intent — what a category is, how two options stack up, which vendors make a shortlist, what could go wrong after signing — and measure your appearance rate inside each group. Collapsing everything into a single headline figure — "we hold about a quarter of the category" — hides the only thing worth knowing: where you show up and where you are absent. Presence spread evenly across every cluster is a fundamentally different position than the same aggregate arriving entirely from one branded query and nothing else. The total is identical; the health is not.
What "healthy" actually depends on
There is no universal healthy percentage, and anyone who quotes you one is selling a benchmark they made up. What counts as healthy depends on the structure of your category.
In a fragmented category with many credible players, a model will name several vendors per answer, and appearing in a meaningful fraction of relevant answers is a strong result. In a category with two or three entrenched incumbents, the model tends to name the same short list repeatedly, and a challenger's realistic near-term target is to break into the roster on a chosen subset of questions rather than everywhere at once. Your stage matters too: a seed-stage brand that appears at all on high-intent comparison questions is healthier, relative to its ambition, than a market leader who is missing from them.
The honest framing is directional. Healthy means present in the questions that matter to your buyers, trending up or holding against your named competitors, and not shrinking when the model refreshes. It is a position relative to a defined competitor set, not an absolute score.
The shape of a healthy distribution
Level tells you how much; shape tells you how durable. Three properties separate a healthy distribution from a fragile one.
Breadth: your presence is spread across intent clusters rather than concentrated in one. If every appearance you have comes from definition questions and none from selection questions, you are visible where buyers browse and invisible where they choose.
Balance across assistants: a brand that scores well in one assistant and disappears in the others is exposed to a single model's quirks. Healthy visibility holds up whether the buyer turns to Perplexity, Gemini, Copilot, or Google's AI Overviews rather than only ChatGPT, allowing for the genuine differences in how each assistant sources its answers.
Source diversity: if your presence depends on a single citation source — one review roundup, one Reddit thread — it is one edit away from collapse. Healthy visibility rests on several corroborating sources, so no single change can erase you.
Reading share of voice against a locked benchmark
A reading you cannot reproduce is a reading you cannot act on. Because model outputs drift, a share-of-voice reading is only interpretable against a fixed method: the same question set, the same competitor set, the same assistants, sampled the same way, over time. That is what a locked benchmark provides. Without it, you cannot tell a real gain from a model update or a lucky sample.
This is where Magrios starts. It measures your appearance rate across a defined buyer-question set, holds the method constant so the trend line means something, and shows the reading per competitor and per question cluster rather than collapsing it into one vanity figure. You see not just whether you are healthy in aggregate but which clusters carry you and which ones you are quietly absent from.
Turning the reading into work
A healthy share of voice is not a plateau you defend — it is the residue of a habit. The reading is only useful if it points at the next thing to fix. When Magrios shows a cluster where competitors appear and you do not, that gap becomes a specific unit of work: the question to answer, the source to earn, the page to sharpen. You act on it, then re-scan against the same locked benchmark to see whether the cluster moved. The score is the readout; the loop of measuring, closing a gap, and confirming the change is the actual instrument of growth. Watch the shape, not the trophy number, and treat every absence as an assignment rather than an embarrassment.