Share of Voice in AI Answers: How to Count It Honestly

Share of voice came to marketing from advertising, where the denominator was at least in principle knowable. Total category spend, total impressions, total column inches. You could argue about the accuracy of the total, but everyone agreed a total existed and that someone could count it.

Drop that metric into AI answers and the arithmetic looks identical. Count how often a model names your brand, divide by how often it names anyone, call the result your share. It reads like a measurement. On a slide, it behaves like one.

The problem is not in the numerator. It is in the denominator, and almost nobody selling you this number will say so out loud.

The denominator is the whole metric

Share of voice is a fraction, and a fraction is determined at least as much by what sits under the line as by what sits above it. So before you look at anybody's share, ask the question that decides its meaning: share of voice against whom?

There are two ways to answer, and both are broken, in opposite directions.

If you choose the competitor set, you are choosing the result

The natural approach is to declare your competitors up front. Pick the brands you consider rivals, count how often the model names each of you, compute your slice.

Now notice what that gives you the power to do. Add a few rivals that generative models effectively never mention. Your absolute mention count has not changed by a single instance. Your share goes up, because the denominator just got heavier with brands contributing nothing to it. Drop a rival that dominates every answer in your category, on the entirely defensible grounds that they sit in an adjacent segment or a different price tier, and your share goes up again.

Nobody has to be dishonest for this to happen, which is what makes it dangerous. The set gets assembled by someone who sincerely believes those are the relevant competitors, and each individual editing decision is arguable in good faith. But the number that comes out the far end was largely decided when the set was assembled, not when the model was queried. A metric whose value is fixed by an upstream editorial choice is not measuring the world. It is reporting the choice.

The tell is simple, and it applies to any share of voice figure that lands on your desk. Ask whether the list is written down. Ask when it last changed, and who changed it. If those answers are vague, the number is decoration.

If you let the model choose it, the denominator will not hold still

The apparently more objective alternative is to impose no set at all. Run the prompt, extract every brand the model names, let your share be your mentions over all mentions. No editorial thumb on the scale.

Except these models are probabilistic. Run the same prompt again and the cast changes. Some answers name a tight handful of vendors. Others sprawl into a long list taking in adjacent tools, an open-source project or two, and occasionally a company that no longer exists in the form the model is describing. Every one of those brands enters your denominator and dilutes your share, and none of them entered because anything about your market changed.

Your share now moves between runs for reasons entirely disconnected from your visibility, and it moves in the direction hardest to interpret: a model in a verbose mood names more brands, so your share falls, so it looks like you lost ground on a day when nothing happened at all.

An open denominator is not more objective than a chosen one. It has outsourced the choice to a system that makes it differently every time and will not tell you why.

How to count it defensibly

None of this makes the metric worthless. It makes it conditional, and the condition has to be visible.

Declare the competitor set, in writing, and freeze it. The set is a published assumption, not a private one, and it belongs next to the number so that anyone reading the share can see what it is a share of. Changing it is allowed. Changing it silently is not, and any change resets the series: the figures before and after are not comparable and should not sit on the same line.

Build the set on adversarial grounds, not flattering ones. The brands you least want in your denominator are exactly the ones whose exclusion turns your number into fiction. If a company keeps appearing in the answers you care about, it goes in the set, whatever your market map says about tiers and segments.

Report both denominators, separately. Your share against the declared set answers one question: how do I stand among the companies I decided I compete with. A tally of the unexpected brands the model named, kept as its own figure, answers a more urgent one: who is being recommended in my place by companies I had not thought about. Blend them into a single number and you lose both.

Many runs, many prompts, or nothing. A share computed from one answer is not a share. It is a single sample dressed up as a population. Each prompt has to be run repeatedly, and the share computed over the whole body of results. If the figure swings wildly between runs, that instability is a finding and should be shown, not averaged away into a reassuringly flat line.

Read it as a trend, never as a value. The absolute number is an artifact of your denominator and means nothing on its own. What means something is the same number, computed the same way, against the same frozen set, moving in a direction over time. Treating a single reading as a statement about your position in the market is the error the metric invites, and the one it never survives.

What the metric cannot see

Even counted honestly, share of voice is a coarse compression.

Consider two answers. In the first, a model opens by naming your brand as the standard choice in the category and describes it accurately. In the second, your name appears in the tail of a longer list, characterized as the cheap option people settle for when budget is tight. Both are a mention. Both increment the numerator identically. Share of voice cannot tell them apart, and the distance between them is most of what you would actually want to know.

The same flattening hides more than that:

• Position within the answer. Named first and named last read very differently to a human, and identically to the metric.
• Framing. The default, the challenger, the budget option, the legacy incumbent. All of them are simply mentions.
• Accuracy. A model that names you and describes a product you do not sell has raised your share of voice while damaging you.
• Which prompt produced it. A mention on a buying question and a mention on an idle definitional one are not worth the same, and averaging them makes both unreadable.
• Language and market. A global share is the average of a strong home market and near-total absence abroad, and it describes neither. Visibility does not travel across languages, and a share computed across them conceals the exact gap an exporter needs to see.

So it is a low-resolution instrument for detecting movement, which is genuinely useful, and not a certificate of standing, which it cannot issue.

What it is actually good for

Against a stable set and a stable prompt list, run repeatedly, a share that trends upward over months tells you the model's answer space is opening to you. A share that erodes while your raw mention count holds steady tells you a competitor is gaining rather than that you are slipping, which is a different problem needing a different response. A share that is respectable at home and collapses in your export markets tells you where the work is, and that is the thing this metric almost never gets used for.

None of those readings requires the number to be precise. All of them require it to be consistent.

Where the counting gets done

Consistency at this scale is not a manual exercise. Holding a competitor set stable across four platforms and every language you sell in, repeating it on a fixed cadence and keeping the results comparable, is a data problem rather than a browsing session.

That is the work PSentry is built to do: run your prompts across ChatGPT, Claude, Gemini and Perplexity, per language and per market, record which brands are named alongside you, and show how it moves.

A piece built on honest denominators owes the same honesty about its own numbers: the share reported is observed, not engineered. Nothing here nudges a model toward naming you, and no rising slice is promised in advance, since the denominator moves for reasons no vendor controls. The scans behind it run twice a month, which is why this reads as a trend, not a feed to refresh.

Frequently Asked Questions

Should the competitor set be the same in every market?

Usually not, and forcing it to be is a reliable way to produce a meaningless number. The brands a model recommends in one language are frequently not the ones it recommends in another, because the local web is different. Keep a declared set per market, and accept that shares from different markets are not comparable to each other. Each is comparable to its own history, which is the only comparison this metric supports anyway.

Can share of voice be compared across platforms?

With caution, and never as like for like. These systems retrieve and generate differently, some leaning on live pages and others on what they absorbed in training, and they name brands with different degrees of verbosity. A share on one platform and a share on another are two separate series. Watch each one move. Do not subtract one from the other.

Is a rising share good news if nothing else changed?

Not necessarily, and this is the trap the denominator sets. Your share can rise because a competitor's coverage decayed, because the model turned terser and named fewer brands, or because somebody quietly edited the set. Before celebrating, check whether your raw mention count moved. If it did not, the movement happened underneath the line, not above it.

Does a share of zero mean the model has never heard of my brand?

It means the model did not name you in answers to those prompts. Those are different claims. Ask directly about your brand by name and you will usually get a description, because the system will find something to say. Absence from a recommendation answer is a statement about your standing among the options it considers, not about whether the name exists in its memory, and confusing the two leads to fixing the wrong problem.

How often should the number be recomputed?

Slowly enough that you are reading a trend and not noise. These systems do not change their view of a brand quickly, because that view rests on a slow-moving accumulation of pages and coverage. Recomputing constantly produces a jittery chart that mostly reflects the probabilistic behavior of the models, and tempts you to react to movement that was never there.