Share of voice came out of media buying. You could count the ads running in a category, count yours, and get a fraction that meant something, because the inventory was finite and everybody could see it.
People have started borrowing the phrase for AI answers. The instinct behind it is reasonable. The arithmetic underneath it does not carry over.
What the phrase is borrowing and what it leaves behind
In its original setting, the measure worked because the denominator was knowable. Everyone competing for attention in a category showed up in the same finite set of places, and you could count them.
An AI answer has no denominator. There is no fixed set of slots. A response to one question might name six vendors. A response to a slightly reworded version of the same question might name two, and neither of those sets is the population of companies that could plausibly have been named.
So when people talk about share of voice in AI answers, they are describing something closer to a felt sense than a fraction. Across the questions that matter to us, how often are we in the room, and how much of the room do we take up when we get there?
Still worth tracking. It just does not resolve into a percentage the way the old version did.
Why a single answer tells you almost nothing
Ask the same question twice and you will often get two different answers. Ask it on a Tuesday and again on a Friday and the set of names can shift. Change one word in the phrasing and the framing moves with it.
That variance is the first thing that trips teams up. Someone runs a prompt, sees a competitor named and themselves left out, and treats it as a finding worth escalating. Someone else runs the same prompt an hour later and gets close to the opposite.
Neither result is wrong. A single sample drawn from a system with this much movement in it is simply not evidence of anything yet.
Sampling instead of checking
The way most teams get to something usable is by turning the check into a small routine and then leaving it alone long enough to be informative.
- Write a set of prompts that match how buyers actually ask, in their words, including the messy ones with a paragraph of context attached
- Run each one repeatedly over time rather than once
- Run them across more than one system, because they behave differently from each other
- Record what came back, not only whether your name was in it
What you are building is a sample, and like any sample it gets more honest with repetition. One run is an anecdote. The same prompt run twenty times over a few weeks starts to describe a tendency, and a tendency is something you can act on.
Presence is the easy half
Counting appearances is straightforward. Reading them takes longer and matters more.
When your name comes up, look at the sentence wrapped around it. Are you described the way you would describe yourself? Are you in the right category? Are you the recommendation, or the afterthought at the end of a list? Is the use case attached to you one you actually serve well?
Teams often discover their appearance rate is fine and the characterisation is where the trouble is. Being named as the cheap option for small teams when you sell into enterprise is a positioning problem wearing a visibility costume. It will not improve by publishing more of what you already publish.
That reading also tells you what to write next. If the description keeps missing something true about you, the material that would have corrected it is usually either missing or buried three clicks deep on a page nobody links to.
Reading it next to everything else you can see
Two other things sit alongside this and give it context.
The first is your own funnel. If direct and branded traffic holds steady while general search traffic softens, research is probably happening somewhere you cannot watch and finishing at your door. The wider conversation about the shift away from traditional search traffic tends to fix on the decline itself, though the part that changes what you do is where the step went rather than that it went.
The second is where you are in your own work. A team that has never once checked how it appears in an answer is at a very different point than a team running a stable prompt set every month and rewriting pages off the back of it. Most of the value in building AI-search maturity as a sequence comes from that ordering, so you are not measuring something with precision before you have made any attempt to influence it.
Holding it loosely on purpose
Treated as a directional read, this is genuinely useful. Treated as a scoreboard, it will mislead you, because the thing you are sampling shifts underneath you and the denominator was never real to begin with.
The teams that get value from it keep their grip light. They track the same prompts long enough for drift to become visible. They pay more attention to how they are described than to how often. And they treat a gap in the answers as a signal about their own material, which usually means somebody goes and writes the page that was missing. What gets reported upward is the change in description, which moves slowly and is harder to fool than a count.
