How AI Share of Voice Is Calculated: Five Cases That Disagree

Type a share-of-voice number into a leadership deck and it reads as a fact. It is 42 percent, it was 38 percent last quarter, the line is going up. The trouble is that the same set of AI answers, counted by two honest tools, can produce two numbers that are far apart, and the gap almost never comes from the counting.

AI share of voice is a division. It is your brand mentions over all brand mentions across a set of prompts sent to AI answer engines, written as a percentage. That part is settled. What is not settled is the denominator: which brands are in the set, whether non-brand entities get filtered out, how each prompt is weighted, and how many times each prompt is run. Move any one of those and the number moves while your real visibility sits exactly where it was.

This post walks five cases where the same answers give different numbers. Each one carries its figure and its source, and each one is a denominator or a sampling decision rather than a mistake in the arithmetic.

AI share of voice is the count of answers that mention your brand divided by the count of all brand mentions in the same set of answers, turned into a percentage. If a competitor set of ten brands is named across a hundred answers and you appear in thirty of those mentions, your share is thirty percent. Nothing about that is hard.

The hard part is every decision that happens before the division. A share is a fraction, and the argument is always about the bottom of the fraction, not the top. The five cases below are five different ways the bottom of the fraction gets built, and each one produces a number a reader will assume means market position when it means something narrower.

Run the same thirty mentions through two defensible denominators and watch the headline move. Count your thirty mentions against a competitor set of five brands that together collected ninety mentions, and your share is 33 percent. Count the identical thirty against a broader set of twelve brands that collected one hundred and eighty mentions, and the same visibility reads as 17 percent. Nothing about the brand changed between those two numbers. One analyst drew a tighter competitor set than the other, and the fraction did the rest. That is the whole problem in one calculation: a share number carries a hidden argument about who counts, and the percentage on top never shows it.

The cleanest way to read a perfect score is to divide by almost nothing. A brand monitoring dashboard can report a 100% AI share of voice when an empty competitor set is defined, because the denominator holds nothing but the brand itself. Google's own Merchant Center documentation describes the behaviour, and a practitioner flagged it plainly: a 100% reading is not proof of market dominance, it is proof that nobody else was in the set.

The figure to watch here is the one that looks best. A share of an empty market is arithmetically real and analytically worthless. Before a share number means anything, the competitor set has to exist and has to be named beside the result, so a reader can see what the brand was measured against.

This case is common precisely because the default is friendly. A fresh dashboard with an empty competitor list does not refuse to compute a share. It computes the one it can, which is your mentions over your mentions, and prints 100%. The number then travels into a weekly report, and the first competitor a buyer adds the following month drops it to something ordinary, which reads as a decline that never happened. The honest version of the 100% case is not a number at all. It is a note that says the competitor set was empty, so there is nothing to divide against yet.

The next case is subtler, because the denominator is not empty, it is contaminated. A displayed share-of-voice list for a crowded category put the leading brand at 9.5%, then listed Caffeine at 4.6% and the FDA at 2.5% in the same ranking. Caffeine is an ingredient and the FDA is a regulator. Neither is a competitor, and both are sitting in the denominator the leading brand is being divided against.

Every non-brand entity left in the list drags the real brand shares down, because the fraction is being divided by a larger bottom than the market actually contains. Filter the list down to brands and every genuine share rises. The fix is a cleaning step before the count, and the cost of skipping it is a number that looks more competitive than it is.

Entity contamination is harder to catch than an empty set, because the number looks plausible. Nobody questions a 9.5% share the way they question a 100% one. It takes reading the full list to notice that two of the three names beside the leader are an ingredient and a regulator, and most reports never show the full list, only the leader's figure. The reason the contamination happens at all is that the extraction step, which pulls named things out of an answer, does its job well and pulls out everything named: brands, ingredients, agencies, standards bodies. Turning that raw entity list into a competitor set is a separate decision, and a tool that skips it is not wrong about what it found, only about what it called a competitor.

The third case moves the number with no change to the answers at all. On 2026-08-31 a widely used tool switched its AI share of voice from plain Google search volume to an AI-adjusted volume base, so a prompt's weight in the denominator changed overnight. The vendor said it directly: some prompts look bigger, some smaller, the gaps between you and the brands you track have shifted, and nothing about your real visibility changed. The denominator did.

A weighting base is a design choice, and when it changes it drags every share with it. The published methodology for AI-adjusted volume and the calculation walkthrough are worth reading before comparing a number from after the switch to one from before it. Record the base and its date next to any share you report, or a base change six weeks from now reads as a visibility swing that never happened.

Weighting is reasonable in principle. A prompt that millions of people ask should count for more than one almost nobody asks, and adjusting for how much each engine is actually used relative to Google is a defensible attempt at that. The hazard is not the weighting, it is the silent re-baseline. A prompt's number on one engine and its number on another are no longer the same number after the switch, so a chart that spans the change date shows a step that belongs to the method, not to the brand. A brand sitting at 30% of the denominator before the switch can read 24% after it with no change in its own visibility, because the base under the fraction moved. The only defence is to timestamp the base, because a reader six weeks later has no way to separate a real move from a definitional one unless the report tells them a definition changed.

The fourth case is a sampling decision wearing a visibility number. A builder tracking AI visibility noticed that four engines returned the same score, run after run, within a point of each other. The number never moved. The cause was in the prompt set: the prompts contained the brand name, so the model was handed the answer and repeated it. A metric built on prompts that name the brand has a floor and a ceiling at the same place, and it cannot move regardless of what the brand does.

When the team split the prompts, keeping some branded and moving the rest to unbranded category queries, one site dropped from 47 to 13, a fall of 34 points on the same scale. The 13 is the honest figure, because it measures whether the brand surfaces on questions that do not already contain its name. A share number built on branded prompts is measuring an echo.

The distinction matters because the two question types answer different things. A branded prompt, such as asking an engine what it knows about a named product, tests whether the model can describe you correctly once it has been pointed at you. An unbranded prompt, such as asking for the best option in a category, tests whether the model reaches for you unprompted, which is the thing a buyer's own question actually looks like. Both are worth tracking, but folding them into one share hides the one that is hard to move behind the one that is nearly guaranteed. Reported separately, a high branded score and a low unbranded score tell a clear story. Blended, they tell a flattering one.

The last case is the one every single-pass score sits on without showing it. AI answers are not deterministic, so the same prompt can return a different answer on the next run. We measured this directly. Our own step-zero run sent 12 money queries five times each against one model on one day. 60 of 60 calls succeeded, and 7 of the 12 queries flipped which brand was cited across repeats that were identical by construction.

A separate permutation test ran 20,000 random splits of that same query set with no change applied at all, and the null difference centred on zero at a standard deviation of 0.215. The full method and the raw numbers are published in why a single AI visibility score is noise. A score that runs each prompt once reports a single draw from that spread as if it were the centre of it, which is how two honest runs of the same prompt set produce two different shares.

The practical reading of a 0.215 standard deviation is that two measurements taken minutes apart, with nothing changed, can sit most of a fifth of the scale apart purely from the model answering differently. With twelve queries, a single query flipping is more than eight points of movement, so a week-over-week change smaller than the spread is indistinguishable from the engine having a different afternoon. The fix is not exotic. Run each prompt several times, report the average and the spread together, and treat any movement inside the spread as noise until more passes say otherwise. A number with no spread beside it is hiding exactly the quantity that tells a reader whether to believe the trend.

None of the five fixes needs a new tool. They are reporting choices, available in any dashboard that shows its working, and together they turn a share number into something a reader can act on rather than a figure that quietly depends on four decisions nobody wrote down.

State the competitor set, filter the non-brand entities, freeze the weighting base and its date, and repeat each prompt while keeping the engines separate. For the deeper version of this, our measurement procedure writes up how we baseline and re-measure, and the best AI visibility tools comparison covers which tools expose the denominator at all.

A share number is only as good as the method behind it, and there are three honest ways to produce one. A spreadsheet answers whether you appear at all. A monitoring tool answers how the number moves. Only a held-out control answers whether your work caused the move, because it keeps a set of queries deliberately unworked as a comparison.

That last rung is where answer engine optimization run against a control earns its cost. If you want the denominator behind your current number checked, or a control run against it, talk to us. It is the one question a dashboard cannot answer about its own score.

Every figure in this post is either read from a named public case or measured by us, and the read date for all of them is 2026-10-06. The 100% reading, the 9.5% list, the 2026-08-31 base change and the move from 47 to 13 are published cases, each linked in its own section (the 100% case cites Google's Merchant Center documentation) and recorded as reported rather than re-derived by us. The repeat-variance figures, 60 of 60 calls, 7 of 12 queries flipping, and the 0.215 null standard deviation, are our own and published with their raw numbers.

The ranking in this post is by how early each decision enters the pipeline, from building the set down to sampling it, not by how often each failure occurs. That is the main limitation: these are specific, dated cases, and the post does not measure how common each one is across the market, nor does any figure here describe a lift caused by any work. Our own design is underpowered by its own power analysis, with a minimum detectable lift of 51.1 points, so it supports a claim about noise and a claim about the null, and nothing about causation.

Citon publishes its measurement runs, including the parts where the instrument is still too blunt to prove what a buyer would want it to prove. A share-of-voice number is cheap to produce and easy to misread, and the five cases above are the ones that turn a reassuring percentage into a figure that cannot survive the question a careful reader will ask: what, exactly, is in the denominator.