Share of voice in AI search: measuring visibility against rivals

Thursday, 11:40. A marketing lead at a B2B software company types her category's most common question into an AI assistant and reads the answer: four brands named, hers not among them. She screenshots it and drops it into the team channel. The next morning she asks the same assistant the same question, and this time her brand shows up in the third sentence. She now has two observations that contradict each other. The difference is not the result of overnight content work; it is the ordinary variance of a single-sample reading.

Share of voice exists to close exactly that gap. Instead of the impression left by one answer, it gives you a pooled ratio across a defined question set and a defined brand set. This piece walks through the whole measurement in order: what sits in the numerator and denominator, who decides the competitor set and on what evidence, how a question set is designed, how many runs it takes before a number means anything, why mention share and citation share are kept apart, when weighting is legitimate, whether six engines can collapse into one figure, how to read a competitive matrix, what the diagnostic order is when the number falls, and which decision the ratio is finally attached to.

What division does share of voice actually represent?

Share of voice is a pooled ratio inside the brand set you track: your mentions divided by the mentions of every brand in that set. The numerator is trivial. The entire argument lives in the denominator. Add two more brands to the same measurement window and your share drops; remove one and it rises — with nothing at all changing on your side. That is why the figure is not an absolute grade but a comparison ratio that only works when the basis is declared out loud.

"Our share of voice is thirty-four percent" says nothing on its own. It becomes a sentence when three facts travel with it: which brand set, which question set, which window. In Rocketly's GEO product the Share of Voice and Competitor Comparison view refuses to hide the denominator — it prints the set you declared next to the ratio, and when you change that set the break in the series stays visible. Reading the number before and after a denominator change as one continuous line is the fastest way to corrupt your own measurement.

You also need to fix the counting unit before you divide anything. If your brand appears three times inside one answer, is that three mentions or one presence? In practice the binary version is sturdier: one count per answer stops long, chatty responses from inflating whoever repeats their own name. What matters more than the choice itself is that you never quietly switch it halfway through a quarter.

Which competitors belong in the denominator, and who decides?

Build the competitor set from the answers themselves, not from the sales team's memory. The company you actually compete with in deals and the company the model puts in the same sentence as you are often not the same company. Over the first two or three weeks of measurement, collect every brand name that appears in your answers, rank them by frequency, and take the top five to eight as the candidate set. That list usually holds two surprises: a marketplace you never thought of as a competitor, and a newcomer nobody in the room had heard of.

Resist the pull to make the set as broad as possible. With fifteen brands in the denominator everyone's share shrinks and the ratio becomes too dull to react to a real change. With only three, it lurches on every small fluctuation. Five to eight is a workable range in most categories, sensitive enough to move and stable enough to trust. Draw the boundary of the set by which brands the questions actually surface, not by how your category is defined on a slide.

Once the set is fixed, log every change to it: date, who added which brand, and why. Head-to-head prompts feed that list naturally, since the same pairs keep appearing together — and because staying inside the answer on comparison queries follows its own logic, the names that surface there are the most honest source for your denominator.

Which question set do you measure with?

After the denominator, nothing shapes the number more than the question set. Measure only with "best of" prompts and you will flatter a brand that lands in category lists but loses every head-to-head. A durable set covers four question types in balance, each one mapping to a different moment in the buying process.

Question type	Example pattern	What it measures	Reasonable share of the set
Category discovery	"best tools for ..."	Ability to make the list	About one third
Head-to-head	"X or Y, which fits ..."	Position in a direct match	About one quarter
Problem solving	"how do I ..."	Strength as a cited source	About one quarter
Brand validation	"is X reliable, what does it offer"	Tone and attributes	The remainder

Treat those proportions as one worked example, not a rule; shift them toward whichever moment dominates your funnel. Set size is a decision too: thirty questions produce enough signal for a weekly read, while a hundred hold up once you split by market and language. Rocketly's Question-Level Tracking view puts each question under tracking with its own language and market and shows your remaining quota on the same screen; draft ideas are suggested from search demand, but nothing enters tracking until you approve it, so the set cannot quietly grow and dilute your denominator.

Write the prompts in the customer's own words, and keep at least half of them free of your product name. A question containing your brand is one you are expected to win, and a set stuffed with those inflates the ratio artificially. The same reasoning behind question-and-answer pages that get cited applies here: when the prompt you measure and the heading on your page speak the same language, measurement and content finally line up.

How many runs before a share number means anything?

Generative engines can answer the same prompt differently on the same day. A single run is therefore an observation, not a measurement. What makes share of voice meaningful is repetition: running the same set on a schedule and pooling results across a window. A practical starting cadence is at least one run per question per week, read over a rolling thirty-day window — an example setup, not a standard, which you tighten once you can see your own volatility.

There is a simple way to size that volatility. For the first fortnight, run the same question twice on the same day and look at how far the two results diverge. If the gap is wide, widen the window, add questions, or do both. In a small set a single answer flipping can move total share by several points, and mistaking that wobble for a trend puts either a fictional win or a fictional crisis into the board pack.

Days with no measurement are a separate problem, and the most common error sits right here. When a provider does not respond or the allowance runs out, that day has no data; counting it as zero drags the series down and manufactures a decline. The GEO product's An Unmeasured Day Is Not Zero principle prevents exactly that: the chart breaks at that point, the ratio shows a dash rather than a zero, and the number of unmeasured days is stated separately. Change Tracking answers "it named us last week and not this week" by showing the difference between two runs, without dressing single-sample noise as change.

Why are mention share and citation share calculated separately?

Your name appearing in the text of an answer and your domain being listed as a reference underneath it are two different events. Blending them into one ratio makes a brand that gets named constantly but never visited look successful. So calculate two shares: mention share and citation share. The gap between them tells you, on its own, what work comes next.

There are four combinations and each one leads to a different task list. Mentioned and cited: defend the position. Mentioned but not cited: a page structure and quotability problem. Not mentioned but your content is feeding the source list: your brand name is not stated clearly enough inside your own content. Neither: you are simply absent from the category and need to publish. Rocketly's Mention or Citation view counts the two without mixing them, and Evidence-Backed Opportunity Cards grow out of that same split — "a competitor was named on this question and you were not" and "your name appeared but your domain was not cited" are deliberately separate cards.

High mention share with low citation share usually points at weak trust signals rather than thin content. The discussion of what makes an engine choose one brand as its source is the right place to start when you want to close that gap rather than widen your publishing calendar.

Should every question count equally?

A flat average treats every question in the set as equal. That is the most honest and most defensible form of the measurement, but on its own it may not carry a budget decision. When a niche prompt asked a handful of times a month weighs the same as the busiest question in your category, the ratio can push you toward the wrong investment.

The reasonable source of weighting is demand data. The GEO product's Demand Signal puts historical Google Ads volume and Search Console impressions side by side for a question, and states plainly that this is decision context, not a GEO measurement. Demand does not change your share of voice; it helps you choose which questions to attack first. Melting the two into a single composite destroys the explainability of both.

The healthiest practice is to report both: raw share of voice and demand-weighted share of voice. The distance between them is a finding in itself — a high raw share with a low weighted share means you are missing from the crowded questions. The roundup of which KPIs belong on a GEO dashboard is useful for deciding what context fields should sit next to those two numbers.

Can six engines produce a single share number?

Engines answer differently: one returns a list, another recommends a single vendor, a third writes a generic explanation naming nobody. The primary read is therefore always per engine. If you do want a combined figure, keep the weights and their revision dates written down; a blended score is a versioned methodology, not a casual average.

Folding an engine you do not measure into the denominator is its own mistake. Rocketly's Six-Engine Catalog keeps Google AI Overview, ChatGPT, Perplexity, Gemini, Claude and Grok in one list, and each row states whether measurement is live for that engine or an integration is still pending; appearing in the catalog does not mean it is measured, and no number is produced for an engine that is not. Finding the engine where a rival consistently lands ahead of you is the quickest route into why an assistant recommends a competitor instead of you.

Put a market and language split next to the engine split. The same brand can be strong in one language and invisible in another, and that is usually a shortage of sources in that language rather than a content problem. Where answers are not purely text, the scope of measurement widens too: the conditions behind visibility in multimodal search directly affect who gets named in image-bearing answers.

How do you read the competitive matrix?

A single percentage never tells you where the work is. The real read is a question-by-brand matrix: tracked questions down the rows, the brands in your set across the columns, mention and citation status in the cells. That grid opens up the distribution behind the total — a thirty percent share can come from being dominant in twelve of forty questions or from being mediocre in all forty, and those two situations demand completely different work.

Pattern in the matrix	How to read it	First action
Very strong on a few questions, absent elsewhere	Narrow but deep visibility	Extend into adjacent question types
Mentions in many questions, almost no citations	A quotability gap	Fix page structure and answer blocks
The same rival ahead of you in eight questions	A positioning issue, not eight issues	One comparison asset, not eight
Your count is flat but the denominator grew	A new entrant in the category	Update the competitor set log
No brand at all is named on a question	The engine does not recommend vendors here	Consider dropping it from the set

The most valuable use of the matrix is collapsing "the same competitor is ahead of us in eight questions" into a single item of work. The GEO product's Single Action Plan is built that way already, ordering work by dependency rather than by score. If your product pages sit empty across whole rows, the problem may be page format rather than page count — the reasons behind an assistant skipping your product page usually cluster in heading order, summary sentences and missing structured data.

What is the diagnostic order when share falls?

The first place to look is the measurement, not the content. Did the allowance run out, did an engine integration drop, how many days in the window are empty? Rocketly's Measurement Allowance and Health screen splits "no data arriving" into four distinct answers, and Alert Rules include a separate rule type for measurement gaps; keeping a drop alert and a gap alert apart is what saves you a week of chasing a false alarm.

Second comes the denominator. If a brand was added to the set, or a rival simply started appearing in answers, your share falls even when your own performance is unchanged. Third is the source side: the domains feeding the engine's answer may have shifted. The question of how much space forums and community sites occupy in cited sources can explain why your share erodes while nothing on your own site changed at all.

Fourth is how your brand is represented in the model's knowledge base. What category it files you under, and with what description, can drift over time, and stale entries in encyclopedic sources feed that drift. The account of how brand facts are learned from Wikipedia and Wikidata shows how that layer behaves. Finally, read the tone: being named and being named as "expensive but thorough" are not the same outcome, and Tone and Attribute Reading separates positive, neutral, negative and mixed — when the answer text could not be read, it leaves the field empty rather than defaulting to neutral.

Which decision does this ratio attach to?

If share of voice is not decoration on a dashboard, it has to end in a decision. The three it attaches to most directly are content sequencing, splitting budget between classic search and generative engines, and making agency work reportable. All three need the same discipline: period-over-period comparison on the same denominator, the same question set and the same window.

Do not leave reporting dependent on one person. The ready-made GEO visibility report can be taken as PDF, Excel or JSON, turned into a shared link, or sent automatically at whatever frequency you set; the shared report carries no raw answer text and no provider load, so the document a client receives shows the same numbers as your internal screen without carrying the same evidence. Be careful when linking share to traffic as well: AI Referral Traffic figures sit beside visibility, never as its cause. The finding that visitors arriving from AI search convert differently also explains why share and sessions rarely grow at the same rate.

The right moment to start measuring is the moment you realise history cannot be backfilled: generative engines have no Search Console-style archive, so your series begins the day you start. If you want to set your own denominator against your own question set, open a Rocketly account and define your brand, your competitors and your first prompts to start the series today.

FAQ

What share of voice counts as good?

There is no universal threshold, because the ratio depends entirely on the denominator. In a five-brand set, an even split is twenty percent; in an eight-brand set it is twelve and a half. The meaningful target is direction against your own baseline and your rank inside the set, not an absolute number.

How many questions are enough to track?

Around thirty questions produce enough signal for a weekly read in most categories. If you track several markets or languages, each split needs enough questions of its own, otherwise the per-split ratio swings wildly on very little evidence.

Does adding a competitor break my historical data?

Past measurements are not lost, but the ratio loses comparability. Record every set change with its date and, where possible, run period comparisons on the same denominator; because the screen marks where the set changed, the break is easy to explain in a report.

Is share of voice the same thing as search ranking?

No. Ranking measures your position in a list of links; share of voice measures how often you appear inside a generated answer. The same content investment can move both, but the units, the data sources and the reasons for a decline are different.