AI share of voice: calculation, examples and limitations

Where share of voice comes from, and what changes in AI answers

The concept was born in media planning. In the definition Nielsen still uses, share of voice is a brand's slice of total category media spending in a given market, channel and period. The metric gained weight once it was tied to market share: Binet and Field's research on 171 campaigns linked a share of voice above a brand's market share to growth. The denominator there is money spent, a number that exists regardless of who is measuring.

SEO adapted the term for organic search. With no media spend to divide, organic share of voice is estimated from rankings across a keyword set, combining each position with an expected click-through curve and with search volume. The denominator became the total traffic available for that set of terms. It is still an estimate, but it rests on two external quantities: position and volume.

In AI answers, neither quantity exists in the form SEO knows. There is no fixed position: the brand appears inside a paragraph, sometimes in a list, sometimes in a sentence of caveats. To weight your company's basket, what is missing is prompt volume that is at once primary, published by the platform owners, audited, and specific to your audience at question or category level. Third-party estimates segmented by topic and platform do exist, based on licensed panels and modeling; they depend on the sample, consent, coverage, and method used. Keyword volume, in turn, does not tell you how many people ask an AI that question, and the answer changes from one run to the next.

The denominator of AI share of voice is therefore a sample you build yourself: which prompts, on which platforms, in which language, over which period, against which brands. Share of voice in media describes the market; share of voice in AI describes your collection. The number is only interpretable alongside a description of the collection, and only comparable with another collection made under the same conditions.

Visibility and share of voice are not the same calculation

The two metrics appear together and measure different things. Visibility answers a question about reach: in how many of the valid answers did the brand appear? Share of voice answers a question about competition: of all the brand mentions in those answers, how many went to yours?

Dimension	Visibility	Share of voice
Question	In how many answers does the brand appear?	What is the brand's slice of all brands mentioned?
Numerator	Valid answers mentioning the brand	Brand mentions (one per answer)
Denominator	Total valid answers	Total mentions of every brand considered (one per brand per answer)
Sum across brands	Does not add up to 100%; several brands can share an answer	Adds up to 100% across the brand set
What moves it	A change in the AI's answers or in the prompt set	Also the list of brands compared, even with no change in any answer

"Valid answer" needs a definition. A failed collection, an empty response or a Google result with no AI Overview are not answers that omitted the brand; they are observations with no AI answer. Google shows AI Overviews only when the system finds them additive to classic Search, so they often do not trigger. Counting those absences as answers without the brand lowers visibility for no reason; excluding them and reporting how many there were is the honest practice.

The denominator question also applies to the word "mention." In the definitions in this entry, each brand counts at most once per answer, even if the text names it three times. This stops one long, repetitive answer from outweighing ten short ones. It is a choice, not a law; other tools count every occurrence, and a metric with the same name can have a different denominator.

Example 1: one collection, two metrics and a denominator that moves

The following example is hypothetical. Suppose a panel of 50 prompts about e-signature platforms, run on four AI platforms, for 200 collections in total. Ten failed or came back empty, leaving 190 valid answers. Four brands are tracked: yours (Brand A) and three competitors.

Brand	Answers with a mention	Visibility (÷ 190)	Share of voice (÷ 285)
Brand A (yours)	57	30.0%	20.0%
Brand B	95	50.0%	33.3%
Brand C	85	44.7%	29.8%
Brand D	48	25.3%	16.8%
Total presences	285	—	100% (99.9% after rounding)

Brand A's visibility is 57 ÷ 190 = 0.30, or 30.0%. Its share of voice is 57 ÷ 285 = 0.20, or 20.0%, because 285 is the sum of all presences (57 + 95 + 85 + 48). The visibilities add up to 150%, which is not an error: 285 presences across 190 answers means 1.5 brands per answer on average. The shares of voice add up to 100% by construction.

Now suppose the team, reading the answers, notices that a fifth brand shows up often and decides to track it. Brand E has 57 presences, the same as Brand A. Nothing changed in the answers; the brand list changed. Total presences become 342 (285 + 57), and Brand A's share of voice drops to 57 ÷ 342 = 16.7%. Its visibility is still 30.0%.

The drop from 20.0% to 16.7% is 3.3 percentage points, which is a relative change of −16.7% ((57 ÷ 342) ÷ (57 ÷ 285) − 1). A report that presented this as "we lost 17% of our share of voice" would be describing a configuration change, not a change in AI behavior. This is why the composition of the brand set has to be recorded next to the number.

Which denominator: tracked brands or every brand the AI names?

There are two families of denominator, and each has a price.

A closed set uses only the brands you decided to track: yours and a defined group of competitors. It is reproducible, easy to explain and stable across periods, as long as the list does not change. The cost is blindness to whatever was not registered: if the AI starts recommending a new brand that is not on the list, your share of voice does not register the loss. That is why reading the answers remains necessary even when the metric is automated.

An open set tries to count every brand the AI mentions, which requires extracting entities from the text. It captures unexpected entrants and describes the whole conversation better, but it depends on a classifier that makes mistakes: generic names, products named as a negative example, brands from another category mentioned in passing. The denominator grows and shrinks as the AI wanders, and a period-over-period comparison starts to reflect the extractor's instability as well.

At Lastfind, the definition is closed and specific to the product: visibility divides valid answers with a mention by total valid answers, and share of voice divides the brand's mentions by the total presences of the brands registered in the project, counting each brand once per answer. The comparison covers up to ten competitors you choose, not the whole market, and a mention is detected by the brand name in the final answer text, excluding link addresses, reasoning summaries and the organic results beside an AI Overview. Brand and competitor settings are stored with each answer, so editing the project today does not rewrite what was observed yesterday. If your question is "who else is the AI recommending that I am not tracking yet," the answer is not in the metric; it is in reading the stored answers and registering the competitor that shows up. That discovery process belongs to competitor monitoring in AI answers.

Whichever family you choose, the rule is not to mix them. A share of voice computed with four brands cannot be compared with one computed with five, and one extracted through open entity recognition cannot be compared with one from a closed list.

Weighting options and what each one distorts

The basic formula treats every mention as equal. Three weightings are common, and each answers a legitimate question at the cost of a distortion you need to know about.

Weighting by position in the answer starts from the idea that being the first brand named is worth more than being the fifth. A common scheme assigns decreasing weight to order of appearance. The problem is that AI answers do not always order by preference: the first brand may be the one the user typed into the prompt, the one used as a counterexample ("unlike Brand B, which does not offer..."), or the one that opens an alphabetical list. Position weighting rewards order without checking why the order exists, and penalizes brands that appear at the end of long lists even when they are the recommended ones. If it is used, position should be reported separately, and a sample of answers should be read to confirm that order means preference for that kind of prompt.

Weighting by platform tries to reflect that one platform has more users than another. Third-party volume estimates by topic and platform exist, but they are not a primary, audited measure specific to your company's audience. Turning them automatically into weights imports their panel and modeling choices into the metric. Beyond that, a platform that answers with many brands dilutes everyone, while one that answers with a single brand concentrates the share on whoever won; adding the two into one number mixes different mechanisms. An auditable alternative is to present share of voice per platform, side by side, and leave aggregation to an explicit reading such as "the simple average of the four platforms" or "the priority platform only."

Weighting by topic makes sense when one group of prompts matters more to the business than another: purchase-comparison prompts, for instance, are worth more than definition questions. The distortion appears when the weights are set after seeing the results, favoring the topic where the brand did best. Topic weighting is only defensible if the weights are fixed before collection and kept constant across periods.

There is also the quiet choice of counting every occurrence of the brand instead of one presence per answer. Counting occurrences favors long, repetitive answers and turns a text that names the brand five times into five "voices." One presence per answer is more conservative and closer to what the buyer experienced: they read one answer, not five.

Example 2: two periods, same basket — visibility rises while share of voice falls

This example is also hypothetical and continues the previous one, with the four original brands and 190 valid answers in each period. In period 2, Brand A appeared in more answers, but Brand B grew even more.

Brand	Period 1: mentions	Period 1: SOV	Period 2: mentions	Period 2: SOV
Brand A (yours)	57	20.0%	66	19.6%
Brand B	95	33.3%	120	35.7%
Brand C	85	29.8%	95	28.3%
Brand D	48	16.8%	55	16.4%
Total	285	100%	336	100%

Brand A's visibility rose from 57 ÷ 190 = 30.0% to 66 ÷ 190 = 34.7%, a gain of 4.7 percentage points. Its share of voice fell from 57 ÷ 285 = 20.0% to 66 ÷ 336 = 19.6%, a loss of 0.4 percentage points. Both readings are correct and tell different stories: the brand is in more answers, but the answers are naming more brands (336 ÷ 190 = 1.77 brands per answer, up from 1.5), and Brand B captured the largest share of that extra room (25 of the 51 additional presences).

Reporting only share of voice would hide a real gain in reach. Reporting only visibility would hide that a competitor is advancing faster. Neither metric replaces the other, and reading the answers where Brand B got in and Brand A did not is what turns the two numbers into a working hypothesis.

Sample size and noise: what 2 out of 10 becoming 3 out of 10 can tell you

The example is hypothetical. A team runs 10 prompts on one platform in September and sees the brand in 2 answers. In October it sees the brand in 3. Visibility went from 20% to 30%: 10 percentage points more, or a relative increase of 50%. It sounds like a lot, but it is one answer.

A confidence interval shows the size of the uncertainty. Using the Wilson interval for proportions, which is suited to small samples, 2 out of 10 corresponds to a 95% interval of roughly 5.7% to 51.0%; 3 out of 10, roughly 10.8% to 60.3%. The two intervals overlap almost entirely. With these data, the true rate in September could be higher than October's without contradicting anything that was observed.

Repeat the comparison with 200 answers per period: 40 out of 200 (20%) and 60 out of 200 (30%). The intervals become roughly 15.0%–26.1% and 24.1%–36.7%: they still overlap narrowly, but uncertainty is much smaller than in the example with 10 answers. The same 10-point difference is estimated more precisely and is worth investigating with 200, although comparing the intervals alone does not prove that the rate changed. More observations reduce uncertainty; the collection design and analysis method still determine what can be concluded.

One monthly observation of 10 prompts is too sparse to attribute a trend on its own. Before interpreting the change, define the design: the same basket, cadence and comparable periods, a rule for missing or failed answers, and a method set in advance to quantify uncertainty. Then examine a recurring series. Historical variation helps diagnose noise, but it is not a binary threshold: a sustained change may be smaller than the earlier swing, and a larger spike may still be random. Without a design that isolates interventions, the sequence of a publication followed by a rise does not demonstrate a content effect.

Share of voice is even more sensitive than visibility, because numerator and denominator move together: one extra answer that names three competitors moves the denominator by three and the numerator by zero. In small samples, a brand's share of voice can swing several points because of a single answer.

The answer changes between runs, and that is not a collection failure

Part of the noise comes from something SEO never had: the same question, asked twice of the same platform, can receive different answers. Thinking Machines showed that 1,000 runs of the same prompt on the Qwen3-235B-A22B model, at temperature zero, produced 80 distinct completions, because of how servers batch requests. A 2024 study of supposedly deterministic settings found accuracy varying by up to 15% across repeated runs of the same models on the same tasks. In OpenAI's API, setting a seed gives only "(mostly) deterministic outputs," and changes OpenAI makes to its model configurations can affect that determinism.

For anyone measuring share of voice, the conclusion is that one run is a sample from a distribution, not the platform's answer. Within Google, AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary: the same question can be answered differently depending on the Google experience.

Three practices reduce the problem without eliminating it. Collect on a recurring schedule, on different days, so that a period holds several observations of each prompt rather than one. Treat each platform as its own series, because the distributions differ. And keep the original answer from every collection, because a change can only be explained by reading what changed: whether the brand vanished because the answer got shorter, because the AI switched sources, or because it started recommending something else.

How to compare periods without fooling yourself

An honest period comparison holds constant the four things that define the denominator: the prompt set, the platform set, the brand set and the valid-answer rule. If any of them changed, the comparison has to be redone on the intersection alone, or the earlier number has to be recalculated under the new rule. Adding 20 prompts to a 50-prompt panel and comparing next month with last month compares different panels.

Period lengths should also be equal, and the comparison should state how many valid answers each period contains. A period with 190 answers and another with 90, because collection failed on one platform for a week, are not equivalent, and the failures need to appear in the report as failures rather than vanish into the denominator.

When describing the change, use percentage points for the difference between two percentages and relative change for the ratio between them, and say which one you are using. In Eurostat's definition, percentage point is the term used when comparing two percentages, a different quantity from relative change: from 20% to 25% is 5 percentage points and a 25% relative increase (25 ÷ 20 − 1). Both are correct; picking the larger one to look better, or the smaller one to look stable, is the error.

Finally, compare segments before comparing the total. An aggregate share of voice can stay flat while the brand gains on one platform and loses on another, or grows on definition prompts and slips on purchase prompts. Since weightings are hard to justify, segmentation is the substitute: show the number per platform and per topic, and leave the aggregate as a summary, not a result.

How to report share of voice to leadership

Decision-makers want a number and a direction. Both can be delivered without misleading anyone, as long as the number carries its context on the same line rather than in a footnote.

A defensible report presents, for each platform and for the total, the period's visibility and share of voice, the change in percentage points against the previous period of equal length, the number of valid answers and failures, and the list of brands compared. When the brand list changed, the report says so before the number, and presents the previous period recalculated with the new list. When the method cannot distinguish change from variation, the report presents the uncertainty instead of labeling the movement as a trend.

What the number cannot promise also needs to be said. AI share of voice measures presence in a sample of answers; it does not measure recommendation, traffic or revenue. A brand can have a high share of voice because the AI names it as an example of what to avoid. The distinction between a mention, a recommendation and a cited source is a subject for another read; here it is enough to remember that the metric counts mentions.

The proof that the number is real is the original answer. A report that lets someone open the answer where the brand appeared or stopped appearing is auditable; one that shows only the percentage asks for trust. Evidence from the original answer is also what lets a leader ask the right question: "which answer did we get into, and why," instead of "why did we drop two points."

How to test a share of voice report you did not produce

Sooner or later a report arrives from an agency, a vendor or another team with a ready-made share of voice. The tests below fit into one meeting and need no access to raw data; each stands on its own and catches a specific problem.

• Ask what the denominator is. If the answer is "total answers," the number is visibility, not share of voice; if it is "total brand mentions," ask which brands.
• Ask for the brand list of the current period and of the previous one. If the lists differ and the previous period was not recalculated with the new list, the change cannot be interpreted.
• Ask how many collections failed and how a search with no AI Overview was handled. If failures and absences entered the denominator as "answer without the brand," visibility is understated.
• Ask for the number per platform before the total. If only the aggregate exists, ask how the platforms were combined; "weighted by audience" without a disclosed source and method is not auditable.
• Ask how many runs each prompt had in the period. One per prompt per month means the report measures a sample of one answer per question.
• Ask for the change in percentage points and the relative change side by side. If only one appears, check which one was left out and whether it is the less impressive one.
• Ask how many valid answers support the change, and request the recurring series, constant design, and method defined before seeing the results to quantify uncertainty. Use week-to-week variation as a noise diagnostic, not an automatic boundary between trend and no trend.
• Ask whether the metric counts one presence per answer or every occurrence of the name. Counting occurrences favors long answers and inflates whoever is repeated.
• Ask when topic or position weights were set. If it was after seeing the results, the weight serves the result.
• Ask to open one answer where the brand got in and one where it dropped out. If the report cannot, the percentage cannot be audited, and any reading of "recommendation" or "commercial result" from it is extrapolation.

Where to start

Before calculating anything, write down on one page what defines your denominator: the prompts, the platforms, the language, the valid-answer rule, the brands compared and the counting rule, one presence per brand per answer. Decide whether the brand set is closed or open and accept the cost of the choice. Define before collection how uncertainty will be quantified and what evidence from the series, design, and method supports reporting a trend; do not choose the criterion after seeing the results.

Then collect on a recurring schedule, keep every answer, and compare equal periods with the same basket. Present visibility and share of voice side by side, per platform, with the change in percentage points and the sample size on the same line. The value may rise or fall; the benefit of this process is making the denominator and the method's limits transparent, so readers can redo the calculation and assess the conclusion.

Platform documentation checked in October 2026.