AI visibility share of voice: a transparent calculation method

AI visibility share of voice needs an explicit denominator. A useful definition is the share of eligible brand appearances that belongs to one brand within a documented set of AI-answer observations.

This article defines an observation-brand pair method. Each brand contributes at most one pair per eligible answer, regardless of how often its name is repeated. The resulting percentage describes that answer set and that brand set. It is not market share or a count of everything people see in an AI product.

This is a methodology reference with a fully hypothetical example. It is not an industry study and does not claim that AnswerProbe currently displays a dedicated share-of-voice percentage in its reports.

Define the observation and brand set first

A planned observation is one question-and-answer-surface check expected within a chosen measurement window. Give it an ID and preserve its question, time, market and relevant context with the returned answer.

Decide in advance how retries will be handled. Retain the attempt history, and use at most one eligible answer per planned observation under that rule. Do not count repeated attempts as additional independent observations or select whichever answer favors your brand.

Before calculating a percentage, document which results are eligible and which brands you will count. Keep those rules and the comparison set fixed within the stated version. The numeric denominator can still change when the observed answers contain more or fewer qualifying brands.

Use the buyer-question selection guide to define a relevant question set before collecting answers.

Use this eligibility rule for the method described here:

Result status	Coverage record	Contribution to brand-pair calculation
Successful and eligible	Count as a usable answer	One pair for each qualifying brand present
Unavailable	Show as a missing result	No pairs; do not invent a non-mention
Failed	Show as a failed attempt	No pairs; do not invent a non-mention
Excluded	Show separately with the exclusion reason	No pairs under this method

An eligible answer can contain no tracked brands. It still contributes to eligible-answer coverage and the denominator for mention rate, but contributes zero pairs to the share-of-voice calculation. Never exclude an otherwise eligible answer merely because your brand is absent.

Here, coverage means eligible usable answers divided by all planned observations. It is not merely the fraction of requests that returned any response. Keep unavailable, failed and excluded counts visible alongside it.

For broader guidance on observation setup, use the AI answer tracking guide.

Count each brand once per eligible answer

If an answer names your company four times, that creates one pair for your company. If the same answer also names two other eligible brands, it contributes three pairs in total.

This avoids treating repeated prose as repeated independent exposure. It also creates an important property: an answer mentioning more eligible brands contributes more pairs than an answer mentioning fewer brands.

Therefore this is a share of observed brand appearances, not an equal-weight average of each answer's recommendation strength. It does not assign extra weight for prominence, sentiment or purchase intent.

Before counting, define the canonical name and accepted aliases for each brand. Count a text match only when the surrounding answer identifies that company or product. A generic phrase such as “help desk software” is not a brand alias. Review ambiguous names against the preserved answer, and apply the same rule to every brand. A brand named only inside a citation URL is not a text mention under this method.

The calculation is:

Brand share of voice = that brand's eligible observation-brand pairs ÷ all eligible observation-brand pairs × 100.

Reproduce the hypothetical example

Everything in this table is invented for explanation. Brand A is the subject. Brands B and C are the other two members of a fixed comparison set. The table is not a claim about any real company.

Nine observations were planned. Six produced eligible successful answers. A value of 1 means a qualifying text mention was present at least once; 0 means it was absent from an eligible answer. N/A means the answer is outside the eligible set, not that the brand was absent.

Observation	Status	Brand A	Brand B	Brand C	Pairs
1	Successful	1	1	0	2
2	Successful	1	0	1	2
3	Successful	0	1	1	2
4	Successful	0	1	0	1
5	Successful	0	1	1	2
6	Successful	0	0	0	0
7	Unavailable	N/A	N/A	N/A	0
8	Failed	N/A	N/A	N/A	0
9	Excluded	N/A	N/A	N/A	0
Eligible totals	Six eligible answers	2	4	3	9

Observation 9 is excluded because the captured response belongs to the wrong question, violating the predefined question-matching rule. Its zero pair contribution is an accounting exclusion, not a negative brand finding. Observation 6 is a valid answer with no tracked brands and remains in the eligible-answer count.

The three relevant calculations are:

• Eligible-answer coverage: 6 eligible successful answers ÷ 9 planned observations = 66.7%.
• Brand A mention rate: 2 eligible answers mentioning A ÷ 6 eligible answers = 33.3%.
• Brand A share of voice: 2 A-pairs ÷ 9 total brand-pairs = 22.2%.

Brand B's pair-based share is 4 ÷ 9, or 44.4%. Brand C's is 3 ÷ 9, or 33.3%. Rounded displayed percentages may not sum to exactly 100%; the underlying fractions do.

These values answer different questions. Mention rate asks how often A appeared in a usable answer. Share of voice asks how much of the counted brand appearance total belongs to A. Coverage shows how much of the fixed planned observation set produced eligible usable answers.

Why competitor selection changes the result

Keep Brand A's two appearances unchanged, but add Brand D to the eligible set. Suppose D appears in three of the same eligible answers. Total pairs increase from nine to twelve.

Brand A's pair-based share becomes 2 ÷ 12 = 16.7%. Its mention rate remains 2 ÷ 6 = 33.3%. Nothing about A's own appearance count changed; the comparison set changed.

For a trend, either keep the eligible brand set fixed or clearly version the set and explain the effect. If you recalculate old windows using a new set, label that restatement rather than presenting it as the original measurement.

A newly discovered competitor can be recorded for later review without silently adding it to the current denominator. A publisher cited as a source, a generic product category and a competing software company are not interchangeable.

Keep citations and recommendations separate

A source URL attached to an answer is a citation. Its domain might belong to the subject brand, a competitor, a publisher or a documentation host. Counting cited domains produces a different measure from counting brand mentions in answer text.

A mention is also not automatically a recommendation. An answer may describe a company as unsuitable for the stated requirement. That negative mention still counts as one pair if the brand match is valid. Pair-based share of voice records appearance; it does not establish endorsement or positive sentiment.

If you want a recommendation-only or sentiment-weighted measure, define and validate that method separately. Do not silently add those meanings to this percentage.

Handle empty and incomplete sets honestly

If no eligible brand appears in any eligible answer, the pair denominator is zero. Report share of voice as undefined or unavailable, with the underlying counts. Dividing zero by zero does not produce a meaningful 0% share.

If eligible answers exist but none mentions Brand A, its mention rate is 0%. Its share of voice is 0% only if some other tracked brand contributes pairs, making the pair denominator positive.

If no eligible successful answers exist, mention rate is undefined. With planned observations but no eligible answers, coverage is 0%. If no observations were planned, coverage is undefined too. Report the counts and statuses rather than filling every empty metric with zero.

When coverage differs between periods, show it next to the percentages. A larger count does not automatically remove bias: missing answers may belong disproportionately to one question or surface. The percentage describes the eligible answers you have, not what the missing answers would have said.

A repeatable calculation checklist

1. Freeze the inputs. Record questions, surfaces, measurement window, intended market and language, brand-set version, alias rules, eligibility and retry policy.
2. Reconcile coverage. Check every planned observation ID. Separate eligible answers, unavailable results, failed attempts and exclusions with reasons.
3. Review the mentions. Inspect preserved answer text, resolve ambiguous aliases and create at most one pair for each brand in each eligible answer. Keep citation and recommendation labels separate.
4. Calculate from counts. Sum the pairs for each brand and across the fixed comparison set. Calculate coverage and mention rate from their own denominators. If aggregating surfaces, pool the underlying counts; do not simply average their percentages.
5. Publish the context. Show the numerator, denominator, question and surface scope, window, brand-set version, coverage and rounding beside the percentage. Include a way to inspect the evidence.
6. Check comparability. Before calling a change an improvement, inspect changed questions, brands, matching rules and missing coverage. Label changed scope or restated historical results explicitly.

Use the percentage within its documented limits

For a defensible comparison, retain question wording, answer surfaces, market, language, dates, eligibility decisions, brand-set version and collection coverage. Provider or consumer-interface differences should be disclosed where relevant.

This method can help describe which brands recur in a controlled set. It cannot reveal platform-wide impressions, consumer demand, market share or the cause of a later sale.

AnswerProbe's illustrative sample report is a static preview using synthetic fixture data, not a live scan of a company. It demonstrates the presentation of answer evidence. This article describes a calculation method and does not claim that the sample or current product provides a dedicated share-of-voice metric.

Use the underlying evidence and a clearly stated method whenever you derive another metric from observations. The denominator should remain understandable even when the headline percentage is attractive.

Frequently asked questions

Is AI visibility share of voice the same as mention rate?

No. In this method, mention rate divides answers naming your brand by eligible answers. Share of voice divides your brand's pairs by all pairs from the stated brand set. In the hypothetical example, Brand A has a 33.3% mention rate and a 22.2% share of voice.

Does repeating a brand name increase its share?

Repeating it within the same eligible answer adds no extra pair. An answer that names Brand A once and another that names it five times each contribute one A-pair. Separate eligible observations can each contribute a pair.

Do negative mentions count?

Yes, when the text clearly identifies the tracked brand. Calling a product unsuitable is still an appearance under this definition. Evaluate sentiment and recommendations separately before deciding whether that appearance is useful.

Should missing answers count as zero visibility?

No. An unavailable or failed result does not show whether a brand would have appeared. Keep it in the planned-observation coverage record, and exclude it from eligible-answer mention-rate and brand-pair calculations. A valid answer with no tracked names is different: it remains an eligible answer with zero pairs.

What does 100% share of voice mean?

It means every counted pair in this comparison set belongs to that brand. One brand appearing in one of six eligible answers, with no other tracked brands appearing, would have 100% pair-based share and only a 16.7% mention rate. It does not mean the brand appeared in every answer or dominates the market.

Can I compare this percentage with another tool's score?

Only after reconciling definitions and scope. Check the brand set, questions, surfaces, time window, treatment of missing answers, matching rules and weighting. A recommendation-weighted score or share of cited domains measures something different, even if it uses a similar label.