How to Measure Share of Voice in AI Search

To measure Share of Voice in AI search, divide your brand’s qualifying appearances by the qualifying appearances of all brands in a defined comparison. First specify whether an appearance means a prose mention or a citation. Then keep the answer set, competitor roster and counting rule consistent. Without those definitions, two percentages with the same label can describe different things.

The practical task is to build a record you can recalculate, not just copy a dashboard percentage. The method below uses a transparent, presence-based convention: each brand can contribute one mention and one owned-source citation per captured answer. It is a reporting method you can reproduce, not a claim about every platform’s formula.

1. Define what your Share of Voice measures

A mention means the answer names the brand in its prose. A citation means the answer includes a source link. For this method, a brand receives citation credit only when that link belongs to a verified owned domain. A third-party article discussing a company remains a third-party source, not an owned-domain citation for that company.

Keep these observations separate. An answer can mention your brand and cite your site, mention it without citing your site, cite your site without naming the brand in prose, or do neither. None of those states alone establishes a recommendation.

Before collecting answers, write down your scope:

• The market and language.

• The exact buyer questions and whether they are branded diagnostics or unbranded discovery questions.

• The AI platform and surface, with model or mode where available.

• The collection dates and number of captures per question.

• The brands, aliases and verified owned domains included in the comparison.

• The event counted, deduplication rule and treatment of missing captures.

For a ChatGPT-only comparison, include only the eligible ChatGPT captures in both numerator and denominator. If you also collect Gemini or other surfaces, calculate their results separately before considering a combined report. A platform-specific percentage should not quietly contain observations from another platform.

2. Create an answer-level ledger

The following is a fictional teaching example, not customer data or a benchmark. Imagine four captured ChatGPT answers to four unbranded buyer questions for Sweden in English, collected in one reporting window. The roster contains Brand A, Brand B and Brand C. Each brand counts at most once per answer in each column, even if its name or domain appears repeatedly.

Captured answer

Brands named in prose

Brands with owned domains cited

Answer 1

A, B

C

Answer 2

A

A

Answer 3

B, C

B, C

Answer 4

B

None

Brand A illustrates three states: mentioned only in Answer 1, mentioned and cited in Answer 2, and neither in Answers 3 and 4. Brand C in Answer 1 illustrates the remaining state, cited without a prose mention. Counting the source link as a prose mention would erase that distinction.

In your working ledger, retain the full answer, prompt, capture date, platform and cited URLs alongside these labels. The compact table above shows the coding result; the underlying text is what lets another person check it. Resolve ambiguous brand names and domain ownership before including them, rather than guessing attribution.

3. Calculate three percentages with different denominators

The ledger contains four captured answers, six brand-answer mention observations and four brand-answer owned-citation observations. An answer naming two brands contributes two observations to the mention-share denominator, but it is still only one answer in a mention-rate denominator.

Mention rate: how often does this brand appear?

Mention rate = captured answers naming the brand ÷ eligible captured answers × 100.

Brand A appears in two of the four answers, so its mention rate is 50%. This answers a presence question about Brand A. It does not describe Brand A’s share of the mentions received by the entire roster.

Mention-based Share of Voice: how is brand presence distributed?

Mention-based Share of Voice = the brand’s answer-level mention observations ÷ all roster brands’ answer-level mention observations × 100.

Brand A contributes two of the six observations, giving it 33.3% of mention-based Share of Voice. Brand B contributes three, and Brand C contributes one. Repeating a brand name within an answer does not increase this presence-based count.

Owned-citation share: whose domains are cited?

Owned-citation share = the brand’s answer-level owned-citation observations ÷ all roster brands’ answer-level owned-citation observations × 100.

Brand A contributes one of the four owned-citation observations, giving it 25% of owned-citation share. Links to sources outside the verified brand roster are not included in this particular denominator. Label the metric accordingly: it is not a share of every source link in the answers.

Brand

Mention rate

Mention-based Share of Voice

Owned-citation share

A

50%

33.3%

25%

B

75%

50%

25%

C

25%

16.7%

50%

Brand B leads mention share in this example, while Brand C leads owned-citation share. The table does not identify which brand was recommended most, received visits or won customers. Those questions need their own evidence.

4. Reconcile the counts before reporting the result

Start by checking that the brand totals reproduce the ledger. Then address the cases that can silently change the calculation:

• Missing capture: mark it unavailable, not “brand absent.” Report the missing coverage. For a before-and-after comparison, use matched question/platform captures where possible and disclose exclusions.

• Captured answer with no roster brands: keep it in the eligible-answer denominator for mention rate. It contributes no brand observations to either share denominator.

• No qualifying observations anywhere: the relevant share is undefined because its denominator is zero. Do not report it as a measured 0% share.

• Repeated names or links: deduplicate within each brand-answer pair under this method. If you choose occurrence counting instead, document it and do not compare the result directly with presence-based counts.

• Ambiguous source ownership: preserve the link as a source observation but leave brand attribution unresolved until checked. A retailer page is not automatically a manufacturer-owned page.

• Uneven captures: pooling gives more influence to questions or platforms with more observations. State that weighting, or report separate groups rather than implying an equal comparison.

A useful reconciliation question is: “Could another analyst get the same result from these saved answers using these rules?” If not, fix the coding or scope before debating whether the percentage is good.

5. Interpret the result without inventing a benchmark

A defensible target needs a comparable baseline and a business-relevant question set. This four-answer example supplies neither a performance benchmark nor a basis for declaring any percentage universally good. An apparently strong share on questions that already name your brand answers a different question from discovery on unbranded buying prompts.

Read the counts beside the percentage over time. A share can rise because competing brands receive fewer qualifying observations, even if your own count is unchanged. Adding a brand to the roster or changing the prompt mix also changes the comparison. Preserve a stable cohort for the trend and label expanded coverage separately.

Use the result to choose an investigation. If your brand is named but its owned sources are absent, inspect which sources support the answer and whether your existing pages answer that buyer question. If an owned page is cited but the brand is not named, inspect the surrounding wording before calling the citation a recommendation. If neither occurs, check existing coverage before commissioning another article.

Record page changes and collection dates, but describe later movement as an observation, not proof that the page change caused it.

Report a result someone else can verify

For the fictional dataset, a concise report would read:

Within four captured ChatGPT answers to four unbranded buyer questions for Sweden in English, Brand A appeared in two answers. Its mention rate was 50%, its share of the roster’s six mention observations was 33.3%, and its share of the roster’s four owned-citation observations was 25%. These are presence-based counts for Brands A, B and C, not estimates of all ChatGPT usage or business outcomes.

In a real report, add the actual collection dates and link the underlying answer records. Start by coding your own captured answer set with the same ledger, then reconcile the totals before using Share of Voice to decide where to investigate next.