An AI share of voice guide: counting mentions, working out shares and weighting the questions that sell

Q1Start from coded answers

The calculation begins after the coding step: every answer already marked with the brands it names. Count one mention per brand per answer, and leave refused answers out of the count entirely, reporting them separately.

Q2The plain calculation

Illustrative figures: 20 questions, 5 runs each, 100 answers; 8 buying questions give 40 of those answers.

Each brand’s share is its mentions divided by all brand mentions. Here your brand holds 15 percent of the conversation.

Q3Weight the questions that sell

Percentages rounded to whole numbers.

Counting buying answers twice lifts your share from 15 to 19 percent, because you do comparatively well where purchases are decided. The plain figure hid that strength.

Q4Check the margin before celebrating

With 100 answers, a shift of a few points can be noise. Before reporting a change, compare it with the margin for the sample, and look for the same direction across assistants and months. How samples and margins work is explained in our report on measuring AI visibility.

Q5Report it the same way every month

• Keep the question set, the weights and the counting rules fixed.
• Show plain and weighted shares side by side.
• Split by assistant, since each behaves differently.

The service version of this method is on AI share of voice.

Q6How precise a share can be

Calculated with the standard formula for a proportion. Halving the margin needs roughly four times as many answers, which is why sample size is agreed before the first round.

Q7Share by assistant

Report shares separately for each assistant before combining them. A brand can hold a strong share in one and a weak one in another, because each draws on different sources. If you combine them into one figure, weight each assistant by how much your own audience uses it, and state the weights. An unweighted average treats every assistant as equally important to your buyers, which is rarely true.

Q8First-mention share

When an answer lists several brands, the first one often gets the most attention. First-mention share counts how often your brand is named first. It moves more slowly than overall mention share and is a useful second measure, especially for buying questions where answers list three or four options.

Q9Common calculation mistakes

• Counting refusals as zero mentions, which deflates every brand's share.
• Changing questions between rounds, so shares are no longer comparable.
• Reporting a single total without the sample size.
• Comparing shares from different assistants as if they were the same audience.
• Celebrating a change smaller than the margin.

Q10Presenting share of voice

Show each share as a range, next to the previous rounds, with the sample size beneath. Add one sentence explaining what probably drove any change larger than the margin. Keep weighted and unweighted figures side by side. Readers who see the uncertainty trust the figures more, not less. Rival comparisons are covered in our guide to AI competitor analysis.

Q11Share by question group

Made-up figures. A brand strong on trust questions but weak on discovery is known to people who already know it, and invisible to newcomers. The pattern points to where work is needed.

Q12Share among recommended brands

Some answers mention many brands but recommend one or two. Calculating your share only among recommendations, rather than all mentions, shows how often assistants actually point buyers to you. This figure is usually smaller and more volatile than mention share, but it is closer to what affects sales. Report it alongside mention share, not instead of it.

Q13When the denominator changes

If an assistant starts refusing more adult questions, there are fewer answers to count, and every brand's share can shift even though nothing changed in the market. Track the number of usable answers each round, and read shares against it. A share that rises because rivals are now refused is not the same as a share that rises because you earned it.

Q14Shares across markets

Brands selling in several countries should calculate shares separately for each market and language. Combining them hides the markets where you are weak. If a combined figure is needed for a summary, weight markets by revenue or by your audience in each, and show the separate figures underneath. Citation patterns behind the shares are analysed in citation source analysis, and monitoring between rounds in prompt set monitoring.

Q15A worked monthly report

Consider this. "In September, across 300 answers on four assistants, your mention share was 15 percent (margin about 4 points), up from 12 percent in June. Your share among recommended brands rose from 8 to 11 percent. Rival A held steady at 35 percent. Refusals fell slightly on one assistant. The rise is just outside the margin and follows the new beginner guide; we will confirm in December before drawing conclusions." A report like this is short, honest about uncertainty and useful for decisions.

Q16Share targets, set carefully

Targets for AI share of voice should be modest and tied to actions. A target of a few points over a year, on the question groups that matter most, is realistic for most brands; doubling a share in a quarter rarely is. Pair every target with the actions expected to move it, and review both each round. Rival analysis is covered in AI competitor analysis.

Q17Share of voice with a small budget

Brands with little budget can still track share of voice. Choose the ten core questions that matter most to your sales, run them three to five times each month on the two assistants your buyers use most, and code the answers with a simple shared sheet. The margins will be wide, so look only for large or lasting changes across several months. Even a small, consistent series beats occasional screenshots shared in chat, and it can grow into a fuller set later as budget allows.

Q18In short

Start from coded answers, calculate shares by assistant and question group, weight openly, show margins, keep refusals apart and report the same way every round, so changes mean something.

FAQQuestions