Share of voice in AI answers: how to count it, and how to grow it

What share of voice measures that the named rate does not

The named rate asks: in how many answers does the business appear? Share of voice asks: of all the names the assistants gave, how many were this business? The two can move apart. A business named in half the answers, always as the fourth of five names, has a respectable named rate and a thin share. A business named in fewer answers, but first and alone when it is named, can hold more of the space.

The space is small, which is why the share matters. An assistant's answer to a buyer's question names two or three businesses and stops. There is no second page to fall to. Every name given is a place taken, and share of voice is the count of how many of those places are yours across the questions your buyers ask. The glossary's share of model is the related measure kept per assistant: the share of one assistant's answers that name you at all.

How to count it

From the same record the rest of the measurement uses: a fixed panel of ten to twelve buyer questions, each put to ChatGPT, Claude, Gemini and Perplexity three times, with every business named in every answer written down in order. Nothing extra has to be collected. Setting up prompt monitoring describes the record.

1. Count every business name in every answer. A hedge that names no one contributes nothing to the count, which is the honest treatment: it was a place nobody took.
2. Count how many of those names are yours, under any spelling of your name.
3. Divide the second by the first. That is your share of voice for the month, across the whole panel.
4. Do the same for each competitor who appears more than once. The list of names with their counts is the most useful table in the report.
5. Repeat per assistant and per question type. A share that comes entirely from the questions that name you is not a share of the buyers who have not heard of you.

An example, with invented numbers, to show the arithmetic. Twelve questions, four assistants and three runs make 144 answers. Say those answers name 380 businesses in total, and yours appears 38 times: your share is 38 in 380, one name in every ten given. Say the business named most often appears 76 times: it holds twice your space. Those two counts, side by side, tell you more than either rate alone.

Weighting: a first mention is worth more than a fifth

A plain count treats every name the same, and a buyer does not. The first name in an answer is the one read with the most attention, and a name given as a recommendation carries more than a name given in a list. So weight the count: a mention counts, a recommendation counts for more, and first place counts for most, and the same weights applied to every name in the panel give a weighted share that tracks prominence as well as presence. Our methodology keeps its own score simpler: the share of the twelve questions in which the assistant named the business, with being named first and who else was named reported beside it.

Pick the weights once and never change them between months. The exact numbers matter less than their stability, because the report is a comparison with last month, and a change of weights is a change of instrument.

How to grow it: start from the questions you are losing

The record already says where the share goes. Sort the panel's questions by how many names each produced and by whether any were yours. The questions that name three competitors and not you are the work list, and each one comes with its evidence: the names, in order, and often the sources the assistant cited.

Take the losing questions one at a time and make the smallest change that gives the assistant what the competitor's page gave it. Then wait for the next monthly run. Chasing every question at once makes it impossible to tell which change moved which count.

What moves share of voice, and what does not

What moves it is the same short list that decides being named at all. In our audit of 9,483 small business websites (September 2026), 61.5% had no question-and-answer content, 43.1% were inconsistent about who and where they are, and 64.7% showed a crawler no review signal. The audit's own probe points the same way: on inspection, the businesses Claude named were the ones whose sites carried answer-shaped content and a consistent identity. A competitor that has those and you do not holds the share you want for a reason you can see.

What does not move it: publishing more pages that answer no one's question, adding markup that describes the website, repeating keywords, or buying placements an assistant has no reason to read. The one controlled study we cite, Aggarwal et al. (KDD 2024), found keyword stuffing did not help on its benchmark, while adding citations, quotations and statistics raised a source's visibility in generated answers by up to about 40%. Specific, attributable content is the lever the evidence supports.

The limits of the number

A panel of twelve questions is small, and a share computed from it moves a few points between months for no reason at all. Read the trend over three months before believing a change, and read it beside the per-competitor counts, which tell you whether a move is yours or theirs. And the number is only as good as the panel: questions written the way you describe yourself will flatter you, and questions written the way your buyers type will not.