What AI share of voice is, and why a high one can still lose the sale

The standard definition is an advertising one. The Marketing Accountability Standards Board's dictionary describes share of voice as the relative communication presence a brand achieves, most commonly calculated as a percentage of category media spend, impressions or gross rating points. It also gives a broader reading built on mentions: “the proportion of the category discussion that the brand achieves”. AI share of voice is that broader reading, where the discussion is the answers an assistant gives a buyer. Of the parts that make up AI visibility, it reads one: whether you are named, not whether you are recommended or described accurately.

The definition leaves three decisions to whoever does the counting: what counts as a mention, which brands are in the total, and which answers the count runs over. Change any one of them and the same answers produce a different share, which is why the arithmetic under a share of voice matters more than the size of it.

Those are the three decisions the definition leaves open, settled this way.

• What counts as a mention. Being named at all: as the recommendation, as one of several names, or named and warned against. A brand named several times in one answer has been named in one answer, so repetition inside an answer does not move the figure.
• Which brands are in the total. Yours and the rivals you track. A brand the answers name that is not in your set is in nobody's figure.
• Which answers are counted. Every answer from ChatGPT, Gemini, Claude, Perplexity and Grok, and each question is put to each assistant more than once in a measurement, with every repeat counted. Your count and every rival's cover the same answers: the answers to your buying questions.

Because every brand is counted over the same answers, the figures for you and all the rivals you track add up to a hundred, give or take rounding. Each one is that brand's part of the naming on your buying questions.

Share of recommendation is a separate reading with its own denominator: the share of answers to buying questions in which the assistant tells the buyer to choose your brand. Share of voice asks whether you were in the conversation; share of recommendation asks whether you won it. The four things an answer can do to a brand are set out in mentioned is not recommended.

The first screen of a Delphi measurement puts four figures on one strip: share of recommendation, the number of answers that named you and did not choose you, the Delphi Index, and share of voice. The two shares divide by different things, so neither can be read off the other.

• The rivals in the total. The total is you plus the rivals you track, so the same answers give you a smaller share against a long set of rivals than against a short one. A share counted against a different set, or against every name in the answers, is a different quantity even on the same answers.
• A rise that is not yours. The ratio rises when rivals are named less, with nothing changing for you, and falls when a rival is named more. Read it beside your appearance rate, which counts how often you turn up against the answers rather than against the other brands, so it stays where it was.
• The verdict. Share of voice counts an answer that warns a buyer off you the same as one that recommends you, because it measures presence, and a warning is never counted as a recommendation. How you are described and whether you are recommended are separate readings, set out in mentioned is not recommended.

Two more limits apply to any figure read from AI answers, not to share of voice alone: which questions it was counted over, and how many times each was asked. Both are covered in why two AI visibility tools disagree.

• Both low. The assistants rarely name you. First find out whether they can read your pages; what to fix first sets out the order.
• Share of voice high, share of recommendation low. You are in the answer and the sale goes elsewhere. The work is on why the answer prefers the rival: what the answers say about you, what they say about the rival, and the pages they were built from. Why your competitor shows up first takes that case apart.

On your own pages it starts with whether an assistant can read them, and continues with pages that tie the brand to the use cases buyers ask about and answer the questions you are absent from. Beyond your own pages are the roundups, best-of lists and community threads cited in answers to the buying questions you lost, which a measurement records wherever an answer names its sources. Those are places to earn a presence through coverage and genuine contribution, not places you control, and Delphi never recommends undisclosed staff accounts, incentivised posts or seeded threads to get there.

Where Delphi can tell a cited page belongs to a publisher or review site, the plan names it and drafts the outreach, with marked gaps for the contact and the offer only you hold; for a community thread it names the thread and leaves the post to somebody at your company, under their own name. What it drafts and what it refuses to are set out in does AI visibility software write the content.

A media budget moves the advertising version of share of voice, and what advertising in ChatGPT can and cannot buy is set out from OpenAI's own statements.