AI share of voice: how much of the AI’s answer is about you

In August 2026 we asked ChatGPT, Gemini and Claude three plain questions, in Spanish, about yerba mate in Argentina: which one is the best, which one is most recommended, and which single one they would pick. Each model answered each question eight times. That is 72 answers.

Playadito was named in all 72. By any measure of presence, it was everywhere. But the three models also named 45 other brands, and all of those mentions together came to 506. Playadito’s 72 are 14% of 506. Being in every answer and owning the conversation are two different things, and share of voice is the number that measures the second one.

1. 01

   Only questions that do not name you

   A question like “what do people say about Brand X?” will mention Brand X by design. Share of voice is computed only over organic questions, the ones that name neither your brand nor a declared rival. We check the question’s category and its actual text, because a question filed under a discovery category can still name the brand.

2. 02

   One mention per brand per answer

   Each answer adds one to your count if it names you, and one to each other counted brand it names. A brand named three times in the same answer still counts once. An answer that never arrived (an error, an empty reply, a search engine that produced no AI summary) names nobody, so it adds nothing to either side.

3. 03

   Count against a declared set

   Since 18 September 2026, the denominator is your brand plus the competitors you declared. AI models now name four to twelve entities per answer, including regulators, marketplaces and platforms. With all of them in the denominator, the number measured how talkative the model was, not how you compare with your rivals. Scans completed before that date keep the rule they were reported with, so no delivered report changes.

4. 04

   Drop names that are not a choice

   Retailers, marketplaces and generic terms that are not one of your declared competitors are left out of the count. A declared competitor is always counted, even if its name looks generic.

5. 05

   Report the sample and the band

   Every share is shown as k of n, with a Wilson 95% interval around it. Two brands whose bands overlap are reported as tied, not ranked.

6. 06

   Split by surface

   Chat models (ChatGPT, Gemini, Claude, Meta’s Muse) answer from what the model knows. AI search engines (Google AI Overviews) answer from live web results, often paraphrasing your own pages. Refractia keeps the two apart, because a brand can be invisible in one and dominant in the other.

Data: Refractia’s public Índice de Marcas en la IA, edition of August 2026 (generated 10 August 2026), category “Yerba mate”. Three questions, three chat models with web search off, eight runs each: 72 answers, every one of which named at least one brand. The questions never name a brand, so every answer is organic.

Below, the same five brands under two denominators. The open set counts every name the models said (506 mentions across 46 names). The declared set imagines a brand that declared the other four as its competitors (293 mentions across those five). The mention counts are real; only the choice of set is an illustration.

The same brand reads 14% or 25% depending only on who sits in the denominator. That is why Refractia always says which set a share is measured against, and why it measures against the rivals you declared.

Engine by engine

Playadito was named in all 24 answers of each model. Its share of voice still differs by model, because the models name very different numbers of brands:

A screenshot of ChatGPT naming a competitor is not a benchmark. It is one answer, from one model, on one day, asked one way. The yerba mate numbers above hold because every brand was measured under the same conditions:

1. 01

   The same questions for every brand

   One set of category questions, and every brand is counted on that set. If each brand gets its own questions, you are comparing questions, not brands.

2. 02

   The same country and language

   Answers change from one market to another. Write the questions in the market’s language and name the country in full (“in Argentina”, not “AR”).

3. 03

   The same period

   Every brand in the same run. Comparing this month’s measurement of you with last quarter’s of a competitor mixes the brand effect with model updates and a changed web.

4. 04

   The same models and the same number of runs

   Answers vary from one run to the next, which is why the index asks each question eight times per model. Five models for you and two for them, or three runs against one, is not a comparison.

5. 05

   The full set of names

   Count every brand the AI names, not only the rivals you declared. In the yerba mate example the models named 45 other brands; the declared set covered 293 of the 506 mentions.

6. 06

   By question and by model, not just in total

   A category holds several conversations. In the bank category of the same index, the most named bank overall is not the most named one when the question is about personal loans; and across the 62 categories, the models disagree sharply about at least one brand in 46.

• Whether you show up at all. A share is relative. You can hold 14% while appearing in every answer, or hold 50% while appearing in only a few. Read it next to visibility, which counts answers, not mentions.
• Whether you are named first. Playadito was the first brand named in 30 of 72 answers. Share of voice treats first and last place the same.
• Whether the mention is good. A mention inside a warning counts the same as a recommendation. Sentiment and description are separate measures.
• Your market share. Models answer from their training data or from search results, which can be out of date. A brand that changed hands or closed can still be named; a new one can be missing.
• Anything outside the questions asked. The number belongs to its question set, country, language and date. Change the questions and the number changes.
• Movement caused by the model, not by you. If a model starts naming more brands per answer, your share falls even if nothing about your brand changed. The Gemini row above is that effect in one table.
• Precision it does not have. The band is a binomial approximation of a share across many brands. It is honest enough to say “between 11.5% and 17.5%”; it is not a guarantee.

The open set also contains names that are not yerba brands at all, such as a food fair the models named once. With a declared set, those never reach the denominator.