AI brand visibility: how often the AI names you
One yerba mate brand was named by Claude in all 24 of its answers, by Gemini in 2, and by ChatGPT in none. Same questions, same day. Visibility is the number that catches this, as long as you read it model by model and next to its sample.
Short answer
AI brand visibility is the share of AI answers to your category questions that name your brand, counted only over the answers the AI actually gave. If ChatGPT, Gemini and Claude answered 72 times and 61 of those answers named you, your visibility is 61 of 72, or 85%. It tells you how often you are in the answer, not how much of the answer is yours: that is AI share of voice.
The same brand, 0 of 24 and 24 of 24
In August 2026 we asked three AI models, in Spanish, which yerba mate is best in Argentina, which is most recommended, and which single one they would choose. Each model answered each question eight times.
CBSé appeared in every one of Claude’s 24 answers. Gemini named it in 2 of 24. ChatGPT never named it. Blended, CBSé reads 26 of 72, or 36%, a number that describes none of the three models. Visibility is simple to compute, and easy to misread when the models disagree this much.
How Refractia computes it
Formula
Visibility = answers that name you ÷ answers the engine actually gave, per engine, reported as “k of n” with a 95% error band.
- 01
A fixed set of questions
Each scan asks the same category questions in the brand’s country and language, so two scans, or two brands, are measured on the same ground.
- 02
Engine by engine
The count is kept separately for each engine. A question asked several times to the same engine counts as naming you when you appear in more than half of those runs. One mention in three tries is not counted as a reliable appearance.
- 03
Non-answers leave the denominator
An error, an empty reply, or a search engine that produced no AI summary is not an answer that left you out. It is a missing measurement, so it is removed from the count instead of being scored as a zero. An engine that did not answer at all disappears from the table rather than showing 0%.
- 04
Sample and band, always
Every percentage is shown with its k of n and a Wilson 95% interval. An engine asked once per question carries a wider band than one asked several times, and the band says so.
- 05
Chat and search kept apart
Chat models (ChatGPT, Gemini, Claude, Meta’s Muse) answer from what the model has learned. AI search engines (Google AI Overviews) build the answer from live web results and often repeat your own pages. A mention means something different on each surface, so Refractia never blends them into one story.
Worked example: yerba mate in Argentina
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 (ChatGPT, Gemini, Claude) with web search off, eight runs per question per model: 72 answers. In the index each answer is its own observation, so there is no more-than-half rule here.
| Brand | All models | ChatGPT | Gemini | Claude | Named first |
|---|---|---|---|---|---|
| Playadito | 100% (72 of 72) · band 94.9–100% | 24 of 24 | 24 of 24 | 24 of 24 | 30 of 72 |
| Rosamonte | 85% (61 of 72) · band 74.7–91.2% | 19 of 24 | 18 of 24 | 24 of 24 | 1 of 72 |
| La Merced | 83% (60 of 72) · band 73.1–90.2% | 18 of 24 | 23 of 24 | 19 of 24 | 4 of 72 |
| Taragüí | 76% (55 of 72) · band 65.4–84.7% | 16 of 24 | 15 of 24 | 24 of 24 | 1 of 72 |
| CBSé | 36% (26 of 72) · band 26.0–47.6% | 0 of 24 | 2 of 24 | 24 of 24 | 0 of 72 |
Source: edition-v2-2026-08, category yerba_mate. Bands are Wilson 95% intervals. Per-model samples are a third of the total, so their bands are roughly √3 wider.
Two readings. Rosamonte and La Merced, at 85% and 83%, sit inside each other’s band: they are tied, not ranked. And Rosamonte is named in 61 answers but first in only one, while Playadito is first in 30. Visibility counts presence; it does not count rank.
What visibility does not tell you
- Whether you are the answer or a footnote. Being named fifth counts the same as being named first. Look at position (Rosamonte: 61 of 72 present, 1 of 72 first).
- How much of the answer is yours. Playadito is in 100% of answers and holds 14% of all brand mentions. That second number is share of voice.
- Whether the mention helps you. A warning and a recommendation both count as a mention. Sentiment and description are measured separately.
- What any single model thinks, if you only read the blend. CBSé’s 36% averages 0, 2 and 24 out of 24. Always open the per-engine view.
- Your market share or your sales. Chat models answer from training data that can be months old. A brand can be well known and rarely named, or the reverse.
- Anything about questions we did not ask. Visibility belongs to its question set, country, language and date.
- Appearances below the threshold. In a client scan, a brand named in one of three runs counts as not appearing for that question. That is deliberate, and it means a borderline brand can read lower than its raw mention count.
Frequently asked questions
What is AI brand visibility in one sentence?
It is the share of AI answers to your category questions that name your brand, counted only over answers the AI actually gave, reported as k of n.
Is visibility the same as share of voice?
No. Visibility counts answers that name you (Playadito: 72 of 72). Share of voice counts what part of all brand mentions is yours (Playadito: 72 of 506). Read them together.
Why are failed or empty answers left out?
Because a missing answer is not evidence that the AI ignored you. Counting it as a zero would make an engine outage look like the AI losing interest in your brand.
Why do you report each AI model separately?
Models disagree, sometimes completely. In our yerba mate data one brand went from 0 of 24 answers in ChatGPT to 24 of 24 in Claude. A blended number hides that.
Why does every percentage come with “k of n”?
Because 100% of 24 answers and 100% of 1,000 answers are claims of very different strength. The sample and the error band tell you how far to trust the number.
Does a mention in Google AI Overviews mean the same as one in ChatGPT?
No. AI search engines build their answer from live web results, often from your own pages. Chat models answer from what they learned. Refractia reports the two surfaces separately.
Keep reading
How often does the AI name your brand?
We ask your category questions to the AI models your customers use and show where you appear, engine by engine, with every number next to its sample. Free, no commitment.