What is AI shopping visibility? A definition and how to measure it
AI shopping visibility is how often assistants recommend a brand when asked what to buy. A precise definition, the measurement method and October 2026 numbers.
When a shopper asks an assistant "what is the best carry-on suitcase under $150?", the reply is a short piece of writing that names a few products. There is no page two. A brand is either in that piece of writing or it is not. AI shopping visibility is the name for how often a brand is in it.
What does AI shopping visibility mean?
The definition has four parts, and a number that leaves one out is hard to interpret.
- The unit is an answer. One question sent to one assistant, with one reply. Not a search, not a session, not a shopper.
- The event is a recommendation. The reply presents the brand as a pick. A brand named only as a comparison or a warning is a mention, and the index does not count it.
- The denominator is a defined set of answers. Which assistant, which market, which questions, which month. Visibility is
answers that recommend the brand ÷ answers sampled, so a bare percentage means nothing without "of how many, and where". - A brand counts once per answer, however many times the reply repeats the name.
Two companion measures travel with it. Position is where in the reply the brand comes up (first, second, later). Sources are the web pages the assistant linked or retrieved while answering. Share, position and sources answer different questions: how often, how prominently, and from what evidence.
What it is not: traffic, sales, sentiment, product quality or search ranking. A brand can be recommended often and sell little, or the reverse. The methodology states the same limits for the Dopbot Index.
How is it different from a search ranking?
A search ranking is a position for one query on one results page. AI shopping visibility is a rate across many generated answers. Four differences matter in practice.
| Search ranking | AI shopping visibility | |
|---|---|---|
| Unit | A page's position for a query | A brand recommended inside a written answer |
| Length of the list | A results page with many links | A short list: 2.3 brands per answer in October 2026 (excluding the alternative-to-a-brand question, 1,271 answers) |
| Repeatability | A position you can look up | A rate you estimate, because the same question can return different answers |
| Where it is measured | One search engine at a time | One assistant at a time, and they overlap only partly |
The short list is the point. In the October 2026 sample, 84.4% of answers recommended one to three brands, and 3.4% recommended none (excluding the alternative-to-a-brand question, 1,271 answers across both engines). A brand outside those two or three slots gets no partial credit.
The assistants also differ from each other. In the same sample, the brands recommended by OpenAI's API with web search (gpt-6-luna) and by Perplexity's Agent API with the Sonar model (perplexity/sonar) overlapped by 47.0% on average, measured per category as brands both recommended divided by brands either recommended, then averaged over 160 categories with equal weight. It is not answer-by-answer agreement. Brands that one assistant recommends are often absent from the other's answers in the same category, so visibility is reported per assistant. From here, "OpenAI" and "Perplexity" mean those two API engines.
How is AI shopping visibility measured?
A measurement method, with the Dopbot Index's choices noted:
- Write the questions. Five per category, one for each intent: best overall, a budget with a price ceiling, premium, a specific use case, and a fifth question (October 2026: an alternative to a named brand; from November 2026 in the 40 updated US categories: a buying-advice question that names no brand).
- Ask each assistant, with web search on, through its developer API, with no account, history or memory attached.
- Read each reply. A model lists the brands the reply names and marks each as recommended or only mentioned. Retailers and marketplaces are not brands. Names are normalized and a hand-kept merge list joins variants of one brand.
- Count. For each brand, the number of answers that recommend it out of the answers sampled, reported as "k of n".
- Attach an interval when you report a rate. With small n, a 95% Wilson interval shows the range of true rates the count is consistent with. The index's category pages show the count as "k of n" and use the interval for badge eligibility. The glossary defines the terms used here.
- Record the sources per assistant, and repeat on a schedule.
Question type changes the result. Across all five question types in October 2026, both engines together, the mean number of brands recommended per answer was:
The alternative question names a brand in the prompt and returns about twice as many brands, which is why the index excludes it from brand-count measures. Of the 316 answers to it, we could find the named brand in 296, and it was recommended in 21 of those. The other four questions return 2.2 to 2.6 brands. "Best overall" answers recommended no brand in 9.1% of the 318 answers, more than any other type (specific use case was next at 4.1%). The index does not record why.
What did the October 2026 sample show?
The edition collected 1,587 answers (1,600 attempted; 13 Perplexity answers failed) in 160 categories: 800 from OpenAI and 787 from Perplexity, in the United States (100 categories), Mexico (30, in Spanish) and Canada (30). Three patterns are relevant to a definition.
Slots are scarce and differ by category. Excluding the alternative question, the median category had 8 distinct brands recommended across its answers, with a range from 3 to 21 (160 categories, about 8 answers each). In the October 2026 sample, US air purifiers produced 4 distinct brands in 8 answers and US mattresses 12 in 8 answers. These are two examples, not a ranking of categories: eight answers is a small sample.
Budget and premium shortlists differ. Per category, with both engines together, the brands in the budget and premium answers overlapped by 8.2% on average (Jaccard, usually two budget and two premium answers per category, 155 of 160 categories). That describes the shortlists observed for those questions, not an effect of wording and not what shoppers prefer. The budget vs premium post covers it.
The evidence behind the answers is recorded differently. OpenAI's answers link 2.2 pages on average. Perplexity's search retrieved 27.3 source URLs per answer on average, and its API does not say which of them the reply used. Those are not comparable counts, which is why this post does not pool them. The October 2026 pillar post shows both side by side.
What does a visibility number not tell you?
- It is the API, not the app. The index samples developer APIs because that allows fixed questions on a schedule. As of October 2026, Perplexity's documentation says its API uses the same search system as its interface, with differences in configuration, so outputs may differ. API answers vs app answers covers the gap.
- It is a small sample per category. About 10 answers per category in October 2026 (about 8 excluding the alternative question). One answer moves a brand's share by about 10 points (12.5 points on the 8-answer base).
- It does not prove a source mattered. A domain in a source list shows that the page was linked or retrieved, not that it shaped the recommendation.
- Search may differ in the apps. OpenAI's documentation says that, as with other tools, the model can choose whether to search based on the prompt. In October 2026 the index asked OpenAI's API to search on every request, so an app answer may rest on different evidence than a sampled one.
- It does not cover Google. The index does not measure Google's AI Overviews or AI Mode, and Google's documentation says, as of October 2026, that those two may use different models and techniques and show different links.
- Prompts are ours. We wrote them. They approximate how shoppers ask, and no shopper was observed.
Why should a brand track it?
Because the answer is short, a brand can be missing from the shortlists that matter to it even when its search rankings look unchanged; the two are measured separately. Tracking tells you which assistants and which question types include you, which sources they link or retrieve, and which competitors share the slots. It does not promise that any change will make you appear. Start with the free manual method, or use the free snapshot if you prefer.
For definitions of the other terms used here, see the glossary.
Frequently asked questions
Is AI shopping visibility the same as being cited?
No. Citations are links that an assistant attaches to a reply, and they are recorded differently by each assistant. Visibility, as defined here, counts whether the brand is recommended in the text. A brand can be recommended with no link to its site, and a site can be linked without its brand being recommended.
Why is visibility measured per assistant?
In the October 2026 sample the two engines' brand lists for a category overlapped by 47.0% on average (Jaccard, excluding the alternative-to-a-brand question, about 4 answers per engine per category, so sampling noise is part of the gap). The overlap is partial, so each assistant gets its own count.
How we write: drafted with AI tools from Dopbot Index data and public sources; every number and outside fact is traced to its source in a claims record and checked before publishing. More about how we write; corrections to hello@dopbot.com.