The answers in the October 2026 edition are not short. The average answer from OpenAI's API with web search (gpt-6-luna) ran 1,165 characters, and the average from Perplexity's Agent API with the Sonar model (perplexity/sonar) ran 1,286 (each engine's full set of answers, every question type: 800 for OpenAI and 787 for Perplexity). Yet most of them recommend only a couple of brands. The scarce resource in an AI shopping answer is not space. It is the shortlist, and this post measures how long that shortlist is.
What counts as a recommended brand?
An AI model reads each answer and lists the brands it names, then marks whether the answer presents each one as a recommendation or only mentions it, for example as a comparison or a warning. Retailers and marketplaces are not brands and are left out. A brand counts once per answer, and at most 15 brands are kept per answer. The rules are on the methodology page.
So "brands per answer" below means recommended brands, not every name that appears. In the October 2026 edition each category had five questions per engine. The headline figures leave out the alternative-to-a-brand question, the edition's fifth question, which names a brand in the prompt (316 answers; it gets its own section). That leaves 1,271 of the 1,587 collected answers: OpenAI 640 and Perplexity 631.
How long is the shortlist?
In the October 2026 edition, excluding the alternative question, an answer recommended 2.3 brands on average and the median was 2. Most answers recommended one to three brands.
Two readings follow. A median of 2 means at least half of the answers recommended two brands or fewer. And answers with seven or more brands are so rare that the 15-brand cap almost never applies.
Does the engine change the count?
Slightly. Perplexity's answers recommended slightly more brands on average (0.3 more), and OpenAI's more often recommended none.
| Engine | Answers | Mean | Median | None | 1 to 3 | 4 to 6 | 7 or more |
|---|---|---|---|---|---|---|---|
| OpenAI | 640 | 2.2 | 2 | 4.8% | 85.5% | 9.1% | 0.6% |
| Perplexity | 631 | 2.5 | 2 | 1.9% | 83.4% | 14.6% | 0.2% |
Recommended brands per answer, October 2026 edition, excluding the alternative-to-a-brand question.
The difference in means is 0.3 brands. The clearest gap is in the middle band: 14.6% of Perplexity answers recommended four to six brands, against 9.1% of OpenAI's.
Does the question change the count?
By the mean, about as much as the engine does; by the share of answers with no brand or with one to three brands, more. The table covers all five question types, with the alternative question shown as a labelled contrast.
| Question type | Answers | None | 1 to 3 | 4 to 6 | 7 or more |
|---|---|---|---|---|---|
| Best overall | 318 | 9.1% | 72.3% | 17.9% | 0.6% |
| Budget (price ceiling) | 318 | 0.3% | 93.4% | 6.3% | 0.0% |
| Premium | 317 | 0.0% | 84.9% | 14.5% | 0.6% |
| Specific use case | 318 | 4.1% | 87.1% | 8.5% | 0.3% |
| Alternative to a named brand | 316 | 0.0% | 13.6% | 72.8% | 13.6% |
Share of answers by number of recommended brands, OpenAI and Perplexity together, October 2026 edition.
The median was 2 for each of the first four types and 5 for the alternative question. Three details stand out.
- Budget answers are the most uniform. 93.4% recommended one to three brands, and 0.3% recommended none.
- The broad question is the one most likely to recommend no brand. 9.1% of best-overall answers recommended none, and answers with no brand at all are a topic of their own.
- Premium wording is specific to the October 2026 edition. Most premium prompts were phrased with a price floor. From November 2026 the 40 US categories that are still updated use a wording with no price floor, so these numbers describe the October wording.
The alternative question lists about twice as many brands as any other, and 13.6% of its answers listed seven or more. That is why it is kept out of the headline figures and reported on its own.
What does a two-brand shortlist mean for a category?
Across the 160 categories in the October 2026 edition, answers excluding the alternative question recommended 9.2 distinct brands per category on average, in about 7.9 answers each. Each answer therefore names about a quarter of its category's brands: roughly 2.3 of 9.2.
There is an identity behind this. Add up every brand's share of answers inside one category and the total equals the average number of brands per answer. At 2.3, all the shares in a typical category add up to about 230%, so the shares are small slices of a short list. A brand named in 3 of a category's 8 answers (37.5%) already holds more than the roughly 25% that an equal split would give. The category figures are examples of the arithmetic. With about 8 answers each, they are too coarse to rank brands.
By market, the means are close, and the no-brand share is where they part:
| Market | Answers | Mean | None |
|---|---|---|---|
| United States | 797 | 2.4 | 1.8% |
| Mexico | 237 | 2.2 | 10.5% |
| Canada | 237 | 2.3 | 1.7% |
Recommended brands per answer, OpenAI and Perplexity together, October 2026 edition, excluding the alternative-to-a-brand question.
Mexico's questions are in Spanish and its categories and prompts differ from the other two markets, so the table is descriptive. It does not say why Mexico's no-brand share is higher.
How do I count the shortlist in my own category?
- Write four questions: best overall, a budget with a price ceiling, premium, and a specific use case. Leave brand names out of them.
- Ask each question several times in each engine you care about (at least five times each, ideally ten), in a fresh session each time, and save the full answers.
- For each answer, list the brands it recommends. A brand it only compares or warns against does not count. An answer with no brand scores 0 and stays in the count.
- Work out the mean, the median and the share of answers with no brand for each engine. Split by question type only when each type has at least ten answers.
- Repeat on another day. One pass is a draw, and the uncertainty shrinks only as answers accumulate; how many AI answers you need works through intervals for a share of answers.
A mean near 2.3 across many answers is what the index saw across its 1,271 answers; a single category's figure from a handful of answers can sit well away from it. If you see 5 or more, check whether a prompt named a brand.
What this cannot tell you
- It is the API, not the app. The apps may use different model versions, tools and personalization, and may list more or fewer brands. See API answers vs app answers.
- It counts brands, not products. An answer that recommends two products from one brand counts one brand.
- It rests on an extraction step. The model that lists brands can miss one or split a name. Across all 1,587 answers, 333 brand mentions were marked as mentions only and are not counted.
- It is one sample. In the October 2026 edition each category had about 8 answers without the alternative question, 4 per engine, and 13 Perplexity answers failed and are in no denominator.
- It does not tell you what any shopper saw. The answers are to prompts the index wrote.
For the rest of the edition's numbers, see What 1,587 AI shopping answers show. The engine overlap behind these shortlists is in OpenAI vs Perplexity.
Frequently asked questions
Why is the median 2 when the mean is 2.3?
A minority of answers name four or more brands, and they pull the mean above the middle value. In the October 2026 edition, excluding the alternative question, 12.2% of 1,271 OpenAI and Perplexity API answers (11.8% for four to six brands and 0.4% for seven or more) named four or more.
Does an answer with more brands help a brand more?
The data cannot say. It counts how many brands an answer names, not what a reader does next. A longer list has more slots for brands. Whether a slot is worth more or less when the list is longer is not something this data can show.
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.