In most categories of the October 2026 edition two things are true at once. A small group of brands comes back across answers, and a longer list of brands appears once. This post puts numbers on both, then checks how much of the pattern small samples would produce on their own. Brands are never ranked here, and no brand is named as the top of anything: with about 8 answers per category, the data cannot support it.
What is being counted?
All figures cover OpenAI's API with web search (gpt-6-luna) and Perplexity's Agent API with the Sonar model (perplexity/sonar), pooled within each category, in the October 2026 edition and excluding the alternative-to-a-brand question. That leaves four questions per engine in each of 160 categories (US 100, Mexico 30, Canada 30). A category has about 7.9 answers: 8 in 151 categories and 7 in the 9 where a Perplexity answer failed.
Three measures describe a category:
- Distinct brands: brands recommended in at least one of the category's answers, either engine.
- Most recommended brand's share: answers that recommend the most recommended brand (or brands tied for it) divided by the category's answers.
- Named once: brands recommended in exactly one answer, as a share of the category's distinct brands.
Only recommendations count. A brand an answer merely mentions does not, and a brand counts once per answer (methodology).
How many brands share a category?
On average 9.2, with a median of 8. The range is wide, from 3 brands to 21.
Gas grills sit at the narrow end. In 8 answers the engines recommended only 3 distinct brands, and at least one brand appeared in all 8. Resistance bands sit near the wide end: 8 answers, 18 distinct brands, the most recommended in 3 answers, and 12 brands named once. Both are examples, not findings. Their category pages (gas grills, resistance bands) are still the October 2026 edition, but they count all 10 answers including the alternative-to-a-brand question, so their totals are higher than the counts here.
How much of a category goes to its most recommended brand?
The mean share was 62.3% and the median 62.5%. In 80.0% of categories the most recommended brand was in at least half of the answers.
| Most recommended brand's share of the category's answers | Categories (of 160) |
|---|---|
| Under 50% | 32 |
| 50% to 74% | 59 |
| 75% to 99% | 58 |
| 100% | 11 |
October 2026 edition, OpenAI and Perplexity together, excluding the alternative-to-a-brand question. With about 8 answers each, one answer moves a category's share by 12.5 points.
In 11 categories at least one brand was recommended in every answer sampled. Those categories averaged 6.2 distinct brands, below the overall 9.2. A brand named in every answer of a small sample is not an outcome the data can promise to repeat. In each category, the brand named first in an answer was one of the category's most recommended brands in 49.2% of the answers that recommend at least one brand, averaged over the 160 categories.
How long is the tail?
In the average category, 48.4% of the brands were named in only one answer. The median was 50.0%. In 88 of the 160 categories at least half of the brands appeared once, and in 7 none did. In plain counts, a typical category had about 4.3 brands that recurred in two or more answers and 4.9 that appeared once.
The tail is partly a sampling effect. A brand named once could have been named in a second, third or fourth answer with more sampling, and brands that were never named are not in the count at all. So the distinct-brand count rises with the number of answers asked. Adding the alternative question's two answers per category lifted the average from 9.2 to 12.2 (with 9.9 answers per category instead of 7.9). That jump comes from a question that asks for lists, so it overstates the effect, but the direction holds.
Is this concentration, or just small samples?
Take the most recommended brand's share first. The maximum of a few small counts is high even when no brand is preferred. To see how high, the index data was run through a deliberately simple model: in each category, keep its actual number of answers and its actual list of distinct brands from the counted questions, and let every answer pick two or three brands (averaging 2.3, the real mean for these questions) at random, each brand equally likely. Repeated many times over the 160 categories, the most recommended brand's share averaged about 54.3%.
The real figure is 62.3%. The model's 54.3% is a high baseline: its categories end with 7.8 distinct brands on average, fewer than the real 9.2, because only part of each category's brand list gets drawn. Giving each model category the longer brand list from all five questions (9.4 distinct brands in the model) lowers the baseline to about 47.2%. So roughly 47% to 54% of the 62.3% is what 8 answers produce on their own, and the remaining 8 to 15 points could reflect real preference; with this sample that is a weak signal either way.
The tail runs the other way. Under the equal-chance model, about 26.6% of the brands were named once, against 48.4% in the real answers. With the longer lists the model names about 38.5% of brands once; the real 48.4% is still higher. Real answers have more one-off brands than equal chance would give, and also a recurring head. The brand distribution is more uneven than random picks from a fixed list, at both ends. The model is a yardstick, not a description of how any assistant works. Its brand lists come from the real answers, which already carry whatever preferences exist.
What separates tight categories from crowded ones?
The groups below are defined by the number of distinct brands, so part of the pattern is arithmetic: a category's 8 answers hold about 18 brand slots (8 x 2.3), and fewer distinct brands means those slots repeat more. Categories with few brands have a high top share and few one-offs; crowded ones show the reverse.
| Distinct brands in the category | Categories | Most recommended brand's share (mean) | Brands named once (mean share) |
|---|---|---|---|
| Up to 6 | 46 | 77.6% | 31.6% |
| 7 to 8 | 36 | 63.3% | 49.9% |
| 9 to 11 | 40 | 56.6% | 54.1% |
| 12 or more | 38 | 48.6% | 61.2% |
October 2026 edition, OpenAI and Perplexity together, excluding the alternative-to-a-brand question, 160 categories.
By market the picture is similar, with small differences:
| Market | Categories | Distinct brands (mean) | Most recommended brand's share (mean) | Brands named once (mean share) |
|---|---|---|---|---|
| United States | 100 | 9.2 | 64.3% | 47.5% |
| Mexico | 30 | 9.1 | 55.7% | 50.4% |
| Canada | 30 | 9.3 | 62.0% | 49.2% |
Same edition and exclusion. The markets have different categories and prompts, and Mexico's are in Spanish, so this is descriptive.
What should a brand do with this?
- Do not read a category share as market share. A brand in 3 of a category's 8 answers is in 37.5% of those answers, not 37.5% of anything sold.
- Treat one appearance as weak evidence. Nearly half a category's brands appear once in a single run. Being named once does not show a pattern, and being absent from one run does not show absence. Repeat the questions before concluding either way. How many AI answers you need covers the intervals, and share of answer the definition.
- Count your category's distinct brands first. Where a few brands filled every slot in your run, a brand outside them was not named in that run; where many brands share the slots, the top share is lower and one-offs are common, so a single appearance says little.
- Measure per engine. In each category the two engines' brand lists overlapped by 47.0% on average (brands both recommended divided by brands either recommended, excluding the alternative-to-a-brand question). See OpenAI vs Perplexity.
To profile your own category:
- Ask four questions without brand names (best overall, budget, premium, specific use case) in each engine, several times.
- List the brands each answer recommends.
- Count the brands in at least two answers, the brands in one answer, and the share of answers holding the most recommended brand.
- Note the number of answers next to each figure. A share from 8 answers moves by 12.5 points with one answer.
What this cannot tell you
- It is the API, not the app. See API answers vs app answers.
- It is one run. The edition asked each question once per engine. 13 Perplexity answers failed in the edition, 9 of them among the questions counted here, and none is in any count.
- Concentration is not shown per engine here. The measures pool OpenAI and Perplexity answers within a category.
- It does not explain why. The data does not measure why some brands recur.
The edition's other headline numbers are in What 1,587 AI shopping answers show, and the shortlist length per answer in How many brands does an AI shopping answer recommend?. The alternative question's effect on counts is in Asking AI for alternatives to a brand.
Frequently asked questions
Does a high top share mean a category is locked up?
Not from this data. In a simple random-pick model, a category with about 8 answers still gives its most recommended brand a share of roughly 47% to 54% on average, against 62.3% observed, and the data shows nothing about how stable a category's brands are over time.
Does a long tail mean a new brand has a chance?
It shows that brands outside the recurring few do get recommended. It does not show how, or how reliably: a brand named once in 8 answers may not be named again on the next run.
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.