Every category in the October 2026 edition ended with the same kind of fifth question: an alternative to one named brand. For gas grills it read "alternatives to Weber grills", which you can see on the gas grills page. The question went to OpenAI's API with web search (gpt-6-luna) and to Perplexity's Agent API with the Sonar model (perplexity/sonar), once each per category. Of the 320 answers asked for, 316 came back: 160 from OpenAI and 156 from Perplexity, because 4 Perplexity answers in the US failed.
New runs of the index from November 2026 will not ask this question. This post explains why, using what those 316 answers contained, and what the findings mean for anyone tracking AI answers or writing comparison pages.
What did the 316 answers recommend?
They recommended far more brands than any other question. The table sets them beside the other four question types, pooled, with OpenAI and Perplexity together.
| Measure | Alternative question | The other four questions |
|---|---|---|
| Answers | 316 | 1,271 |
| Recommended brands per answer, mean | 5.0 | 2.3 |
| Median | 5 | 2 |
| Answers with no brand | 0.0% | 3.4% |
| Answers with 1 to 3 brands | 13.6% | 84.4% |
| Answers with 4 to 6 brands | 72.8% | 11.8% |
| Answers with 7 or more brands | 13.6% | 0.4% |
October 2026 edition, OpenAI and Perplexity together. Brands are those the answer recommends, not names it only mentions.
Every one of the 316 answers recommended at least one brand, and 86.4% recommended four or more. For the other questions, 84.4% of answers recommended one to three. An alternatives question asks for a list, and the answers give one.
Is the named brand on its own alternatives list?
Rarely. The brand named in the question was recommended in 21 of the 296 answers where we could match it, which is 7.1%. A 95% Wilson interval puts that between 4.7% and 10.6%, though the answers are not independent (they share categories), so the real uncertainty is wider.
"Matched" has a narrow meaning. We looked for a brand in the category's extracted brand lists whose name appears in the prompt text. In 20 of the 316 answers there was no such match, and those are left out of the 296. For the other 275 answers (296 minus 21), the data shows only that the named brand was not recommended. The methodology says it is usually mentioned but not counted as a recommendation. This data cannot split "mentioned but not recommended" from "not named at all".
One hypothesis is that a question asking for alternatives to X leads answers to leave X out of the list. The index did not test that, because the October 2026 edition has no version of the question without the brand name. What the data shows is how often the named brand was recommended, and what that means for measurement.
Why did the index drop the question?
Three reasons. The second shows in the numbers, the first is in the wording, and the third is a judgment about sample size.
- It named a brand in the prompt. In every category of the October 2026 edition the question names one brand, and that brand was recommended in only 21 of the 296 answers where we could match it. The data cannot show whether the wording causes this. Naming a brand makes it a different measurement from the other four, which name no brand.
- It inflated category counts. Each category has about two of these answers (one per engine), yet they carry twice the brands per answer. Compare the category measures with and without them:
| Measure | All five questions | Without the alternative question |
|---|---|---|
| Answers per category (mean) | 9.9 | 7.9 |
| Distinct brands recommended (mean) | 12.2 | 9.2 |
| Engine overlap (mean) | 53.6% | 47.0% |
| Brands named in only one answer (mean share) | 42.1% | 48.4% |
| Most recommended brand's share of answers (mean) | 62.7% | 62.3% |
| Recommended brands per answer, pooled over answers (mean) | 2.9 | 2.3 |
October 2026 edition, 160 categories, OpenAI and Perplexity. Engine overlap is brands recommended by both engines divided by brands recommended by either, averaged over categories. The last row is pooled over answers (1,587 and 1,271), not averaged over categories.
About two extra answers per category added 3.0 distinct brands to the average category (12.2 against 9.2) and lifted the engine overlap by 6.6 points. They barely moved the share of the most recommended brand. A reader comparing a category's brand count in October with a later count would have been comparing different question sets.
- Two answers per category cannot say much. One per engine is not enough to tell which brands are the usual alternatives to a name.
From November 2026, 40 US categories are re-sampled with a buying-advice question that names no brand in place of this one. The other 60 US categories, all 30 Mexico categories and all 30 Canada categories stay as the October 2026 edition, with the alternative question included. The two designs are not like for like, so no October-to-November comparison is made.
What should you take from this?
If you are the named brand, do not use an alternatives prompt to judge your own visibility. In this sample, 275 of 296 matched answers did not recommend the brand the question named. Track your brand with questions that name nobody, such as best overall, budget, premium and a specific use case. Read an alternatives question as a map of the brands placed beside yours.
If you are not the named brand, the question offers about twice as many slots per answer as the others (5.0 against 2.3). This data does not say which brands fill them or why, and it cannot rank anyone. It can tell you whether you appear, by asking the question yourself.
If you write comparison pages, this post has no evidence on whether they change what assistants say. The index did not test that. Treat any claim that they do as a hypothesis, and test it with a before and after measurement on enough answers (see how many AI answers you need).
If you design prompts for tracking, four rules follow from this data:
- Keep each question type separate. Never pool an alternatives prompt with category prompts.
- Name no brand in the prompts that measure your share.
- If you do track an alternatives prompt, name one competitor per prompt and report it as its own series.
- Say in every report which question a number comes from. A brand count with no question type attached is easy to misread. Share of answer covers the denominator.
How do I test alternatives to a brand myself?
- Pick one competitor and one engine. Write the question the way a person would: "alternatives to [competitor] [product type]".
- Ask it in a clean session. Save the full answer.
- Mark your brand in each answer as recommended, mentioned only, or absent. Do the same for the competitor you named.
- Repeat enough times to put a number on it. Even 296 answers gave an interval of 4.7% to 10.6% (wider once shared categories are counted), so a handful of answers shows little.
- Compare the three outcomes across engines, never together. The engines recommend different brands even for the same question; see OpenAI vs Perplexity.
What this cannot tell you
- API, not app. The apps may answer an alternatives question differently. See API answers vs app answers.
- One sample per engine. In the October 2026 edition each category has about two alternative answers, one per engine.
- No brand ranking. This post counts brands and does not name or rank them. Each category has too few alternative answers to support a ranking.
For how many brands answers recommend to other questions, see How many brands does an AI shopping answer recommend?
Frequently asked questions
Should I track an alternatives prompt at all?
Only as its own series, and only if "who sits next to my competitor" is a question you need answered. It will not tell you how visible you are, because it names a brand and lists about twice as many brands as other questions.
Does being listed as an alternative to a big brand help?
The data cannot say. It shows that these answers list many brands and rarely the named one. It has nothing on what a reader does with such a list.
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.