When AI shopping answers recommend no brand at all

A share of answers is a fraction, and its denominator includes answers that recommend nobody. In the October 2026 edition, excluding the alternative-to-a-brand question, 43 of 1,271 answers recommended no brand. The answers came from OpenAI's API with web search (gpt-6-luna) and Perplexity's Agent API with the Sonar model (perplexity/sonar). This post counts those answers, shows where they cluster, says what the data cannot explain, and covers what a brand should do about them.

What counts as "no brand"?

An AI model reads each answer and lists the brands it recommends. Retailers and marketplaces are left out because they are stores, not brands, and a brand an answer only mentions (as a comparison or a warning) is not a recommendation (methodology). An answer with no recommended brand scores zero and stays in the denominator of every share.

So a "no brand" answer could be several things: advice on how to choose, an answer that names only stores, an answer that names brands without recommending any, or an answer where the extraction found nothing. The data records the count of recommended brands, not the reason, so it cannot split these. Each of the 43 is simply an answer with a zero.

This post leaves out the alternative question because it names a brand in the prompt and none of its 316 answers had zero brands.

How often does it happen?

Overall 3.4%: 43 of 1,271 answers (95% Wilson interval 2.5% to 4.5%). By engine:

October 2026 edition, excluding the alternative-to-a-brand question. Intervals treat answers as independent, which understates the uncertainty.

A Fisher exact test on these counts gives p of about 0.005 for the gap between the engines. That is indicative only, because answers within a category and prompt are not independent.

Where does it cluster?

By market, Mexico stands apart:

October 2026 edition, excluding the alternative-to-a-brand question.

Outside Mexico the rate was 1.7% (18 of 1,034 answers). The Mexico sample has 30 categories asked in Spanish, with its own categories and prompts, and the engines and counting rules are the same as elsewhere. The data cannot separate language, category mix, prompt wording and how well brands were extracted from Spanish text, so none of them can be named as the cause. The index pages for Mexico are in Spanish.

By question type, the broad question stands out:

October 2026 edition, OpenAI and Perplexity together. The alternative question had 0 of 316.

Counts are reconstructed from the shares in the data pack and the answer counts, and they add up to 43 under each of the three splits. The splits overlap (a Mexican answer is also from one engine and one question type), so they cannot be added, and the data has no cross-tabulation of them. Compare each split's share of the 43 with its share of all answers:

Why would an answer recommend no brand?

The data does not say, and the following are hypotheses to test, not findings.

• The question gives little to choose by. In the October 2026 edition, best overall had no price ceiling and no stated use, and it had the highest rate. Budget and premium prompts, which set a constraint, had almost none. An assistant may answer an open question with how to choose instead of what to buy. The order of the rates is consistent with that, which is as far as the data goes.
• The answer names stores or product types. Retailers are not counted as brands.
• The answer names brands without recommending them. The data records mentions that were not recommendations, but the pack does not break them out for the 1,271 answers analysed here, so it cannot say how many sit in the 43.
• The extraction missed brands. An AI model reads each answer, and models make mistakes.

What should a brand do about it?

1. Keep no-brand answers in your denominator. Dividing only by the answers that name a brand inflates every brand's share by that share multiplied by 0.035 overall (a 20% share shows as 20.7%), by 0.100 on the best-overall question and by 0.118 in the Mexico sample. A tracker that drops zero answers makes a broad question look friendlier to brands than it is. Share of answer covers the definition.
2. Report the rate by question type. A no-brand share of 9.1% on one question and 0.3% on another means they are not the same measurement.
3. Read the sources of your own no-brand answers. When an answer names nothing, check its source list. For OpenAI that is the pages its answer links; for Perplexity, the pages its search retrieved. A domain in a list does not show that it shaped the answer, but it shows where the advice might have come from.
4. Consider a criteria-led buying guide. If an assistant answers an open question with how to choose, the pages that give clear criteria are the ones it may draw on. This is a reasoned guess. The data has no link between no-brand answers and their sources, and no evidence that a guide changes what any assistant says. The guide is useful to people either way: state who each option suits, the thresholds that matter, the trade-offs, and where your products fit with facts that can be checked. Buying guides and FAQs that AI assistants can quote covers how to write one.

To measure your own rate:

1. Ask the same open question ("best [product]") and a constrained one (a budget, or a use) in each engine, several times.
2. Read every answer. Sort the no-brand ones into four groups: advice only, stores only, brands mentioned but not recommended, and errors or refusals.
3. Divide each group by all the answers, per engine and per question.
4. Treat a rate from fewer than a few dozen answers as rough; how many AI answers you need works through the intervals.

What this cannot tell you

• It is the API, not the app. The apps may behave differently when no brand fits. See API answers vs app answers.
• It is one sample. The October 2026 edition asked each question once per engine, and 13 Perplexity answers failed and are in no count (9 of them outside the alternative question, which is why Perplexity has 631 answers here against OpenAI's 640).
• It counts, and does not explain. The brand lists were extracted by a model, and the reasons for a zero were not recorded.
• It is not about the Mexico market as a whole. It is 30 categories, asked in Spanish, in one edition.

For how long a typical shortlist is, see How many brands does an AI shopping answer recommend?. The edition's other headline numbers are in What 1,587 AI shopping answers show.

Frequently asked questions

Is a no-brand answer a refusal or an error?

The data cannot say. It records only how many brands each answer recommended, not why a count was zero, and it has no field for a refusal. A failed API call produces no answer, so it is in no count; a zero is an answer that came back and recommended no brand.

Will a buying-advice question produce more of these answers?

The October 2026 data cannot say, because this post uses only that edition, which has no buying-advice question. Its best-overall question names no brand and had a 9.1% no-brand rate, but it asks a different question. From November 2026, the 40 US categories that are updated monthly use a buying-advice question that names no brand, labelled "Buying advice" on their pages once the November update is published (methodology). Compare its no-brand rate only with the monthly updates, not with October's.