What 1,587 AI shopping answers show: the October 2026 Dopbot Index

We sent 800 shopping prompts to each of two AI engines in October 2026, 1,600 requests in all, and collected 1,587 answers. The other 13 failed, all of them Perplexity's. This post sets out the edition's main numbers with the engine and the number of answers beside each one, and then says what they cannot show. Everything here describes answers to prompts we wrote. We did not observe shoppers or their searches.

How was the sample built?

The edition covers 160 categories: 100 in the United States, 30 in Mexico (prompts in Spanish) and 30 in Canada. In the October 2026 edition each category has 5 prompts: best overall, a budget with a price ceiling, premium, a specific use case, and an "alternative to [a named brand]" prompt. Each prompt was asked once per engine, which gives about 10 answers per category (160 categories x 5 prompts x 2 engines = 1,600 requests).

The two engines are OpenAI's API with web search (gpt-6-luna) and Perplexity's Agent API with the Sonar model (perplexity/sonar), with web search on for both. Requests carry no account, search history or memory, and the web search location is set to the market's country. The methodology has the full rules.

The 13 missing answers are Perplexity requests that failed, so Perplexity's counts rest on 787 answers in all, not 800, and on 631 when the alternative question is left out.

How many sources does each engine show per answer?

The two engines record sources differently, so we report them side by side and never add them together. For OpenAI, a source is a page its answer links. For Perplexity, it is a page its search retrieved, and the API does not say which of those pages the answer relies on. That difference in what is recorded is part of why the second column is so much longer; the data cannot say how much of the gap it explains.

An average taken across both engines would describe neither of them. Our companion post on the websites in these answers lists each engine's most frequent domains with its own denominator, and what an API sample can and cannot tell you explains the gap to the apps.

How many brands does an answer recommend?

Excluding the alternative-to-a-brand question, which we explain below, the 1,271 remaining answers recommended a mean of 2.3 brands and a median of 2. OpenAI's 640 answers averaged 2.2 and Perplexity's 631 averaged 2.5. Most answers are short lists: 84.4% recommended one to three brands, 11.8% recommended four to six, 0.4% recommended seven or more, and 3.4% recommended none.

For a brand, the practical reading is that an answer to a shopping question has room for about two names, so most brands in a category are absent from most of its answers. A count of recommendations needs many answers behind it before it means anything.

Do the two engines recommend the same brands?

Only partly. We measure overlap per category: the brands both engines recommended, divided by the brands either engine recommended, averaged over categories with equal weight. It is not answer-by-answer agreement.

Excluding the alternative question, the mean overlap across the 160 categories was 47.0%. On average across the 160 categories, 3.9 brands were recommended by both engines and 5.3 by only one. By market the mean overlap was 47.5% in the United States, 46.3% in Mexico and 45.9% in Canada. Including the alternative question, which lists more brands per answer, the mean is 53.6%. In this edition each category has about 8 answers behind it without that question, so the overlap for any one category is noisy; the mean over 160 is steadier. The full overlap analysis shows why this means a brand should measure each engine separately.

Do budget and premium questions get the same brands?

Rarely. In 155 of the 160 categories we have 2 budget answers and 2 premium answers, one per engine for each prompt. For each category we compared the brand lists of those answers (shared brands divided by brands in either) and averaged: the mean overlap is 8.2%, and in 70% of categories the budget and premium answers shared no brand at all. For comparison, the best-overall answers overlapped with the budget answers by 23.8% and with the premium answers by 26.8%.

Read this as the observed overlap of short brand lists from about four answers per category, not as proof that price wording changes recommendations or as a statement about what shoppers prefer. It still suggests a check worth running: whether your brand appears for the kind of question you want to be in. The budget versus premium post has the detail.

What did the alternative-to-a-brand question teach us?

October's fifth prompt asked for alternatives to a named, well-known brand in the category. Its 316 answers recommended 5.0 brands on average, against 2.3 for the other four prompts. The named brand itself was recommended in only 21 of the 296 answers where we could find it, 7.1%.

Because the prompt names a brand, its answers are about replacing that brand, and they recommend roughly twice as many brands, which inflates brand counts and raises the mean engine overlap (53.6% with it, 47.0% without). That is why the measures in this post exclude it unless we say otherwise. From the US run of November 2, 2026, the 40 categories that are still updated replace it with a buying-advice question that names no brand.

What do the Mexico and Canada samples show?

They are descriptive, and the markets are not like for like: Mexico's 30 categories are different categories, with prompts in Spanish. Excluding the alternative question, 10.5% of Mexico's 237 answers recommended no brand, against 1.8% of the United States' 797 answers and 1.7% of Canada's 237. We do not attribute the Mexico gap to language, because categories and prompts differ too.

For the 24 categories that appear in both the US and Canadian lists under the same name, the brands recommended in the two countries overlapped by a mean of 47.6% excluding the alternative question, and 54.9% with all five prompts. It is one illustration, not a finding. The Mexican and Canadian pages are at /mx/index and /ca/index.

Which websites sit behind the answers?

Per engine, with a different denominator each:

A strict own-site match means a domain whose first label equals a recommended brand's name (samsung.com for Samsung). It misses country domains, parent companies and other store names, so it probably undercounts. The retailer and forum lists were also made by hand and are not complete. A domain in a source list does not show that it influenced the recommendation. Wirecutter's numbers and the full domain tables are in their own posts.

What this data cannot tell you

• It is the API, not the app. The apps can differ in model version, instructions, search tools and personalization, and they know more about the user.
• Categories have few answers. In the October 2026 edition, with about 10 answers per category, one answer moves a brand's share by about 10 points. Per-category numbers are examples, not rankings.
• 13 answers failed. All were Perplexity's, so Perplexity's counts rest on slightly fewer answers than OpenAI's.
• The prompts are fixed. In the October 2026 edition each was asked once per engine. A different wording, or the same wording asked again, would give different answers.
• Presence is not influence. A domain in a list tells you the page was linked or retrieved, nothing about whether the answer used it.
• Only 40 categories keep updating. From November, 40 US categories are re-sampled monthly with 40 answers each. The other 120 (60 in the US, 30 in Mexico, 30 in Canada) stay as this October edition. The two designs are not like for like, so this data shows no change over time.

What can a brand take from it?

1. Expect short lists and measure with enough answers. About two brands per answer means a few answers tell you little. Our free manual method explains how many to collect.
2. Track each engine separately. Under half of the brands overlap, so showing up in one engine says little about the other.
3. Check budget and premium questions on their own. Their shortlists barely overlap in this sample.
4. Keep your own product facts accurate. OpenAI's links included a strict own-site match in 46.6% of answers, and the engine links only about two pages per answer. We cannot say your site caused a recommendation, but a page that is linked should be right.
5. Do not read one domain as a lever. Perplexity's retrieved lists are long and varied; a domain's presence there is a pattern in sampled answers, not a switch.

Frequently asked questions

Where can I see the numbers for my category?

Each category has a page in the US index, the Mexico index or the Canada index with its prompts, the brands recommended and counts per engine. The monthly report is at /reports/2026-10 and the brand and source counts behind it can be downloaded as CSV files at /data/2026-10.

How does Dopbot decide that a brand was recommended?

An AI model reads each answer and lists the brands it names, in order, and marks whether the answer presents each as a recommendation or only mentions it. Retailers and marketplaces are left out, and a brand counts at most once per answer. The methodology describes the merge list for brand names that come back in more than one form.