Gemini listens to YouTube and Reddit, ChatGPT almost never does

What we asked, and why

A buyer who wants a robot vacuum, a CRM or a carry-on suitcase now asks an assistant before a search engine. The answer names a few brands and stops. For a brand, and for the agency that works for it, two questions follow: which brands does the answer name, and what did the assistant read before naming them. We wanted numbers on both, on more than one assistant, because most reports on this subject look at ChatGPT alone.

So we wrote 100 buying questions in 10 sectors, with no brand name in any of them: phones, computers, home appliances, software for small businesses, money, travel, health and fitness, learning, home entertainment, cars. We asked each one once to ChatGPT, Perplexity and Gemini, with web search on, on 13 November 2026. Every question ended with the same sentence, “I'm in the United States”, so the three assistants answered the same buyer. We kept every answer, the brands it recommends and the pages it cites. The full study has the tables, the method and the data file.

Each assistant has its own ears

This is the clearest result. Gemini cites YouTube in 53 of its 100 answers and Reddit in 48. Count the answers that cite one or the other and you get 74 of 100. ChatGPT, asked the same questions on the same day, cites YouTube once and Reddit never: 1 answer of 100.

ChatGPT leans on publications that test products. Its most cited sites are RTINGS (18 answers), Tom's Guide (15), TechRadar (11), Wired (8) and Consumer Reports (7). Gemini cites those too, but after YouTube and Reddit: Forbes (18), PCMag (16), CNET (15), RTINGS (14).

The practical reading is simple. The two assistants can recommend the same vacuum cleaner and have learned it from different people. A brand that is well reviewed by test publications is in the material ChatGPT uses. That says nothing about the videos and the forum threads Gemini uses. We could not include Perplexity here: its answers name their sources in words, such as “Wirecutter's top pick”, but only 3 of 99 carried a link we could read.

The brand's own site is rarely the source

ChatGPT cited 210 pages across its 100 answers. 51 of them belong to a brand that the same answer recommends. Gemini cited 659 pages, and 34 of them belong to a recommended brand. Put another way, about three pages in four for ChatGPT, and nineteen in twenty for Gemini, are somebody else's page: a reviewer, a comparison site, a forum, a video.

This does not make a brand's site useless. A page has to be readable for an assistant to use it at all, and product pages are where prices and specifications come from. But the recommendation itself, the sentence that says “buy this one”, is built on what third parties wrote. A brand that only works on its own pages is working on the smaller part of the answer.

ChatGPT answers with a short list

ChatGPT names 2.4 brands per answer on average. Gemini names 5.4 and Perplexity 6. Take the 96 questions where all three recommended at least one brand. Question by question, Gemini and Perplexity recommended 737 brands between them. 512 of those are not in ChatGPT's answer to the same question.

We have to be careful with this one. We asked through the APIs, with one web search allowed for ChatGPT and a fast model, and a longer answer from a larger model would probably name more brands. So this is not a law about ChatGPT. It is a measured fact about these 300 answers: the same question returned two names in one place and six in another. Where an answer holds two names, being the third best known brand is the same as being absent.

They agree more than we expected

We started this study expecting three assistants to give three different answers. The data does not support that. In 86 of the 96 comparable questions, the brand ChatGPT puts first is also recommended by both Gemini and Perplexity. All three put the same brand first in 40 questions. Two of the three share the first brand in 47 more. Three different first brands happened in 9 questions only.

So the disagreement is not about which brands are good. It is about the order, the length of the list and the sources. That is a less dramatic headline than “the assistants contradict each other”, and it is the one the answers support. We would rather publish the smaller claim.

What this changes for a brand or an agency

This part is our reading, not a measurement. First, the work that gets a brand recommended happens mostly off its site: reviews, tests, forum threads, videos. An agency that only reports on the client's own pages is reporting on a small share of what the assistants read.

Second, one assistant is not a stand-in for the others. Checking ChatGPT and concluding for “AI” misses that Gemini listens to different people. Third, a single check is a snapshot. We asked once, on one day. Asked again next week, some answers will differ, which is why the same questions have to be asked on a schedule to mean anything. That is what FixFlow does for one site: the same questions, every week, on the same three assistants, with the count of who is cited next to named competitors.

Method and limits

One answer per question and per assistant, collected through the APIs and not the consumer apps: OpenAI's Responses API with web search, Perplexity's Agent API, Gemini with Google Search grounding. ChatGPT and Perplexity were allowed one search per question; Gemini decides how many it runs. An answer given without a web search was asked again, then dropped. 299 of 300 answers came back.

The brands were listed from each answer by a language model and merged when two names mean the same seller, for example Alienware and Dell. We checked the lists by hand on 20 answers. A page counts as a source when the answer links it, and as a brand's page when the site name contains a recommended brand. The questions are in English, about products sold in the United States: nothing here says what the same assistants answer in French or for a buyer in Spain. In a first run without the country, Gemini answered for a Spanish buyer and the other two for an American one, which is why we added it. The questions and every answer are in the data file on the study page.