What ChatGPT and Gemini recommend: 96 buyer questions, analyzed (2026)

More buyers now ask an assistant before they search. If you sell a booking app or a Shopify plugin, or you are the dentist down the street, the useful question is not whether AI search matters. It is: what exactly does ChatGPT say about my category, what does Gemini say, and where do they get it from? Nobody publishes that for small categories, so we measured it.

The short version

ChatGPT and Gemini rarely agree on what to recommend, and neither repeats itself reliably. ChatGPT reads the brands' own websites. Gemini leans on other companies' blogs, Reddit and YouTube, and skips the web entirely for over a third of questions. Established specialists win most mentions, and local businesses get the most random treatment of all.

How we ran the study

We picked 12 categories that look like the businesses we build tools for: seven kinds of small-business software (salon booking, freelancer invoicing, Shopify review apps, newsletter platforms, AI website builders, project management, a CRM for service businesses), three consumer apps (a habit tracker, a family password manager, a budgeting app for couples), and two local services (a family dentist in Austin and a dog groomer in Denver).

For each one we wrote eight questions the way a buyer types them, such as “We're a 5-person agency outgrowing Trello. What software gives us timeline views and client collaboration without being as bloated as ClickUp?” or “Where can I find an emergency dentist in Austin, TX that takes new patients on short notice?” Then we asked all 96 to both engines on October 10, 2026: GPT-5.4 mini through OpenAI's API with web search switched on, and Gemini 3.6 Flash through Google's Gemini API with Google Search grounding switched on. Each engine decided for itself whether to search.

Two honest caveats up front. These are API answers, not screenshots of the ChatGPT or Gemini apps, which can use other models, your history and your location. And the code that asked the questions is the same code behind AI Mention Watch, the weekly tracker we sell. That is the point: you can see exactly what it measures. The full method, including every judgment call, is at the end.

1. ChatGPT and Gemini agree on about a third of what they recommend

For the same question, the two engines' lists of recommended products overlapped by 34% on average. Of every product named for a question, 28% came from both engines, 29% from ChatGPT alone and 43% from Gemini alone. They put the same product first in 42% of questions.

Part of the gap is volume. Gemini wrote answers more than twice as long (576 words on average against ChatGPT's 243) and named more options in software categories. In local services the two named about the same number of businesses, just almost never the same ones.

Agreement depends heavily on the category. Where a market has a settled set of leaders, the engines mostly name the same ones: budgeting apps for couples, newsletter platforms and password managers all overlap by more than half. Where there are many interchangeable options, they scatter. Habit trackers overlap by 15%, and Austin dentists by 2%.

2. Ask the same question twice and you get a different list

We asked 24 of the questions a second time, a few minutes later. ChatGPT named only 48% of the same products again, and its first pick changed in 39% of repeats. Gemini was steadier with its products: it kept 74% and changed its first pick 25% of the time.

The twist is in the last row. Gemini kept naming the same products while citing mostly different pages to justify them: only 19% of its sources repeated. ChatGPT reread more of the same pages (45%) and still changed its list. Either way, one screenshot of one answer tells you very little about where you stand. That is why an AI visibility tracker asks several questions, to several engines, every week.

3. They read different parts of the web

ChatGPT searched the web for 88 of 96 questions, and 75% of its citations pointed at the official website of a product or business in the category, usually one it named in that same answer. Gemini searched for only 59 of 96, and when it did, it spread its citations across other companies' blogs, publishers, Reddit and YouTube.

Four things stand out:

• Reddit is a Gemini habit. 22% of Gemini's answers cited a Reddit thread. ChatGPT cited Reddit once in 96 answers. YouTube showed the same split: 16% of Gemini answers, no ChatGPT answers.
• Being cited is not being recommended. Gemini repeatedly leaned on “best X” posts written by small companies that it then did not recommend: getfinny.app in five budgeting answers, wiserreview.com in four Shopify answers, heybooked.com in three salon answers. Their articles shaped the list without putting their own product on it.
• Review sites matter less than you would expect. G2, Capterra and GetApp appeared in 4% of ChatGPT answers and 9% of Gemini answers. ChatGPT preferred app store pages (14% of answers), mostly the Apple App Store and the Shopify App Store.
• Gemini reads wider. It cited 279 different domains against ChatGPT's 131. And for 37 questions it read nothing at all, answering from what the model already knew. For those answers, what got into the model's training data mattered more than anything on the web that morning.

4. Established specialists win, small products get the long tail

In the ten software categories we sized every product by its estimated monthly US search visits. Products with a big footprint (over a million visits) were 11% of the names and got 20% of mentions. Established specialists in the middle got 56%. Small products were 46% of the names but only 24% of mentions, usually named once and never again.

That is not a closed door. 51% of software answers named at least one small product, and across all 12 categories 58% of the 371 names we saw appeared exactly once. The top of each category is far more fixed:

5. Local businesses get the least predictable answers

For the Austin dentist and Denver groomer questions, ChatGPT and Gemini overlapped by 7%, and in 9 of 16 questions they named no business in common. Sixty-three different practices and individual dentists were named across 16 answers, and none appeared in more than four. Most local citations were the businesses' own websites, followed by directories like Healthgrades and Zocdoc.

For a local business that cuts both ways. Nobody owns these answers yet, so a clear website and complete directory profiles can get you in. But a single check means even less here than it does for software.

What this means if you run a small business

• Check both engines, with several questions, more than once. They disagree with each other and with themselves. A tracker samples; a single screenshot misleads.
• Your own site is the first thing ChatGPT reads. Say plainly what you do, who it is for and what it costs, on a page that answers the question a buyer asks.
• For Gemini, get into what it reads. The comparison posts, Reddit threads and YouTube reviews in your category are the pages that shaped its lists. We turned the source data into a step-by-step playbook: how to get recommended by ChatGPT.
• Small is not invisible. Half of software answers named a small product. If the words are new to you, start with LLM visibility, explained plainly.

Limits of this study

It is one morning, 96 questions, US English and two API models, with 16 answers per category, so a per-category number moves a lot with one answer. Deciding what counts as a recommended product, and which pages count as a company's own site, takes judgment; we wrote every rule down. And AI answers change week to week. Read this as a dated snapshot, not a ranking.

Methods in full

• Questions: 8 per category, written by the same question generator AI Mention Watch uses from a one-line description of a made-up product, then read by hand. 17 of the 96 name a product, because real buyers ask for alternatives.
• Engines: OpenAI Responses API, model gpt-5.4-mini (2026-03-17 snapshot) with the web search tool; Gemini API, model gemini-3.6-flash with Google Search grounding. No system prompt, no location, no history. Asked on October 10, 2026 between 07:59 and 08:12 UTC, four requests at a time. No call failed.
• Products: extracted from each answer by a Gemini JSON call, then merged to brand level (Wix Bookings counts as Wix) with context mentions removed: operating systems, payment methods, insurers, directories, and the platform a question is about.
• Sources: the citation links each engine returned, up to eight per answer, typed by reading every domain by hand.
• Size: each software product's official domain, looked up in DataForSEO's US organic traffic estimate. A feature of a large company's product counts as that company.
• Cost: about $7 in API calls for all 240 answers, including the repeats.

Questions people ask

Do ChatGPT and Gemini recommend the same products? Mostly not. In our October 2026 test of 96 buyer questions, their product lists for the same question overlapped by 34% on average, and they put the same product first 42% of the time. For local businesses the overlap fell to 7%.

Where does ChatGPT get its product recommendations? When it searched (88 of 96 questions), 75% of its citations pointed at the official website of a product or business in that category, usually one it named in the same answer. It used app store pages often and Reddit almost never: one answer in 96 cited Reddit.

Does Gemini always search the web before recommending something? No. In our test Gemini ran no web search for 37 of 96 buyer questions and answered from what the model already knew. When it did search, it cited other companies' blog posts, publishers, Reddit, and YouTube far more than ChatGPT.

Can a small business get recommended by ChatGPT or Gemini? Yes. In the 10 software categories, 51% of answers named at least one product with a small search footprint. But each small product was usually named only once, while established specialists collected 56% of all mentions.