Do ChatGPT, Gemini and Claude agree on which software to recommend?

Mostly, no. We asked ChatGPT, Gemini and Claude the same 40 buyer questions across 10 software categories, such as "best feature flag tool for a small engineering team" and "cheapest transactional email API". Between them they named 290 different products. Only a quarter of those came up in all three assistants, more than half were named by just one, and the three agreed on the top pick in just one category. Asking again a day or three later changed the answer more than a third of the time.

If you sell software, the practical meaning is simple: being recommended by one assistant tells you little about the other two, and one check tells you little about next week. Here is what we found, how we measured it, and where the method has limits.

How we ran it

We track 10 software categories for our public AI picks pages: screenshot APIs, transactional email, product analytics, privacy-friendly analytics, uptime monitoring, error tracking, feature flags, form builders, newsletter platforms and AI coding assistants. Each category has 4 buyer questions written the way buyers ask, including one "X alternatives" question about a well-known product.

We asked each question to ChatGPT (GPT-5 mini), Gemini (Gemini 2.5 Flash) and Claude (Claude Haiku 5.5) through their APIs, without web search, between November 7 and 10, 2026. That's 120 answers: the latest one for every question and assistant. We pulled the product names out of each answer and merged spellings of the same product, such as JotForm and Jotform, Plausible and Plausible Analytics, or Amazon SES and AWS SES. Some questions were asked twice, one to three days apart, which gave us 36 repeat pairs to compare.

Two limits are worth saying up front. The consumer apps may use larger models than these, and they often search the web before answering, so this measures how the models' own knowledge ranks products, not every answer a buyer could see. And 120 answers is a snapshot, not a census. All 156 answers, including the earlier ones in the repeat pairs, are in a spreadsheet you can download with the products we counted in each, so you can check our reading.

Finding 1: they rarely name the same products

Of the 290 products named, 72 (25%) were named by all three assistants, and 165 (57%) were named by only one.

Error tracking and uptime monitoring had the most common ground. AI coding assistants had the least: 49 products named, and only 4 of them by all three assistants.

Finding 2: they pick different winners

For each assistant, we took the product it named most often across a category's four questions as its #1 pick, with ties broken by how early in the answer it appeared.

All three agreed in 1 of 10 categories. Two of the three agreed in six more, and in three categories each assistant had a different favorite. Counting pairs of assistants, they shared a #1 pick 9 times out of 30.

Finding 3: Gemini gives you the fewest chances

The assistants don't list the same number of products. ChatGPT named 8.6 products per answer on average, Claude 7.3, and Gemini 4.9. A shorter list is harder to get onto: if you're not one of the five or so products Gemini thinks of first, you're simply not in its answer.

Finding 4: ask again and the list changes

For the 36 questions we asked twice, one to three days apart, the two lists overlapped by only 55% on average. 84 of the 231 products named in the first answer, or 36%, were missing from the second. Nothing about these products changed in a few days. The answers move on their own, which is why the same question gets different answers and why a single check is a snapshot, not a verdict.

Finding 5: the best-known product usually leads, but not always

Each category includes a question like "LaunchDarkly alternatives", where the product in the question is rarely named in the answer. That pulls the best-known product's overall share down, so to compare fairly we looked only at the other three questions in each category, 9 answers each.

On those questions, the best-known product led or tied its category most of the time: Sentry was named in 100% of answers, Substack 89%, UptimeRobot 78% and GitHub Copilot 56%, each the top of its category. Typeform tied Jotform at 89%.

But not always. LaunchDarkly was named in 44% of answers, while Flagsmith was named in 89%. SendGrid was named in 78% and Amazon SES in 100%. Mixpanel was named in 67% and PostHog in 100%. Being the best-known name helps, but it doesn't guarantee a place in the answer.

What this means if you sell software

1. Check all three assistants, not only ChatGPT. More than half of the products in our data were named by just one of them. Our guide to tracking Claude covers why each needs its own check.
2. Measure over time, not once. With a third of a list changing in days, the useful number is your share across many answers. Share of voice explains how to read it.
3. Plan for the shortest list. Gemini's answers name the fewest products, so it rewards being among the clearest, most specific options in your category. How to get recommended by ChatGPT covers the pages that help with all three.
4. Don't assume your brand will carry you. Challengers like Flagsmith and PostHog are named ahead of better-known products in their categories.

See the data, and check your own product

The answers behind this study are in the data download: one row per answer, with the assistant, the model, the date, the products we counted and the full answer. The AI picks pages show the latest answers for every category and question, refreshed every two weeks. The feature flags list is a good place to start.

To see what ChatGPT and Claude say about your own product, run the free AI visibility check: 5 buyer questions in about a minute, no sign-up. To track your questions across ChatGPT, Gemini and Claude every day, start a 7-day Pro trial with no card.