# We asked ChatGPT, Claude and Gemini the same 300 questions. For local businesses, they rarely agreed

> We asked ChatGPT, Claude and Gemini 300 questions three times each. For local businesses, they gave three different top picks in 74.3% of cases.

Canonical: https://www.preferly.ai/blog/chatgpt-claude-gemini-same-300-questions
Published: 2026-11-12
Last updated: 2026-11-12

By the Preferly team

**Short answer:** for local businesses, mostly not. We asked ChatGPT, Claude and Gemini the same 300 questions, three times each, for 2,700 answers in all. On a local question, the three put three different businesses first in 74.3% of cases, and all three gave the same top pick in just 3.6%. For software, they agreed far more often.

This post is the summary. The full tables, method and data are in the study, [Do ChatGPT, Claude and Gemini recommend the same businesses?](https://www.preferly.ai/research/who-ai-recommends)

## How we asked

We wrote 250 questions about local businesses, in 10 industries across 25 US cities, and 50 about software for a small business, one per category. A typical question was "Who is the best family law attorney in Houston, TX?" No question named a business. Each assistant answered every question three times between November 9 and 11, 2026, with web search on, and after each answer it listed every business it named and its top pick.

## Six numbers from the study

A question-run is one question in one run, answered by all three assistants: 750 for local questions and 150 for software.

| What we measured | Result |
|---|---|
| Local question-runs with three different top picks | 557 of 750 (74.3%) |
| Local question-runs with the same top pick from all three | 27 of 750 (3.6%) |
| Local businesses named by one assistant only | 87.4% of names |
| Software question-runs with a name all three share | 125 of 150 (83.3%) |
| Local questions where ChatGPT gave the same top pick in all three runs | 208 of 250 (83.2%) |
| ChatGPT's cited addresses, local, that were a named business's own site | 66.5% |

## They rarely agree on local businesses

In about three of every four local question-runs, each assistant put a different business first. The share moved only between 71% and 77% from one run to the next, so it isn't the result of one odd day. Even a looser test, whether any single business appeared in all three answers, passed in only 91 of 750 question-runs (12.1%). Most names came from one assistant alone: 87.4% of the businesses named for a local question in a run were named by just one of the three.

## Software is the opposite

For the 50 software questions, all three named at least one of the same products in 83.3% of question-runs. They gave three different top picks in 14.0% and the same top pick in 37.3%.

## They also disagree with themselves

Asked the same local question in all three runs, ChatGPT gave the same top pick every time for 83.2% of questions. Gemini did for 39.6% and Claude for 39.2%. For software it was ChatGPT 76%, Gemini 84% and Claude 68%. Each of these findings gets its own post over the next few weeks.

## They read different sources

Of the addresses each assistant cited for local questions, 66.5% of ChatGPT's were a named business's own site, against 33.2% of Gemini's and 7.4% of Claude's. Claude's most cited sites were directories, review and booking sites. They also answered in different shapes: on a local question, Claude named 6.5 businesses per answer on average, Gemini 4.4 and ChatGPT 1.5.

## What it means for a business owner

- **One assistant naming you says little about the other two.** For most local questions, the three named different businesses first.
- **A single check is one draw.** Claude and Gemini changed their top pick between runs for most local questions. ChatGPT was steadier, but not fixed.
- **The sources differ, so the work differs.** ChatGPT cited businesses' own sites most. Claude and Gemini leaned more on third-party sites.

To see what AI says when your customers ask, run the [free check](https://www.preferly.ai/start). It asks 3 AI models the questions your customers ask, without naming you. To find out who AI names instead of you and whether what it says matches your website, get the [full $99 AI visibility audit](https://www.preferly.ai/ai-visibility-audit), paid once. It asks 10 AI models.

## Limits

- **No right answer.** "Best" has no ground truth. This study measures agreement, not accuracy, so it can't say which assistant is right.
- **No cause.** We measured that the sources differ. We didn't test whether the sources cause the different picks.
- **These channels, these dates.** ChatGPT was run in its app, Claude as agents with web search and Gemini through its API, in US cities, in English, over three days in November 2026. A small follow-up test on 48 questions found that the channel changes the answers, though that test changed two things at once for ChatGPT and Claude: the channel, and asking one question per request instead of 20 per file. We didn't run the consumer Claude or Gemini apps.
- **Automatic name matching.** A hand check of 120 questions disagreed with the matcher on 6 of 360 top-pick pairs (1.7%).

_Data: Preferly, "Do ChatGPT, Claude and Gemini recommend the same businesses?" (2026), https://www.preferly.ai/research/who-ai-recommends. 2,700 answers from ChatGPT, Claude and Gemini to 300 questions, November 9 to 11, 2026. The per-question data is free to reuse under CC BY 4.0 with that credit._

## Questions

### Which assistant gives the best recommendations?

The study can't say. A question about the best business has no single right answer, so we measured whether the assistants agree with each other and with themselves, not whether any of them is accurate.

### Did the questions name any business?

No. Every question asked for a recommendation, for example "Who is the best family law attorney in Houston, TX?", without naming a business.

### Can I use the data?

Yes. The per-question data is free to reuse under CC BY 4.0. Credit Preferly and link to the study page.

## Sources

- https://www.preferly.ai/research/who-ai-recommends