Do ChatGPT, Claude and Gemini recommend the same businesses?

1. In 74% of 750 local question-runs, ChatGPT, Claude and Gemini each gave a different top pick, and all three gave the same top pick in 3.6%.
2. Of the businesses named across 750 local question-runs, 87% were named by only one of ChatGPT, Claude and Gemini.
3. For software, ChatGPT, Claude and Gemini named at least one of the same products in 83% of 150 question-runs.
4. Asked the same local question three times, ChatGPT gave the same top pick every time for 83% of 250 questions, Gemini for 40% and Claude for 39%.
5. Of the addresses cited for local questions, a named business’s own site made up 66% of ChatGPT’s 1,197, 33% of Gemini’s 2,958 and 7.4% of Claude’s 3,883.
6. In a channel test on 48 local questions, the same model through its API gave the same top pick as its main channel for 12 of 48 (ChatGPT) and 10 of 48 (Claude).

We asked ChatGPT, Claude and Gemini the same 300 questions, 250 about local businesses and 50 about software, three times each: 2,700 answers. The numbers below are pooled over the three runs. A question-run is one question in one run, answered by all three assistants.

1. For local businesses, the three assistants rarely agree. In 74% of the 750 local question-runs, about three times in four, ChatGPT, Claude and Gemini each gave a different top pick (71% to 77% in each run). All three gave the same top pick in 3.6%.

2. Most businesses named came from one assistant only. Of the businesses named for a local question in a run, 87% were named by just one of the three. A name all three shared appeared in 12% of local question-runs (9.2% to 16% in each run).

3. Software is the opposite. For the software questions, all three named at least one of the same products in 83% of question-runs, and gave three different top picks in 14%.

4. Each assistant also disagrees with itself. Asked the same local question in all three runs, ChatGPT gave the same top pick every time for 83% of questions, Gemini for 40% and Claude for 39%. For software: ChatGPT 76%, Gemini 84% and Claude 68%.

5. They answer in different shapes, from different sources. For a local question, Claude named 6.5 businesses per answer on average, Gemini 4.4 and ChatGPT 1.5. Of the addresses each one cited for local questions, 66% of ChatGPT’s were a named business’s own site, against 33% of Gemini’s and 7.4% of Claude’s.

6. Claude cited one directory in 49% of its local answers. lantern.llc was cited in 119, 124 and 123 of Claude’s 250 local answers in the three runs. ChatGPT and Gemini never cited it. This depends on how Claude searches: asked 48 of the same questions through Claude’s API with its built-in web search, Claude cited it in none, against 22 for those questions in the main run.

7. The same assistant can answer differently through a different channel. In a small test, ChatGPT through its API gave the same top pick as ChatGPT in the app for 12 of 48 local questions. See the same assistant through a different channel.

Italian restaurants are the exception: the three assistants shared a name in 52% of question-runs and gave the same top pick in 27%. Real estate agents were the most split, with three different top picks in 95%.

The assistants split less often on small cities than on large ones: three different top picks in 63% of question-runs for small cities, against 81% for large cities.

For software, Gemini returned no cited addresses for 147 of its 150 answers, so its software sources can’t be analyzed.

Preferly, November 9, 2026 to November 11, 2026. 2,700 answers from ChatGPT, Claude and Gemini to 300 questions.

This is a small test: 48 of the local questions and 12 of the software questions, two runs each. We asked each assistant through its own API with its own web search, one question per request, with search on and then off, and compared the top picks with the main runs (ChatGPT in the app, Claude as agents). Gemini’s main channel is already its API. For ChatGPT and Claude the test changes two things at once: the channel, and one question per request instead of 20 per file.

• Across the three assistants, three different top picks came up for 21 of 48 questions through the APIs, against 32 of 48 in the main run.
• With search off, Claude declined to name a top pick for 31 of 45 local questions. ChatGPT declined for 1 of 48, Gemini for 2 of 48.
• On these questions, 6% of the addresses Claude cited were a named business’s own site in the main run, against 34% through the API.
• Software barely moves. Between two search-off runs the top pick was the same for 12 of 12 questions for ChatGPT, 12 of 12 for Claude and 12 of 12 for Gemini. ChatGPT in the app and through its API agreed on 9 of 12; Claude’s agents and its API on 8 of 12.

Again: the test is 48 local questions, and for ChatGPT and Claude it changes the channel and the number of questions per request at once, so it can’t say which of the two moved the answers. It says nothing about the consumer Claude or Gemini apps, which we didn’t run.

• Being recommended by one assistant says little about the other two. For most local questions, the three named different businesses first.
• A single check on one day 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. Your own website matters most for ChatGPT, which cited businesses’ own sites most. For Claude and Gemini, third-party directories and review sites matter more.
• Check the channel your customers actually use. In our 48-question test, ChatGPT through its API named the same top pick as the ChatGPT app for 12 of 48 local questions, so a tool that asks through an API and the app your customers use can name different businesses.
• See what each assistant says about your business with the free check. It asks 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, see what AI says about your business ($99 audit). The full report asks 10 AI models.

Questions. 300 questions: 250 about local businesses, in 10 industries across 25 US cities of three sizes (large, mid-size and small), and 50 about software for a small business, one per category. No question named a business.

Runs. Each assistant answered every question three times, between November 9, 2026 and November 11, 2026, with the question order reshuffled for each run. Web search was required for every question. After each answer, the assistant itself wrote a NAMED line listing every business it named and a TOP PICK line with its first choice.

The assistants.

• ChatGPT: gpt-6.1-sol, in the ChatGPT app with search on, 20 questions per file.
• Claude: Claude Sonnet 5.5, run as agents with web search, one search per question, 20 questions per file.
• Gemini: gemini-3.8-flash, through the OpenRouter API with web search on, one question per request.

Matching names. We matched names across answers automatically, after normalizing case, punctuation, legal suffixes such as LLC, and category words such as “law firm”.

How to read the percentages. Agreement percentages are out of question-runs: 750 local and 150 software. “Same top pick in all three runs” is out of questions: 250 local and 50 software, for each assistant.

• The assistants weren’t run identically. Gemini answered one question per request; ChatGPT and Claude answered 20 per file. In a test on 60 questions, a Gemini answer given in a batch of 20 had the same top pick as the one-per-request answer for 22 of 48 local questions, against 27 of 48 between two one-per-request runs. Batching shifts Gemini’s answers somewhat.
• Claude was run as Claude Sonnet 5.5 agents with web search, not in the claude.ai app, whose answers may differ. We asked 20 of the local questions in the app, which answered as Claude Opus 5.5: it cited lantern.llc in none of them.
• The results describe these channels, not the assistants everywhere. Asked through their APIs, the assistants answered differently: see the channel test, which is small and changes two things at once for ChatGPT and Claude.
• Name matching is automatic. In a hand check of 120 questions, the check disagreed with the matcher on 6 of 360 assistant pairs, and counted three different top picks for 79 questions where the matcher counted 81.
• Each industry row is 75 question-runs from 25 cities, so differences between industries of ten points or so aren’t reliable.
• US cities and English only, in one time window. The assistants change, and their answers will too.
• A question about the best business has no ground truth. This study measures whether the assistants agree, not whether any of them is accurate.

One row per question and run (900 rows): the question’s id, group (local or software), industry or software category, city and city size, the run, the question, how many businesses each assistant named, the distinct names, the names from one assistant only, and whether the three gave three different top picks, the same top pick, and a name all three share (1 or 0).

Download the data (CSV, CC BY 4.0)

The business names and the answers themselves aren’t published: this page reports totals only, so it doesn’t rank or name any business. The sites it names, lantern.llc and the most cited sites above, are sources the assistants cited, not businesses we asked about.

The data is free to reuse under the Creative Commons Attribution 4.0 International license (CC BY 4.0). Please credit: Preferly, "Do ChatGPT, Claude and Gemini recommend the same businesses?" (2026), https://www.preferly.ai/research/who-ai-recommends.

How to cite this study. Preferly, “Who AI recommends: Do ChatGPT, Claude and Gemini recommend the same businesses?”, 2026, https://www.preferly.ai/research/who-ai-recommends.