Do ChatGPT, Gemini and Perplexity recommend the same businesses? The AI Agreement Index (2026)

Key findings

• 81% of the businesses recommended were named by only one of the three assistants. 15% were named by two, and only 3.9% by all three.
• In 68% of the questions, not a single business was named by all three assistants. In 18%, not even two of them agreed on one business.
• Each assistant answers on its own. On average, the three answers to one question named 11.3 different businesses, when each assistant gave about 4.6. Any two assistants shared only 13 to 15% of their recommendations (ChatGPT–Gemini 12.9%, ChatGPT–Perplexity 12.7%, Gemini–Perplexity 15.1%).
• It depends on the trade. For pizzerias and coffee shops, the assistants agreed on at least one place in most cities (70% and 60%). For dentists and car repair garages, they never did: 0 of 10 cities each, with 90% and 91% of the businesses named by one assistant only.

What it means for a local business

1. Being recommended by one assistant says little about the others. Check ChatGPT, Gemini and Perplexity separately: your customers don't all use the same one.
2. Trades where reputation is spread across many similar businesses (dentists, garages, plumbers) get the most scattered answers. There, the businesses that stand out are the ones described the same way in many places: website, Google Business Profile, directories, reviews.
3. Famous places converge. Businesses that the press and guides already talk about (some pizzerias and coffee shops) are the ones all three name. For most local businesses, consistent information everywhere is the realistic lever.

See also our guide How to check if ChatGPT recommends your business and the study The words AI uses when it recommends a local business.

Method

• Question: “What are the best places in {city} for: {trade}? Give me up to 5 names.”, asked in English, with the instruction “You are a consumer assistant. Answer in English, concisely, naming specific businesses when relevant.”. These are the settings PickableAI uses in its checks.
• Assistants: ChatGPT (gpt-4.1-mini with OpenAI web search, through OpenRouter), Gemini (gemini-3.1-flash-lite with Google Search grounding) and Perplexity (Sonar), all with live web search. Each question asked once to each assistant on 2 October 2026.
• Pairs: 10 trades × 10 cities were planned (London, Manchester, Dublin, New York, Toronto, Sydney, Melbourne, Auckland, Paris, Berlin). 68 pairs are complete and analysed (204 answers); the others were not collected, or one assistant named no business. Figures by trade are shown only for trades with at least 8 cities.
• Matching names: the businesses in each answer were listed by an AI model (Gemini, without web search), which also grouped mentions of the same business across the three answers (“Joe's Pizza” and “Joe's Pizza Broadway” count as one). We checked a sample by hand.
• Measures: for each pair, the distinct businesses named, and how many assistants named each one. “All three AIs agreed” means at least one business was named by all three.
• Names: business names are not published. The answers name real businesses that did not ask to be part of a study. The figures per pair are in the CSV file.

Limits

• One run per assistant. AI answers vary from one day to the next, even for the same assistant: a second run would give somewhat different lists, and could show more or less agreement.
• Name matching is done by an AI model, checked by hand on a sample only: a few mentions of the same business may have been counted as different (or the reverse).
• 68 pairs, 6 trades, 10 large cities: not representative of every trade, every town or every country. English questions only.
• Up to 5 names per answer: a longer list would give more chances to overlap.