How ChatGPT, Gemini, Perplexity, Claude and Copilot choose which local business to recommend

AI assistants choose a local business in one of two ways. Either they search the web and licensed listing data at the moment you ask, or they answer from what the model learned in training. For questions like "best tacos near me", the major assistants search. Each one searches a different index and reads different ratings: Gemini reads Google Maps, ChatGPT reads the Bing index and, since July 2026, Yelp, Perplexity reads Yelp and Tripadvisor. Most of the sources they cite are business websites and directory listings, and being named is rare.

Search mode and model memory

A language model can answer in two ways. Without search, it answers from patterns in its training data. That data has a cutoff date, so it cannot know that a restaurant changed its hours last month or closed last year. With search, the assistant sends queries to a search index, reads a handful of pages and listings, and writes an answer that cites them.

For local recommendations, search is what matters. Opening hours, ratings, addresses and "open now" are facts that change, and the assistants now fetch them live. Google says Gemini answers questions about places using Google Maps data covering more than 250 million places, including user ratings, with reviews used as a source for the answer (October 17, 2025). OpenAI documents a separate crawler, OAI-SearchBot, whose only job is to find sites for ChatGPT search answers.

Model memory still plays a part. A business that appears often in articles, guides and "best of" lists is more likely to be known to the model before any search runs. But the facts a customer acts on, such as the rating, the address and whether it is open, come from the search step.

Where each assistant looks, and whose ratings it reads

The table below summarizes what public sources say about each assistant as of the date shown in each cell. These arrangements change: treat every line as "true as of this date", not as a permanent rule.

Two practical points follow from this table. First, there is no single "AI rating". A restaurant can have 4.6 on Google, which Gemini reads, and 3.9 on Yelp, which ChatGPT and Perplexity can read. A cross-platform study of 21,841 US restaurants found Google ratings average 0.7 stars higher than Yelp ratings (Li and Hecht, 2020), so the gap is normal. Second, the Bing index matters more than its search market share suggests, because both Copilot and ChatGPT search draw on it.

What kinds of sources get cited

Several industry studies have counted the sources behind AI answers to local questions. They are vendor reports based on US data, so they describe what was observed, not a guaranteed mechanism.

Yext: 86% from sources the business manages

Yext analyzed 6.8 million citations in ChatGPT, Gemini and Perplexity answers to 1.6 million queries per model, collected in July and August 2025. Of those citations, 44% pointed to the business's own website, 42% to directory listings, and 8% to reviews and social media. The company groups the first two as sources a business manages, for 86% in total. The assistants differ: Gemini cited websites most often (52.1%), ChatGPT cited directories most often (48.7%), and Perplexity spread citations more widely, including Tripadvisor. In food service, directories accounted for 41.6% of citations and reviews and social media for 13.3%. Yext sells listing management, which is worth keeping in mind when reading its numbers.

BrightLocal: 58% business websites in ChatGPT

BrightLocal ran 800 manual queries across 20 industries and 20 US cities and recorded the first ten sources of each ChatGPT answer (published December 12, 2024). Business websites made up 58% of sources, mentions in publications 27% (39% of those were Wikipedia), and directories 15%. This was before the OpenAI and Yelp deal, so the mix has likely changed since July 2026.

A smaller follow-up in July 2025 compared Google AI Mode, Gemini, Perplexity and ChatGPT on 20 queries in 10 industries. Business websites were the top source for all four. Yelp appeared as a source in 33% of queries, and Perplexity cited Yelp in every industry tested. For restaurants and entertainment, Instagram, Facebook and YouTube also showed up.

Why the numbers differ

The studies use different query sets, assistants and dates, so their percentages are not directly comparable. What they agree on is the direction: the business's own website and its listings are the main material assistants use for local answers.

How rare it is to be named

A map search shows a list. An AI answer usually names a few businesses. SOCi's 2026 Local Visibility Index measured the difference across more than 350,000 locations of 2,751 chain brands (January 28, 2026). ChatGPT recommended 1.2% of those locations. In Google Maps, 35.9% appeared in the top three. That makes a ChatGPT recommendation about 30 times rarer than a top-three map position in this data. The locations ChatGPT recommended had an average rating of 4.3. According to Search Engine Land's summary, Gemini recommended 11.0% of locations and Perplexity 7.4%.

These are chain locations, not independent businesses, and the study comes from a company that sells local marketing software. It shows a correlation between ratings and recommendations, not that a rating causes one.

The audience is growing. In BrightLocal's 2026 consumer survey of 1,002 US adults, 45% said they use AI tools for local recommendations, up from 6% a year earlier. That is what people say, not measured behavior, but the change is large.

What this means for a business owner

No one can promise that an assistant will name a specific business. What an owner can do is make sure every source the assistants read is complete, consistent and current, and then measure what happens. A practical checklist:

1. Google Business Profile. Gemini and Google AI Mode read Google Maps data directly. Correct name, address, phone, hours, categories, attributes and photos. Google's own guidance says more reviews and positive ratings can help local ranking.
2. Bing Places. It feeds Bing search, which Copilot uses, and ChatGPT search relies largely on the Bing index. Check that it exists, is claimed, and matches your Google listing.
3. Yelp and Tripadvisor. ChatGPT (since July 2026), Perplexity and Apple Maps in the US read Yelp; Perplexity also reads Tripadvisor. Make sure both listings exist, match your website, and show recent reviews.
4. Apple Maps. Siri and Apple Maps show ratings from Yelp in the US and Canada and collect Apple's own ratings on top.
5. Your website. It is the largest single source type in both the Yext and BrightLocal studies. State the basics in plain text an assistant can quote: what you offer, where, for whom, prices or price range, hours, and how to book. Keep it consistent with your listings.
6. Crawler access. Check robots.txt. Blocking OAI-SearchBot means OpenAI will not show your site in ChatGPT search answers; blocking Claude-SearchBot or Claude-User reduces visibility in Claude, per Anthropic's documentation.
7. Mentions in third-party lists. In Whitespark's 2026 survey of 47 local search specialists, presence in expert "best of" roundups was rated the top factor for AI visibility. This is expert opinion, not a measurement.
8. Measure. Write down 10 to 20 questions a customer would actually ask, such as "best brunch in Silver Lake for a group", and ask each assistant the same questions on a fixed schedule. Save the answers with the date. That record shows whether you are named, how you are described, and which sources are cited.

How we approach it

At AIReputa we start with that measurement: a fixed set of questions run across ChatGPT, Gemini, Perplexity, Claude and Grok, with every answer saved. Then we fix the sources each assistant reads, from the Google and Bing listings to Yelp, Tripadvisor, Apple Maps and the website, and repeat the same questions to see what changed. We commit to the work and the measurement, not to a specific answer from any assistant.