How ChatGPT Actually Picks Which Restaurants to Recommend

Most restaurant owners assume ChatGPT has some kind of database of restaurants it draws from. It doesn't. ChatGPT has no internal business database. Every time someone asks "where should I eat in San Antonio?", it runs a real-time search and assembles an answer from scratch.

In January 2026, French research firm Resoneo reverse-engineered ChatGPT's architecture. They documented 424 Feature Gates, 99 Dynamic Configs, and 237 Layer Configs. Combined with SOCi's empirical analysis of 350,000 business locations and Yext's study of 17.2 million AI citations, we now know exactly how the selection process works.

The 5-step pipeline

When someone asks ChatGPT for a restaurant recommendation, here's what happens in sequence:

Intent classification

ChatGPT runs a "Sonic Classifier" that determines whether your question needs a web search. If the confidence score hits 65%, it triggers a live search. Restaurant queries almost always trigger this because ChatGPT knows its training data for local businesses goes stale fast.

Resoneo

Three-source data retrieval

ChatGPT doesn't search one place. It queries three sources simultaneously: SerpAPI (scrapes Google Search results), an internal provider labeled "b1" (pulls Google Places data: ratings, hours, photos, categories), and the Bing Web Index (top 20-30 organic results). This is why Google and Bing optimization both matter for AI visibility.

Resoneo

Source filtering and cross-referencing

ChatGPT scans 5-8 sources in a "deeper dive" step, then selects 3-5 for the final answer. It filters out duplicates, generic directories, outdated lists, sites with no contact information, and businesses with contradictory data across platforms. If your name, address, or hours differ between Google, Yelp, and your website, ChatGPT treats you as unreliable and skips you.

ChatGPT self-report + Resoneo

Quality threshold

ChatGPT applies stricter quality filters than Google. The average star rating of restaurants ChatGPT recommends is 4.3 stars. Gemini's threshold is 3.9. Perplexity is 4.1. Businesses below these thresholds are effectively excluded. But rating alone isn't enough: ChatGPT only recommends 1.2% of businesses per query, compared to Google's 35.9% in its local 3-pack. You need the rating AND the data signals.

SOCi 2026 (350K locations)

Final selection: 2-4 restaurants

From the filtered pool, ChatGPT picks 2-4 restaurants to recommend. It prioritizes businesses with a "quotable differentiator" (something specific it can say about why this restaurant is worth visiting), consensus across multiple sources, and recent, active review patterns. There is no page 2 in AI. You're recommended or you're invisible.

SOCi + Yext

Where each AI gets its data

Not all AI platforms work the same way. Each has different source preferences for restaurants:

ChatGPT

Strongly prefers third-party directory listings. Yelp, Foursquare, and TripAdvisor are primary data feeds. Foursquare alone powers 60-70% of ChatGPT's local results in smaller cities. Business profile accuracy: only 68%.

Gemini

Prefers first-party websites and inherits Google Search ranking signals. Pulls directly from Google Maps with 100% business profile accuracy. If you rank well on Google, Gemini is your best platform.

Claude

Relies heavily on reviews and social media for restaurant recommendations. Cites user-generated content at 2-4x the rate of other models. In food and beverage, Claude cites UGC nearly 10x more than Gemini.

Perplexity

Frequently utilizes niche, industry-specific directories. Shows the most stable citation behavior across sectors. Average star rating threshold: 4.1 stars.

Source: Yext (17.2M citations across all 4 platforms), SOCi (2,751 brands, 350K locations)

What gets you excluded

The research identified specific factors that actively block restaurants from AI recommendations:

Inconsistent NAP data across directories. If your phone number differs between Google and Yelp, AI flags you as unreliable.

robots.txt blocking GPTBot. 5.9% of websites accidentally block AI crawlers. If your web developer added this, ChatGPT literally cannot read your site.

No "quotable differentiator." AI needs a specific reason to recommend you over the other 98.8% of restaurants. "Great food and atmosphere" is not a differentiator.

Image-heavy, text-light websites. AI can't read images. Restaurants that built their websites for Instagram (visual-first, minimal text) are systematically under-indexed.

No recent reviews. 20+ fresh reviews in 3 months increases ChatGPT probability by 2.5x. A restaurant with 500 reviews from 2 years ago loses to one with 200 reviews from the last 6 months.

Why this is fixable

80% of AI preference rules overlap across ChatGPT, Claude, Gemini, and Perplexity (Carnegie Mellon AutoGEO). Core optimization works cross-platform. The fundamentals are the same: consistent entity data, strong review signals, structured website content, and presence in the editorial sources AI retrieves.

None of this requires changing your food, your service, or your prices. It requires your business data to be organized for how AI reads it.

See what ChatGPT says about your restaurant

We test your restaurant across ChatGPT, Claude, and Gemini. Free. 30 seconds.

Get My Free Report

How ChatGPT actually picks which restaurants to recommend