Most restaurant owners we talk to assume AI assistants have a single source of truth for their business — usually Google. The reality is more interesting. ChatGPT, Gemini, and Claude assemble their picture of your restaurant from at least seven distinct data layers, each with different update cadences, different signal strengths, and different vulnerabilities. Understanding which layer matters most for your specific case is the foundation of any honest AI visibility strategy.
Here's the layered model we use.
Layer 1: training-data baseline
When ChatGPT was last trained (the cutoff for gpt-4o is around April 2024, with newer specialized updates), it absorbed a snapshot of the public web. That snapshot included your Wikipedia entry if you had one, your TripAdvisor profile, English-language press mentions, your own website, and a portion of your Google reviews. This is your training-data baseline — the model's pre-existing knowledge of you before any live web search happens.
If your business opened after April 2024, gpt-4o has zero training-data baseline for you. You exist purely through Layer 2 and below.
Layer 2: real-time web grounding
Every modern AI assistant runs a parallel web search whenever the user asks a question that benefits from up-to-date information. ChatGPT calls this "web search," Gemini calls it "grounding," Claude calls it "browsing." The mechanism is similar across all three: a small model (or sometimes the main model) issues 1–5 search queries against Bing or Google, retrieves the top 10–20 pages, and feeds the relevant excerpts into the main model's context.
This grounding pass is where most of your influence lives. If your Google Business Profile description is current and detailed, it lands in the grounding context. If your website has clear positioning and structured data, it lands. If you have a recent positive press mention, it lands. The grounded snippets often outweigh the training-data baseline because they're newer.
Layer 3: review aggregators
TripAdvisor, Google Reviews, Yelp (in the US), Booking.com (for hotels), and OpenTable (for US restaurants) feed AI grounding heavily. Review sentiment density — not just star average, but how concentrated your reviews are around specific, repeated themes — affects how confidently AI describes you. A restaurant with 200 reviews mentioning "tasting menu" and "modern Anatolian" gets cleaner positioning in AI's mental model than a restaurant with 200 reviews scattered across ambiguous descriptors.
Layer 4: Wikipedia and Wikidata
If you have an English-language Wikipedia article, you have outsized AI influence. Wikipedia is the single most-cited grounding source across every major AI assistant we've tested. The article doesn't need to be long; a 200-word stub with proper citations contributes meaningfully. Wikidata (Wikipedia's structured-data sibling) is even more compounding — it canonicalizes your entity in a way that AI's knowledge graphs lock onto.
Most restaurants don't qualify for Wikipedia notability. The ones that do — Michelin-starred, historically significant, chef-led with significant press — should claim the article.
Layer 5: press mentions in trusted outlets
An English-language Eater profile, a Conde Nast Traveler review, an FT How-to-Spend-It feature — these are signal multipliers. AI assistants weight high-trust outlets higher than blog mentions or aggregator listings. The effect compounds: one strong piece in The Guardian is worth more than ten thin mentions on Yelp.
Layer 6: structured data on your own site
Schema.org Restaurant markup with menu sections, opening hours, price range, and accepted payment types gives AI assistants a clean parse of you. Most restaurants don't have this. Adding it is a one-hour task and improves Layer 2 grounding meaningfully.
Layer 7: NAP consistency across the web
Name, Address, Phone — these need to be identical across your Google profile, your website, your TripAdvisor, your Instagram bio, and any other directory listings. Inconsistent NAP confuses AI's entity resolution; the model sometimes treats two slightly-different versions as different businesses.
Worked example: a typical mid-tier London restaurant
Let's walk through what AI sees for a hypothetical mid-tier modern Italian restaurant in Soho, London.
Layer 1 (training-data baseline): opened in 2018. ChatGPT has minimal pre-existing knowledge — perhaps a single mention from a Time Out London listicle that was crawled in 2023.
Layer 2 (web grounding): when a user asks "best Italian restaurants in Soho," ChatGPT issues a Bing search and gets back the restaurant's website, three review aggregator pages, and one recent Eater article. The Eater article is the most influential — it lands first in the context window.
Layer 3 (review aggregators): 340 Google reviews averaging 4.4, 110 TripAdvisor reviews averaging 4-of-5, no OpenTable. Review sentiment density: high for "fresh pasta" and "Negroni," medium for "service," low for "atmosphere" (mixed reviews on the room).
Layer 4 (Wikipedia): none.
Layer 5 (press): Eater feature 2024, Time Out review 2022, single Guardian mention as part of a roundup, no other English-language coverage.
Layer 6 (structured data): partial — Restaurant schema with hours and address, but no menu structured data.
Layer 7 (NAP consistency): consistent across all listings.
What happens when ChatGPT gets asked "best Italian restaurants in Soho"? It cites this restaurant in the second or third position, with the description "fresh pasta and a strong cocktail program" — a paraphrase of the Eater piece, slightly compressed. It rarely uses the words from the restaurant's own website (which the model deems less authoritative than the third-party source).
What this means for action
The leverage is not where most owners assume. Updating your website copy is low-leverage; the AI usually uses the third-party article instead. Generating Wikipedia eligibility is higher-leverage but slow. Pitching one well-positioned English press mention is, at most price points, the single highest-ROI move.
Your audit's action list ranks moves by what specifically affects your layered profile, not generic best practices.