A customer in Amsterdam opens ChatGPT and types: "Where should I go for dinner tonight, somewhere with good pasta but not too expensive."
They do not get ten blue links. They get one answer. A short list. Three or four restaurants, recommended by name, with reasons.
If your restaurant is on that list, you got a customer. If it is not, you do not get a second chance to be considered. This is happening right now, in every city, every night.
Local Falcon's 2026 research found that 83% of restaurants are invisible in ChatGPT, compared to just 14% invisible on Google. That gap is the most important shift in restaurant marketing since the rise of mobile search.
This article explains the mechanism behind that gap, what the academic research says actually works, and what to do about it.
What generative engine optimization is
Generative engine optimization (GEO) is the practice of structuring information about your business so generative engines can understand, cite, and recommend it.
The term was formally introduced in 2024 by researchers at Princeton University and IIT Delhi, published at ACM SIGKDD, the top academic conference in data mining. Their paper distinguished GEO from traditional SEO structurally. Search engines rank a list of pages, and the user chooses. Generative engines synthesize one answer from multiple sources, citing some of them. The "ranking" is invisible. The citation is the entire prize.
For restaurants this changes everything. A customer asking ChatGPT "best ramen near me" does not see ten options to compare. They see two or three names, in a sentence. Every restaurant not in that sentence is invisible to that customer.
How generative engines find your restaurant
There are two ways large language models can answer a question about your restaurant. Understanding them is the foundation for everything that follows.
Parametric knowledge. This is what the model learned during training, stored in its neural network weights. The training data for major models includes Common Crawl (a monthly archive of the public web), Wikipedia, books, Reddit, and curated news. If your restaurant has been mentioned in those sources, the model "remembers" it. It can recommend the restaurant without looking anything up.
For most independent restaurants, this is unrealistic. Training data favors restaurants with media coverage, Wikipedia entries, or repeated mentions in editorial sources. A neighborhood bistro with no press is not in the training data of GPT-4 or Claude. The model has never heard of it.
Retrieval-augmented generation. When the model does not know something, it looks it up. ChatGPT's local-restaurant queries trigger an API call to Foursquare for business data and Mapbox for the map visual. Perplexity and Google's AI Overviews search the live web. The model then synthesizes an answer from what it just retrieved.
The implication is direct: for most restaurants, retrieval is the only path. And recent research shows that even when a restaurant is in the training data, generative engines have a strong bias to use what they just retrieved over what they remember. The retrieved content wins.
There are two ways generative engines find your restaurant: what they were taught and what they look up. For most independent restaurants, only one of those exists.
The signals that determine retrieval visibility are remarkably similar to organic marketing fundamentals: relevance, consistency, prominence, structured data, and authoritative mentions. The difference is which platforms count.
Why most restaurants are missing
When a generative engine retrieves restaurant information, it pulls from a different and broader ecosystem than Google Maps:
The websites of established food publications. ChatGPT routinely cites Eater, Time Out, Resy, Tasting Table, and local equivalents. One mention in a regional food blog can make a restaurant visible in answers for years.
Review platforms with structured data. Tripadvisor, Yelp, and Google reviews are read by language models specifically because they contain quotable human language. Models trained to synthesize answers prefer real customer quotes over marketing copy.
The restaurant's own website. Generative engines read your menu, your description, and your structured metadata. A site with no schema markup, a PDF menu, or vague descriptions sends weaker signals than a site with structured data, an entered menu, and clear category language.
Wikipedia, Reddit, and forum discussions. These are heavily weighted in both training data and retrieval because the content is densely cross-referenced and human-authored.
Restaurants ranking well on Google can still be completely absent from generative engine answers, because the underlying data sources differ.
What the academic research found works
The Princeton and IIT Delhi GEO paper tested optimization tactics on actual generative engines, including Perplexity. Their findings were specific and counterintuitive.
Three tactics consistently improved visibility by 30-40% across queries: adding statistics, adding direct quotations from authoritative sources, and explicitly citing sources by name. Traditional SEO tactics like keyword stuffing performed worse than no optimization at all. Generative engines penalize content that pattern-matches as marketing.
For restaurants, the principle is simple. The website description that says "we serve authentic Italian food in a warm atmosphere" is invisible to generative engines. The description that says "we serve handmade pasta with sauces refined over 23 years by chef-owner Marco Rossi, alongside a wine list of 47 Italian regional wines, in a 32-seat dining room with one Michelin Bib Gourmand mention" gives generative engines specific, citable facts.
The same principle applies to reviews, menu descriptions, and any content the model might encounter. Specific facts are quotable. Vague language is not.
The five sources generative engines check for restaurants
Six years ago, getting found online meant Google. Today the ecosystem is wider and shallower. Five sources matter most for restaurant visibility:
Foursquare. Still the primary live data source for ChatGPT's local queries. Most independent restaurants have never claimed their listing. Claiming it takes 15 minutes and costs nothing. This is the highest-leverage action for visibility.
Bing Places. Microsoft's local business directory feeds Bing search and integrates indirectly with ChatGPT. Coverage is patchy in Europe, but for English-language queries it remains a gateway.
Editorial publications. A mention in Eater, Time Out, The Infatuation, or a local equivalent (in the Netherlands: Misset Horeca, IENS, Smaakvol Eten) is more valuable than a thousand backlinks from low-quality sites. Generative engines trust editorial sources because the content is human-authored and curated.
Review platforms. Tripadvisor, Yelp, and Google reviews are read for the actual review text, not just the star rating. A review that says "best risotto I have had outside Milan" is quotable and citable. Generative engines reuse such phrases verbatim. The implication: encouraging guests to write specific, descriptive reviews matters as much as star count.
Your website with structured data. Schema.org LocalBusiness markup, an entered menu (not a PDF), an FAQ section, and clear cuisine and category language. Without this, your website is a collection of pretty pages that generative engines cannot parse.
Check your generative-engine visibility
The interactive checklist below is built around the actual sources generative engines pull from. Answer honestly. The result tells you where the largest gap is.
This is also the kind of work that restaurant marketing systems increasingly handle in the background, because the underlying signals are too distributed for an owner to maintain manually across a busy service week.
