Why is my restaurant invisible in ChatGPT?

If a traveler asks an AI assistant where to eat, the answer may contain only a few venues from a market of thousands. A restaurant can be open, well reviewed, and easy to find in ordinary search, yet never enter the shortlist. That is not a small branding problem. It is a discovery gap.

A new August 2026 preprint, Invisible to the Machine, offers one of the clearest recent tests of that gap. It checked 4,776 operating cafes, restaurants, and bars across Canggu and Ubud in Bali against 2,208 search-grounded answers from ChatGPT, Claude, Gemini, and Perplexity. The study is limited to two local markets, but its measurement design exposes a mistake many businesses make: they treat being recommended and ranking inside the recommendation as the same event.

Most venues never enter the AI shortlist

The study found that 85.6 percent of venues were never recommended by any of the four systems. Even among established venues with at least 50 ratings, 72.6 percent never appeared. This does not mean those businesses are bad. It means the models had a narrow answer space, and most venues lacked enough visible, consistent evidence to enter it.

The paper found that entry into an answer was associated with review volume, having an own website, listed price information, and third-party web mentions. Star rating did not predict entry at this stage. Among venues that did get recommended, however, rating helped predict first position. In plain English: documentation helps you get into the room, while reputation can help you move closer to the front.

The authors also found 93 recommendations for permanently closed venues. Outright fabrication was rare, at 0.08 percent of mentions. For local businesses, stale information may be the more practical threat. A wrong opening time or old location can do more damage than a visibly absurd answer because it looks plausible to a buyer.

Why ChatGPT, Claude, Gemini, and Perplexity disagree

The systems do not share one local-business ranking. OpenAI's publisher guidance explains that ChatGPT Search uses web crawling and search partners. Anthropic's web search documentation describes source-backed web search in Claude. Perplexity says its Pro Search process can draw from articles, academic papers, forums, and other sources, with direct links in the answer. Gemini also has its own source and grounding behavior. Different retrieval paths mean a restaurant can be visible in one answer engine and absent in another without anything changing on its website.

The Bali study measured low cross-system agreement. Its top-20 Jaccard similarity ranged from 0.33 to 0.54, which means the systems shared only part of their top recommendations. Do not read that as a universal market-wide percentage. Read it as a warning against a single blended visibility number. Your restaurant may have a ChatGPT problem, a Claude problem, or an evidence problem that appears only when the prompt adds a budget, neighborhood, diet, or occasion constraint.

A related analysis from Yext on 17.2 million AI citations also found that the four systems use different citation patterns. That makes the operating answer simple, even if the technology is not: build a consistent public source of truth, then test it separately in each system.

💡Read tip: How to benchmark AI visibility against competitors

What local businesses should fix first

Do not respond by stuffing more keywords into a homepage. The study points to a more basic job: make the business easy to identify, compare, and verify. That means plain facts that can survive retrieval, not just attractive language.

Start with a GeoSnake free business scan to see how AI systems currently describe your business, which competitors they mention, and which facts are missing or wrong. The scan is useful only as a baseline. It does not prove that one prompt represents the whole market, so keep the prompt set fixed and repeat it.

1. Make the identity unambiguous. Put the exact name, address, neighborhood, cuisine or service category, price range, audience, and booking or contact path in visible text.

2. Publish the facts buyers actually filter by. State opening hours, dietary options, reservations, accessibility, menu or service scope, typical spend, and important limits. Add an update date where details change.

3. Keep public sources aligned. Check your website, business profiles, review platforms, partner pages, and local directories for conflicting names, addresses, hours, price ranges, and status.

4. Test constrained buyer prompts across four systems. Ask the same questions for budget, location, occasion, group size, dietary needs, and date. Record mention, recommendation order, citation, factual errors, and stale details.

5. Fix one bottleneck and rerun the test. If you are absent, improve category clarity and public evidence. If you appear but rank poorly, improve fit and reputation. If the facts are wrong, correct the source of truth before chasing more mentions.

This framework also explains why a high star rating alone is not enough. A model may not know what your venue is, where it is, what it costs, or when it is open. Reputation helps after discovery, but it cannot repair missing identity or contradictory facts.

Measure visibility as a funnel, not a trophy

Track four separate outcomes: the venue entered the answer, the venue was recommended for the right constraint, the answer cited accurate evidence, and the buyer reached a working next action. A single visibility score can hide the difference between those states.

For a weekly operating routine, see How marketers track AI visibility weekly. Store the exact prompt, model, market, language, date, cited URLs, and observed errors. That record lets you distinguish a real improvement from normal answer variation.

💡Related: Can AI search read the facts hidden in PDFs and images?

The uncomfortable conclusion is that many businesses are not losing AI recommendations because they need a clever new tactic. They are losing because the public web does not make the business easy to classify and verify. Fix the basic facts, test the exact buyer questions, and improve the stage where your venue disappears.

Run GeoSnake's AI visibility scan to see whether ChatGPT, Claude, Gemini, and Perplexity can find your business today. Then use the results to correct the weakest public signal and measure the same questions again.

FAQs

Does a high rating guarantee that AI search will recommend my restaurant?

No. The study found that rating helped with position among venues already recommended, but it did not predict entry into the answer. Identity, review volume, an own website, price information, and third-party mentions were more relevant to being surfaced at all.

Why does my restaurant appear in ChatGPT but not Perplexity?

The systems use different retrieval and citation paths, and they may respond differently to the same location, budget, or occasion constraint. Test each platform separately and compare the sources, not only the final mention.

What should I do if an AI assistant lists outdated restaurant information?

Correct the public source of truth first, then check major profiles, directories, review pages, and partner listings for the same error. Track the stale URL or statement and rerun fixed prompts after the changes have had time to propagate.