A restaurant can have strong reviews, years of operating history and a loyal customer base and still be almost invisible when someone asks an AI assistant where to eat.

That is the central finding of a new study examining recommendations from ChatGPT, Claude, Gemini and Perplexity across 4,776 restaurants, cafes and bars in two Bali markets. The results suggest that AI-driven discovery is developing a gatekeeping layer of its own — and traditional measures of quality are not enough to guarantee entry.

The research, published as a preprint Aug. 7 and not yet peer-reviewed, analyzed 2,208 search-grounded responses to 96 persona-based queries collected over seven days in Canggu and Ubud. Researchers compared the answers against a complete census of the local market rather than a sample of businesses.

The result was stark: 85.6% of venues were never recommended by any of the four AI systems. Even among established venues with at least 50 ratings, 72.6% never appeared.

AI visibility appears to have an entry problem

The study points to a distinction that matters for businesses trying to understand AI discovery: the factors associated with getting included in an answer were not the same as the factors associated with ranking highly after inclusion.

Entry into recommendations was positively associated with signals of documentation and web presence, including review volume, an owned website, listed pricing information and third-party web mentions. Star rating, by contrast, was not statistically significant at the entry stage.

Once a venue made it into the recommendation set, ratings mattered more. Among recommended venues, a higher rating was associated with a greater likelihood of appearing first.

That is an important operating distinction.

Businesses have spent years optimizing for a search environment in which ranking is the visible problem. AI assistants can create a different problem first: whether the business is present in the system’s consideration set at all.

If the model does not have enough reliable, structured or repeated evidence to connect a business with a user’s request, the business may never reach the stage where its quality signals can help it.

Documentation is becoming distribution infrastructure

The findings fit a broader pattern emerging around AI recommendations.

DineVisible, a commercial platform tracking restaurant recommendations across ChatGPT, Claude, Gemini and Perplexity, has built its measurement model around a similar concept: presence rather than a single fixed rank. It evaluates how frequently venues are named across different prompts and engines, along with position, context richness and source diversity.

That approach reflects a basic characteristic of generative systems. There is no stable equivalent of a single Google position. The same business can appear for one type of request and disappear for another. Different engines can produce different competitive sets, and repeated runs can vary.

For an operator, that changes what “search optimization” means.

A website is no longer just a destination for human traffic. It is part of the evidence layer that machines use to understand what a company is, what it offers, where it operates, what it costs and whether other sources corroborate those claims.

Third-party coverage becomes more than reputation. Menu data becomes more than customer convenience. Accurate pricing, location information and structured business details become inputs into machine interpretation.

That does not mean businesses should manufacture mentions or flood the web with redundant content. It means their public information architecture increasingly affects whether AI systems can confidently place them in context.

The study also found a staleness problem

The research found outright fabrication was rare, accounting for just 0.08% of mentions. But the systems recommended permanently closed venues 93 times.

That may be the more practical risk for local discovery.

A model does not need to invent a restaurant to give a bad recommendation. It only needs to rely on stale evidence.

For businesses, that makes consistency across websites, maps, directories, review platforms and third-party references an operational requirement. For AI providers, it raises a different challenge: recommendation quality depends not only on reasoning but on the freshness of the retrieval layer.

Different AI systems are not seeing the same market

Agreement across the four systems was also limited. The study reported relatively low overlap among the top recommendations produced by different assistants.

That means there may not be one universal AI visibility strategy.

A business that appears consistently in one system can remain absent in another because the engines rely on different retrieval methods, sources, ranking logic and prompt interpretation. Measuring only one assistant can therefore create a false sense of security.

It also suggests that AI recommendation monitoring will need to be treated more like a recurring operational metric than a one-time SEO audit.

Quality still matters. It may just come second.

None of this suggests star ratings or customer experience have become irrelevant. The study found ratings became meaningful once a venue entered the recommendation set.

The more useful interpretation is that AI discovery may operate in two stages.

First, the system has to know enough about a business to consider it. Then it can decide how strongly to recommend it.

That is a very different problem from simply asking why a competitor ranks above you.

For companies that depend on discovery, the new question is becoming more fundamental: Does the machine know enough about us to put us in the room?