AI assistants miss 85.6% of venues in a complete Bali market census

If you want to know whether a place is “discoverable” inside AI recommendations, this paper is useful because it doesn’t sample a convenient list of restaurants or lean on review-site rankings. It builds a complete operational census of 4,776 cafés, restaurants, and bars across two bounded Bali markets — Canggu and Ubud — and then asks four production assistants what they recommend under 96 persona-conditioned queries.

That makes the result much harder to dismiss as a corner case. The authors are not measuring whether a model can find a few famous spots. They are measuring how much of an actual market disappears when the same market is filtered through search-grounded AI answers.

85.6%.

---

What the audit measured, and why this version of “invisible” matters

The central move here is the census frame. Instead of taking a sample of venues and extrapolating, the paper enumerates the complete food-and-drink market within two study polygons using Google Places, then resolves AI-mentioned names back to that census. That gives the authors a concrete denominator for visibility: 4,776 venues, not an inferred market size.

They then run four production systems — ChatGPT, Claude, Gemini, and Perplexity — through their search-grounded interfaces over seven days. The confirmatory wave produces 2,208 search-grounded responses from 96 unique persona-conditioned queries, which in turn yield 12,439 valid venue mentions after extraction, filtering, and entity matching.

That setup matters because “invisible” becomes measurable rather than rhetorical. A venue either shows up in the answer set or it doesn’t, and the audit can distinguish between absence from the recommendation list and low rank within the list.

The paper is explicit about the frame limits. This is a complete enumeration of an operationally defined market, not a metaphysical one, and the invisibility figures are lower bounds. Even so, the design is strong enough to show a real pattern rather than a handful of bad outputs.

---

Most of the market never gets mentioned at all

Across the 4,776-venue census, 85.6% — 4,087 venues — were never recommended by any of the four systems. That is the headline result, and it is the one to keep in mind if you are thinking about AI as a new discovery layer for local businesses.

The gap is not limited to obscure or unreviewed places. Even among established venues with 50 or more Google ratings, 72.6% were never recommended. In other words, the systems are not simply surfacing the “best” places and leaving out the rest. They are leaving out most of the market, including venues that already have some reputation signal.

That is why the paper treats these invisibility rates as floors. The census is tied to a specific operational frame, Bali venue turnover means the market is a snapshot, and some venues outside the Google-listed food-service frame are underrepresented. But the main signal is still clear: AI recommendation coverage is sparse enough that long-tail invisibility is the norm, not the exception.

For practitioners, that changes the question. The problem is not just “how do we rank higher?” It is also “how do we enter the set that gets mentioned at all?”

---

Entry and rank are driven by different signals

The paper’s most useful contribution for operators is that it separates two margins: getting into the answer and getting the top slot once you are in it. Those two things are not governed by the same variables.

At the entry margin, documentation and web presence matter. Review volume has an odds ratio of 1.64, an own website has an odds ratio of 1.92, listed price information has an odds ratio of 1.54, and third-party web mentions have an odds ratio of 1.44. Star rating, by contrast, is null at this margin with an odds ratio of 0.89.

That is the practical signal. The system appears to use documentation as a gate to inclusion. A venue that has more surface area on the web — a site, price information, third-party mentions, review volume — is more likely to get into the recommendation pool. High quality alone is not enough to open the door.

The rank story is different. Among venues that are already recommended, rating predicts first position with an odds ratio of 1.17. So quality does matter, but later in the pipeline: it shapes rank inside the answer, not access to the answer itself.

---

The models are not converging on the same market view

Cross-system agreement is low enough that you should not assume one assistant is standing in for the others. The paper reports top-20 Jaccard overlap in the 0.33–0.54 range, which means the systems are drawing different partial maps of the same venue market.

That matters because a lot of people still talk about AI recommendation as if there were one generic visibility layer. This audit suggests there are several overlapping layers with different retrieval and construction habits. If a venue appears in one system but not another, that may not be noise. It may be a structural mismatch in how each assistant assembles its answer set.

The paper also shows that outright fabrication is rare: 0.08% of mentions. That is reassuring in one sense, but it is not the main failure mode. The bigger problem is omission, not invention. And the systems still recommended permanently closed venues 93 times across 14 confirmed-closed establishments, which is enough to show that stale or erroneous coverage remains part of the mix.

---

The churn looks stochastic, not stable

One of the more interesting controls in the paper is the two-week test–retest holdout. The authors reran 16 queries on all four engines two weeks after the main wave, and the cross-period answer similarity was comparable to same-day rerun similarity.

That pattern points away from slow temporal drift and toward sampling stochasticity. In plain terms: a lot of the variation you see is not because the market changed in two weeks. It is because the systems sample differently each time.

For anyone trying to optimize presence in these surfaces, that is a warning. If answer sets are partially stochastic, then visibility is not a single stable placement to be won once. It is a probability distribution you have to keep feeding with the right signals.

It also makes repeated measurement important. A venue that misses one run may not be permanently absent, but the paper suggests that absence is common enough that single-shot checks will understate the size of the problem.

---

What to do about it if AI recommendations matter to you

Treat AI recommendation as a visibility pipeline, not a reputation leaderboard. The paper’s entry-margin results say that if a venue lacks basic documentation, it is less likely to get mentioned at all — so the first job is to create enough machine-readable surface area for the system to see you.

That means the operational basics matter more than people often admit: a clear own website, visible pricing information, a steady trail of third-party mentions, and review volume. Those are the signals this paper links to inclusion. If you are advising venues, that is where the first layer of effort should go.

Then separate entry from rank. Once a venue is being surfaced, star rating starts to matter for where it lands in the answer. So the ranking problem is about maintaining quality signals, while the inclusion problem is about building the documentation footprint that gets you into the candidate set.

The broader takeaway is not that AI assistants are “wrong” in some dramatic sense. It is that they are incomplete in a way that is easy to miss if you only look at popular places. If you care about discoverability, you need to optimize for being included before you optimize for being preferred.