A Bali AI Audit Found Most Listed Venues Were Never Named

A restaurant can be absent from an AI answer for several reasons. The question may not fit it. The system may retrieve information about competitors. It may use a different name for the same place. Or the answer may have room for only a few recommendations.

To interpret the absence, you need more than the answer. You need a defined set of businesses that could reasonably be considered and a record of the questions asked.

Invisible to the Machine takes that approach in Canggu and Ubud, two Bali markets. It assembled a Google Places-based census of 4,776 food-and-drink venues within defined geographic boundaries, then compared that list with recommendations from search-enabled API setups for four AI systems.

Across its sampled answers, 85.6% of those venues were never named. That is a substantial coverage gap in the audit. It does not mean those venues could never appear for any question, language or consumer-app setting.

The denominator changes the question

A typical spot check begins with a handful of well-known competitors. If the assistant names most of them, it can look as though it covers the market well.

This study began with an operational market list instead. That makes a different question possible: how much of the listed market appears anywhere in the answer sample?

The researchers used 96 persona-conditioned queries and collected 2,208 search-grounded responses in the confirmatory wave. The questions were in English and reflected tourist or nomad-oriented needs. The results describe those questions, those API configurations and that period. They should not be presented as a measurement of every local customer's experience in the consumer apps.

The census has a boundary too. It covers venues represented by the study's Google Places frame and geographic rules. A business absent from that frame cannot enter the denominator. Calling it a census makes the counting method explicit; it does not make it an exhaustive description of every way the local market could be defined.

Appearing and ranking are separate outcomes

The paper distinguishes entry into the sampled recommendations from position among the venues that appeared. That is useful for an audit because these outcomes require different measurements.

If a venue never enters your answer sample, its average position among recommendations is undefined. Assigning a convenient ranking number can hide the more basic fact that it was not named. If it appears regularly but near the end, the problem is different.

The researchers found associations between inclusion and several documentation-related variables, including having a website, listed price information, review volume and third-party mentions. Star rating behaved differently across entry and ranking analyses.

These associations do not establish that adding a website or a price field causes a recommendation. Businesses differ in many connected ways. A visible venue may also be older, larger or better represented across sources. A star rating is a recorded platform signal, not a direct measurement of the quality of the restaurant experience.

The result gives you useful variables to inspect. It does not supply a proven optimization checklist or show that the assistant uses those variables as fixed gates.

Match the business before measuring its visibility

Names deserve attention in any local audit. An assistant might abbreviate a venue, use a previous name or mention a branch without its location. A text match alone can count a real recommendation as a miss, or credit the wrong business.

Keep the original answer and resolve uncertain names against addresses or other identifying information. Record ambiguous matches separately. A confident-looking percentage cannot repair a mistaken business identity underneath it.

Freshness is another concrete issue. The study recorded 93 recommendations of 14 venues confirmed as permanently closed. That is a different failure from not covering the long tail of open businesses. A useful audit should keep stale recommendations, uncertain matches and simple nonappearance separate.

Repeated answers also varied. The study's retest results are a reason to save multiple runs, not to treat one screenshot as a permanent placement. Variation across products adds another reason to identify exactly which surfaces the report covers.

Ask for the frame behind the score

Before acting on a local visibility percentage, ask for the business list, geographic boundary, exact queries, language, product interface and repeated answers. Then ask how names were matched and what counted as appearing.

With those records, you can distinguish a venue missing from the sample from one represented inaccurately or outranked among included alternatives. Without them, “invisible to AI” describes a score whose meaning you cannot inspect.