A business can survive being ranked.
It cannot survive being uncounted.
A complete census of one real market, 4,776 restaurants, cafés, and bars in Bali, tested against 2,208 AI recommendation runs. 85.6% never appeared in a single answer. Not ranked low. Never considered.
On 7 August, a research team filed an audit built on a premise most discoverability studies avoid because it is expensive: count the entire market first. Not a sample. Not the businesses that showed up in a search. Every Google-listed restaurant, café, and bar across two bounded submarkets of Bali, Canggu and Ubud, 4,776 of them, enumerated before a single AI system was ever asked a question.
Then the researchers put four AI systems, ChatGPT, Claude, Gemini, and Perplexity, through 2,208 recommendation runs across 96 persona-conditioned queries: different traveler types, different needs, different budgets. Because they held the full census going in, they could measure something citation studies cannot. Not who ranks where. Who was never considered at all.
The filed number: 85.6 percent of the market, 4,087 real, operating businesses, never appeared in a single recommendation. Across any system. Across any persona. Across all 2,208 runs.
The reflex objection is that the invisible must be marginal — the food stall without a sign. The filing anticipated it. Among established venues with fifty or more ratings, 72.6 percent still never surfaced. These are businesses with standing inventories of customer evidence. The machine did not weigh that evidence and find it light. The machine never reached it.
The two margins
The mechanism the audit documents deserves more attention than the headline number. In this audit, recommendation operates on two separate margins. The first margin is documentation: a working website, review volume, listed prices, third-party mentions — the machine-readable evidence that an establishment exists and can be described. That margin correlates with whether a venue enters consideration at all. The second margin is quality: the star rating. That margin only predicts position among venues that already crossed the first. The authors call the resulting measurement an “invisibility rate.”
Read that structure again, because it inverts two decades of search-era instinct. Quality does not purchase candidacy. Quality is consulted only after candidacy — after the documentation margin has already been crossed. A 4.9-star establishment with thin documentation loses to a plainer competitor with a complete machine-readable record — not by ranking below it, but by never appearing in the same universe.
The record has a name for this. Decision Exclusion — the fifth clinical characteristic of Digital Derangement Syndrome™: decisions made without the entity ever entering consideration. It has always been the least measurable of the five, precisely because absence leaves no trace. You cannot see the recommendation you were never part of. What this filing supplies is the census-grain measurement: when someone finally counted an entire market and checked it against what the machines produce, exclusion was not a tail effect. It was the majority condition.
Two margins, not one. Documentation correlates with whether a venue enters consideration. Rating only predicts position among venues that already cleared that first gate. Nearly every visibility effort a business undertakes — better content, sharper positioning — operates on the second margin. For roughly six businesses in seven in this census, the second margin never came into play.
Optimization is tactical. Installation is strategic. The uncounted 85.6 percent did not have an optimization problem. No optimization budget crosses a margin that documentation, not rating, gates.
Provenance, stated plainly. The study was funded and conducted by Norly, a company that sells review-management and AI-visibility tools to hospitality businesses — a commercial interest directly adjacent to its own research question, which the filing discloses. It has not been peer-reviewed. Its own terms ask marketing derivative works to quote only published effect sizes, and this record does.
The filing carries weight here for one reason: the method — a complete market census, openly documented — can be audited by anyone willing to count. Weigh the funding. Then weigh 4,087 businesses.
The audit counted restaurants — one region, one language, a tourist-and-nomad market small enough and desperate enough to be counted completely. Whether the same two-margin structure holds in other languages, other markets, is a testable question this filing opens, not one it answers. But every market has a census, even the ones nobody has audited yet — a full roster of legitimate competitors, most of whom believe that being good is the same as being findable. Where the machines consult a different roster, no amount of second-margin effort closes the distance.
A business can survive being ranked. It cannot survive being uncounted.
V. Pitenin, “Invisible to the Machine: Auditing AI Restaurant, Cafe, and Bar Recommendation Against a Complete Market Census,” Norly Research, filed 7 August 2026, arXiv:2608.07069. Study funded and conducted by Norly (norly.co); not peer-reviewed at time of filing.