Five Hotels Take Half the Answers. One Recommendation Went to a Demolished Building.

Across six US luxury markets captured on a single day, the three most recommended hotels in each market took 41 percent of every recommendation slot. In the average market, five properties accounted for half of everything AI recommended. In Chicago, Maui, and Napa Valley, four did.

Separately, in Miami, one of the hotels recommended that day had been demolished 108 days earlier.

Both numbers came out of the same fieldwork, and they belong in the same sentence. Recommendations pool into a few names, and the machine doing the pooling is not reliably aware of which buildings are still standing.

What was measured, and how

The AGR Luxury Hotel AI Visibility Index put ten questions travelers ask to ChatGPT, Google AI Mode, and Gemini, across New York City, Los Angeles, Chicago, Miami, Maui, and Napa Valley. Every query was run by hand, logged out, with no account state, in a fresh private window, on the consumer surface a traveler uses rather than through an API. All 180 answers were captured on July 29, 2026, in a single day, and preserved.

Those answers contained 824 ranked hotel recommendations. Each ranked hotel in each answer counts as one recommendation slot, and concentration is reported two ways: the share of a market's slots held by its three most recommended properties, and the number of properties required to reach half of that market's slots. Nothing is modeled. Nothing is projected.

The field is wide. The recommendations are not.

The ten questions span ten traveler intents, honeymoons through client entertainment. Pooled across all of them, the top three properties in a market took 41 percent of its slots.

The wider count sharpens the point. Across all six markets, the three platforms named 152 luxury properties at least once. Twenty-three of them accounted for half of the 824 recommendations. Being named once is common. Being recommended over and over is not.

Maui is the most concentrated market measured: the whole island came back as 14 properties across 60 answers on three platforms. Napa Valley is the most concentrated by owner. Properties belonging to a single collection, Auberge Resorts, took two of every five Napa recommendations. Ask AI about wine country and one company answers twice in five.

Consensus is a separate question from concentration, and there is less of it. Asked to name one hotel per market, the three platforms agreed in only two of six. In New York they returned three different answers to the identical question on the same day.

A very small visible source layer

Every answer's cited sources were logged. All ten of ChatGPT's Los Angeles answers cited the same two Michelin Guide list pages. All ten of its Chicago answers cited the same two Tripadvisor list pages. Gemini cited a single article from one lifestyle publisher on as many as eight of ten answers in a market.

What that establishes is narrower than it first appears, and still significant. These logs show what the systems display as their sources, not the full internal provenance of an answer, which no outside party can see. But the displayed layer is the one a traveler is shown, the one a journalist follows, and the one thin enough to count. In most markets, a platform put the same one or two documents behind every answer it gave about that city. Whoever published those documents is who the machine names when it is asked where the answer came from.

Whatever else that layer is doing, it is not being checked against the world. Mandarin Oriental, Miami closed permanently on May 31, 2025, and was brought down by controlled implosion on April 12, 2026. On July 29, ChatGPT recommended it three times and Google AI Mode twice, once inside its answer naming the top five luxury hotels in Miami. In the same fieldwork, ChatGPT offered The Ritz-Carlton Bal Harbour as an under-the-radar pick while the property sits closed for renovation, a closure disclosed on the same Forbes Travel Guide page these systems cite.

Two documented errors are two documented errors, not a measured error rate. What they establish is that whatever recency checks these systems apply failed on a property that no longer physically exists, and failed again on a closure published in a source they cite.

What credentials buy

An earlier and separate piece of AGR fieldwork, a pilot audit conducted in July 2026 across the same six markets, on the logged-in research modes of ChatGPT and Gemini, 25 traveler questions and 300 captures, found the other half of the picture. In New York, one of the seven Forbes Five-Star hotels in the city appeared in zero of 50 captured answers.

That finding belongs to the pilot, not to the Index, and its protocol was different. What the two studies say together is that institutional quality alone does not produce AI visibility. A rating is awarded by inspectors who visit. It does not put the property in the answer.

The obvious question

If a platform will put the same one or two documents behind a whole city of answers, the obvious question is whether a document built for the purpose can become one of them. AGR published one in New York and watched.

The motive is worth stating, since the reader will supply it anyway. AGR is a luxury hospitality firm and publishes market rankings for the reasons any firm publishes anything, and also to find out something specific. It sells no rooms and no placement, so a ranking of the best hotels in New York City, sourced to Forbes Five-Star ratings, Michelin Keys, and AAA Five Diamond awards, has no inventory behind it and nobody paying to be in it. What it has is a clean, dated, machine-readable record of a market. The discipline behind constructing it that way is Knowledge Formation Optimization, AGR's name for work performed on the public record, property by property and market by market, long before any traveler asks a question.

The Index discloses what followed, in its own terms: in two of the six markets, material published by AGR appeared among the cited sources, including in one market's top-five answer. The Index scores markets and not AGR's visibility, the capture set behind that disclosure is available on request, and a reader is entitled to know when the author's own material shows up in the data.

The New York observations are logged capture by capture, and the record is worth reading for its refusals as much as its results. AGR states there that it cannot demonstrate this work caused any classical search outcome, uses the word spillover for the search behavior it observed, and notes the observation tests none of its own falsification protocol. That restraint carries here without amendment. The observation is the narrow one: a document built to be well formed turned up among the displayed sources of systems writing the answers, in a market where that displayed set is small enough to count.

What this leaves a hotelier

The arithmetic is short. Half of every recommendation in a market goes to about five properties, across questions that span ten different reasons to travel. The documents these systems put their name to are few and identifiable, market by market. And a Forbes Five-Star rating did not keep one New York hotel from appearing in zero of 50 answers.

If a demolished building can hold a recommendation slot for 108 days, a property with a thin, stale, or contradictory public record has no protection either.

So the useful first question is not where a hotel ranks. It is what the machine says about the property, and which document it points to when it says it.