An AI Travel Assistant Recommended a Hotel 108 Days After It Was Demolished

The Mandarin Oriental Miami was demolished in an implosion on April 12, 2026. One hundred and eight days later, ChatGPT and Google AI Mode were still actively recommending it to travelers — ChatGPT three times, Google AI Mode twice — in a single-day audit of 824 AI hotel recommendations run by Americas Great Resorts across six U.S. luxury markets on July 29, 2026, reported by Hospitality Net.

The demolished-hotel finding is the headline, but the concentration pattern underneath it is the bigger structural story: just 23 properties captured roughly half of all 824 recommendations, and the three most-recommended hotels in each market averaged 41% of that market’s slots. Maui was the tightest case — only 14 distinct properties appeared across 60 captured answers. The audit also found a sourcing problem: all ten of ChatGPT’s Los Angeles answers cited the same two Michelin Guide list pages, meaning a handful of upstream sources are doing an outsized share of the work deciding which hotels get discovered at all. A “credential paradox” compounded it — a Forbes Five-Star-rated New York hotel appeared in zero of 50 captured pilot answers, so established quality ratings didn’t guarantee AI visibility.

This isn’t an isolated glitch; it’s a pattern hospitality commentators have been flagging. Martin Soler of Soler & Associates warned specifically about AI-generated hotel-tech content that “looks highly professional yet frequently contains significant factual errors” — the same probabilistic pattern-matching that produces a polished market map can produce a polished, wrong recommendation, and neither comes with a disclaimer. Revinate’s own audit of hotel phone reservations found a parallel accuracy-adjacent gap: 92.2% of unbooked callers leave no digital trace, meaning even a hotel’s own channels routinely fail to capture the data that could correct a bad AI verdict downstream.

The fix isn’t complicated, just neglected: query what AI assistants say about a property regularly, and correct inaccuracies at their source rather than assuming visibility is static once earned.