Skift Research ran one set of travel questions through three large language models and changed the traveller each time. Every model rewrote its description of the trip. The destination list held nearly steady.
The research house published the work on 1 October 2026 as "The Mystery of the Disappearing Destination: How AI Is Transforming Travel Discovery". Skift Research analyst Saniya Zanpure wrote it. Skift's own news item followed on 16 October 2026.
What the study tested
The study used ChatGPT, Gemini and Claude. Each model received repeated prompts under traveller personas that varied by demographics, interests, travel companions and budget. The models changed their language. Travellers with children received detail on family-friendly activities. Travellers on a limited budget received low-cost flight options.
The report rests on three inputs. Skift Research worked from a booking-journey diary study published with McKinsey & Company, a survey of more than 1,000 U.S. travellers, and the 2026 AI Destination Study.
The concentration finding
Skift Research reports that the top five destinations in its recommendation set took 38 percent of all recommendations. The top ten took 48 percent.
The report's chapter list carries the argument in order: "Global Travel Is Already Concentrated", "AI Recommendations Are Also Concentrated", "But Less So Than Travel Demand" and "AI Adapts the Recommendation Pitch, But Not the Recommendation List".
The report places global international arrivals in a handful of well-known places. It places the spread of AI recommendations slightly wider than the spread of those arrivals.
A second test measured the distance between fit and output. The study rated 29 percent of persona-destination pairs as a high fit, and no model surfaced them in a recommendation.
Where AI sits in the booking journey
The report places AI at 21 percent of touchpoints during inspiration, 16 percent during research, 6 percent at booking and 29 percent after the booking is made.
Comfort with the tools follows that shape. Between 40 and 50 percent of respondents were comfortable using AI for ideas, comparisons and practical details. For booking, the figure was 27 percent.
Travellers wanted the shortlist kept. Sixty-three percent preferred a curated set of options to an unfiltered catalogue or a single model-chosen itinerary.
Skift Research describes the sequence as a new gate in the funnel. The traveller moves from awareness to AI consideration, then to search and validation, then to booking.
What the shortlist gate means for daytime inventory
The 38 percent figure turns the destination list into the first filter. A model that settles on five places sets the list a traveller picks a property from. Day-use inventory sits inside that filter.
Two windows in the study carry the operator's attention. The post-booking window is the largest AI touchpoint the study measured, at 29 percent of the journey. The booking stage itself carries 6 percent.
The report names three levers in its chapter list: "What May Drive AI Visibility", "Local Content May Improve AI Visibility" and "AI Visibility Depends on Upstream Sources". The third names the part an operator can act on, because the sources a model reads decide which properties it can name.
Independent audits reproduce the concentration finding on narrower samples. A peer-reviewed study in the journal AI carries DOI 10.3390/ai6090236 and a publication date of 19 September 2025. It generated 6,480 recommendations for 216 traveller profiles and reported popularity and geographic bias across the models it tested. Work in Urban Science carries DOI 10.3390/urbansci9070268 and a publication date of 10 July 2025. It reports a much smaller and more spatially concentrated set of Barcelona locations from ChatGPT than from Instagram. A third study, arXiv 2411.07320, ran about 100,000 travel requests through five models and reports fewer distinct locations in recommendations covering poorer countries.
Two figures point at the same window. The post-booking period carries 29 percent of AI touchpoints in the study, and 29 percent of high-fit persona-destination pairs never reached a shortlist. A property that wants a place on the list needs current, structured content in the sources the models read.