Hotel recall without search: what AI remembers, and how much survives a web search

What I measured

Are the hotels real?

Each name was matched against a directory of real hotels for its destination. Between 68% and 86% of each model’s hotels are in it, and a further 4% to 6% were named by a search engine for the same city. I checked 25 of the remaining names per model by hand, after a Google Places lookup: of the 100, 81 are real hotels, 4 are invented, 14 are uncertain and 1 is not a hotel.

The directory is incomplete (196 hotels for Queenstown, 12,620 for Beijing), so “unverified” means not found in these references. The upper bound counts every uncertain case as invented and adds the sampling error (95% interval). Claude Haiku named no hotel at all in 15 answers, giving neighbourhood advice instead.

The invented cases are mostly real names in the wrong city: “Circa 039 Hotel” in Buenos Aires resembles Circa 39 in Miami, and “Hotel NoMo Athens” borrows from NoMo SoHo in New York. When the prompt names a destination, these models rarely make hotels up.

Memory predicts what search finds

For each destination I counted how often a hotel appears in the remembered answers (its memory visibility) and how often it appears in the searched answers to the same prompts. A destination has about 11 prompts, so a hotel remembered in half of them is one the model reaches for whatever the question.

Unit: one hotel in one destination. Memory visibility: share of that destination’s prompts whose no-search answer names it.

Seen from the search side, 64% of ChatGPT’s hotel mentions go to hotels GPT-6 Luna names somewhere in that destination from memory. The figure is 59% for Muse, 63% for the Gemini app and 46% for AI Mode. Across all four memories, 77% of ChatGPT’s mentions and 67% of AI Mode’s go to a hotel at least one cheap model remembers.

The first hotel is usually a memory

Overlap for a whole list is moderate: 38% of Muse’s remembered hotels are in its searched answer to the same prompt, and 45% of GPT-6 Luna’s are in ChatGPT’s. The top of the list is different.

In two prompts out of three, the hotel a searched answer names first was already in the remembered answer. Hotels a model ranks in its top three from memory are about twice as likely to stay after search as those further down. Search mostly changes the bottom of the list.

Search queries carry the memory

Muse exposes the queries it sends to its search engine, and ChatGPT’s are captured by the landscape. I had each of the 3,681 queries read by a model to find the hotels it names.

Brand or not

I classified every hotel with the landscape’s chain classifier into three groups:

• Big group: Marriott, Hilton, Accor, IHG, Hyatt and similar, including luxury sub-brands such as Ritz-Carlton or St. Regis.
• Luxury brand: Four Seasons, Mandarin Oriental, Aman, Rosewood, Peninsula, Belmond, Shangri-La and similar.
• Independent: everything else, including members of networks such as Leading Hotels of the World or Small Luxury Hotels.

Value shown: independent share.

The strongest memories are also where luxury brands concentrate. Among hotels a model names in half or more of a destination’s prompts, 24% are luxury brands, against 4% among hotels named once or twice.

Value shown: luxury-brand share.

Independents are still the majority at every level, because many famous hotels are independent: Le Bristol, Eichardt’s, La Mamounia, Hotel Monteleone. But the hotels search adds that were not remembered are 64% to 76% independent. For a smaller independent hotel, search is the main way into an answer. For a Four Seasons, memory already does much of the work.

One prompt, eight answers

“Best hotels in Paris”, with the four memories on the left and the four searched answers on the right. The table lists the hotels named by at least two answers.

Le Bristol and the Crillon are the only hotels all four memories name, and both are in Muse’s and the Gemini app’s searched answers. ChatGPT’s six hotels include neither. Le Meurice, which only Haiku remembers, is named by all four search engines.

From June to October

The June study listed the ten hotels GPT-5.4 mini remembered most often in Paris, Dubai, London and New York. Four months and one model generation later, most of them still appear across each city’s 11 prompts:

Muse names nearly all of them, with and without search. GPT-6 Luna names six to eight. In Paris, Mandarin Oriental and Hotel du Louvre, both remembered in more than 70% of June’s runs, are named by none of the October answers from GPT-6 Luna or ChatGPT. The two studies used different prompts, so this checks persistence and does not measure change.

How the remembered answers read

Muse quotes a price in 70% of its answers without searching, more often than when it searches (38%). The other three models rarely quote prices from memory. Muse also mentions 2026 42 times across its no-search answers, often in openings such as “Here are my top picks for 2026”, for lists that come from training data. Claude Haiku tells the reader to check or verify in 39% of its answers.

What this means for hotels

• Being remembered is a head start, not a guarantee. A hotel a model names across many prompts for its city is very likely to appear after search too, and near the top. Memory is built from what was written before the model’s training cutoff, which for current models is several months before release.
• Search is where most hotels get in. Most hotels named by search were remembered weakly or not at all, and the ones search adds are mostly independents. For them, the pages search engines retrieve matter more than the model’s memory.
• Engines differ. Muse searches hotel by hotel, starting from names it already knows. ChatGPT sends one broad query and builds its list from what it finds.

Method and limits

• One answer per prompt and model, at default temperature. Visibility is measured across a destination’s prompts, not across repeated runs of one prompt.
• The memory models are the cheapest of each family. Except for Muse, the searched answers come from larger models, so “memory vs search” is a family comparison. Larger models are likely to remember more.
• The landscape captures are from the week of 19 October, eight days before the memory runs. Gemini and AI Mode lists combine several captures per prompt.
• Names are matched on their distinctive words, after removing words such as “hotel” and the city name. A brand word alone never makes a match. In a hand check of 40 matches, 4 pointed to a sister property, so overlap figures are slightly generous.
• The brand classifier works from the hotel name. A property listed under its own name with no brand in it counts as independent.
• Search queries were tagged by Claude Haiku 5.5. A spot check of 25 tagged queries found no errors.
• The invented-hotel estimate rests on 25 hand-checked names per model, judged by one person.
• Cost: $2.01 for the four memory runs, about $1.47 for hotel extraction, $4.10 for Google Places and under $1 to tag the queries.

Disclosure: I work at Lighthouse, which owns The Hotels Network.