How Muse Spark searches for hotels
I ran the 616 hotel prompts of the AI Hotel Landscape through Meta’s Muse Spark API with web search on, and compared the answers with ChatGPT, Gemini and Google AI Mode on the same prompts.
Its searches return a median of 47 links per answer. It opens two pages and cites exactly the pages it opens. It searches one hotel at a time, mostly hotels it already knows, and names more hotels than any other engine, most of them hotels no other engine names.
What I measured
Muse Spark is the model Meta released in April 2026. The Meta AI app needs a login, so the tracking tools that added Muse this autumn measure it through Meta’s Model API, and so did I. This is the API with its web search tool switched on, which is close to the app but not the same product. I call it “Muse Spark (API)” throughout.
Each of the 616 landscape prompts (“best hotels in Lisbon”, “family friendly hotels in Dublin”, and so on) was sent once on 27 October, with the location set to the United States. All 616 came back without errors. For comparison I used the landscape’s captures of ChatGPT, Gemini and Google AI Mode from the week of 19 October, taken from the US. 607 prompts have an answer from all four engines, and the comparisons use those.
The same prompts were also sent to Muse with search switched off, as part of a study of what four models remember: Hotel recall without search.
It searches hotel by hotel
Muse searched on every prompt. The median answer ran 4 search queries (up to 15) and opened 2 pages. ChatGPT’s median is 1 query per answer. The API returns every query, and I had all 2,896 of them read by a model to find the hotels they name. The first query restates the prompt; only two first queries name a hotel, both for prompts that already do (“hotels near Bellagio”). The later ones go through hotels one at a time: 74% of them name a specific property.
Hotel names in queries identified by Claude Haiku 5.5; a spot check of 25 queries found no errors. Apart from a few landmark prompts such as “hotels near Marina Bay Sands”, the prompts name no hotel.
For “best hotels in Lisbon” it started with “best hotels in Lisbon 2026”, then searched for Condé Nast Traveller’s list, then for hotels by name. For “affordable hotels in Berlin under $200” its later queries included “Motel One Berlin Hackescher Markt price per night”. The median answer searches 3 hotels by name. The words added to those names show what it checks:
12% of queries add a year, almost always 2026 (322 queries) or 2025 (165). None uses a site: operator, and only 47 name a brand without a property.
Search results, opened pages, citations
The API reports three things for each answer:
- Search results. Each search comes back with a list of links. The API returns the links only, with no title or text, so it does not show how much of each result the model looked at.
- Opened pages. The model can also fetch a specific page. Most of the pages it opened (1,179 of 1,219) were links from its own search results.
- Citations. Links attached to passages of the final answer. One page can be cited many times: a Mumbai answer has 37 citation marks pointing to 3 pages.
Across all 616 answers, every cited link was a page Muse had opened: 1,149 of 1,149. No link was cited from a search result list alone, and 94% of opened pages were cited. The answers split the same way: 480 opened pages and cited them, 129 opened nothing and cited nothing, and 7 opened a page without citing it.
So the median answer has 47 search results it did not cite and 2 cited pages. ChatGPT’s consumer answers show 19 and 5. In both cases the uncited list is what the search engine returned, so the shares below compare the same two things.
The answers with no citation are the ones where Muse opened no page. They are complete answers: the Berlin one lists hotels with addresses and prices, and links to none of the pages those prices came from.
What search returns and what it cites
Classifying every link with the landscape’s domain categories shows a split between the two lists. Muse’s searches return online travel agencies and review sites. The pages it opens and cites are mostly small travel blogs and hotel guides. The most cited domains were thehotelguru.com, budgetyourtrip.com and santorinidave.com.
| Category | Muse, search results | Muse, cited | ChatGPT, cited |
|---|---|---|---|
| Small sites | 43% | 53% | 19% |
| Online travel agencies | 16% | 7% | 10% |
| Review sites | 10% | 2% | 34% |
| Editorial | 6% | 15% | 14% |
| Hotel's own site | 6% | 8% | 6% |
| Metasearch | 7% | 2% | 1% |
Share of unique URLs per answer. “Small sites” is everything outside the landscape’s named categories, mostly travel blogs and hotel guides.
The largest gap is on review sites. They are 34% of ChatGPT’s citations, almost all Tripadvisor. They are 10% of Muse’s search results and 2% of what it cites.
A US location, foreign sites
With the location set to the United States, 35% of the links Muse’s searches returned were on non-US country domains. The most frequent domain overall was tripadvisor.in, followed by the New Zealand, Canadian and South African editions of Tripadvisor. For ChatGPT the share is 13%.
Muse’s most frequent country domains in search results: .uk 1,901, .in 1,262, .ca 1,166, .nz 1,044, .ie 1,011, .au 919.
The effect is on the search-result side. Cited links are on foreign domains about as often for Muse (8%) as for ChatGPT (7%).
Where the hotels in its answer come from
Muse’s answers name 6,895 hotels in total. I compared each with the hotels Muse searched for by name in that answer, and with the hotels the same model names for the same prompt with search off.
Most hotels in the answer were neither searched for by name nor remembered without search: they come from the result lists of its broader searches. The hotels it does search for by name are mostly ones it already knows: 53% of them are in its no-search answer to the same prompt. Searching does not guarantee a place in the answer: 26% of hotels searched by name were left out.
Memory also shapes the top of the list. In 67% of answers, the first hotel of the searched answer was already in the no-search answer. A hotel Muse names for half or more of a destination’s prompts without searching appears in its searched answers 97% of the time.
Longer lists, and more of its own
Muse names more hotels than the other engines: a median of 10 per answer, against 6 for ChatGPT, 4 for Gemini and 3 for AI Mode.
61% of Muse’s hotel mentions are named by none of the other three engines for the same prompt, and 7% by all three. The top of its list is closer to the others than the tail: across the whole list, 35% of Muse’s hotels also appear in ChatGPT’s answer; counting only as many of Muse’s first hotels as ChatGPT named, the share is 44%.
Brand or not
I classified every hotel with the landscape’s chain classifier: big groups (Marriott, Hilton, Accor and similar, with their luxury sub-brands), luxury brands (Four Seasons, Mandarin Oriental, Aman, Rosewood and similar) and independents.
Value shown: independent share. Members of networks such as Small Luxury Hotels count as independent.
Muse’s mix is close to ChatGPT’s. Its longer lists do not lean towards chains. Where it differs is in the independents: a luxury-brand hotel in a Muse answer is usually also named by another engine, an independent usually is not.
| Parent brand | Mentions in Muse answers |
|---|---|
| Marriott | 456 |
| Accor | 350 |
| Hilton | 232 |
| IHG | 198 |
| Four Seasons | 186 |
| Hyatt | 172 |
| Mandarin Oriental | 79 |
| Shangri-La | 54 |
By destination
The overlap with ChatGPT varies by city. It is lowest in London and the Maldives and highest in Auckland and Sydney.
| Destination | Muse hotels per answer | Also in ChatGPT | Search results | Pages cited |
|---|---|---|---|---|
| London | 12.9 | 24% | 40.9 | 2.3 |
| Maldives | 13.4 | 24% | 39.5 | 2.4 |
| New York | 12.8 | 25% | 36.5 | 2.5 |
| Cape Town | 10.7 | 46% | 49.6 | 2.0 |
| Queenstown | 9.3 | 47% | 62.2 | 1.4 |
| Sydney | 9.7 | 49% | 51.1 | 1.4 |
| Auckland | 8.2 | 52% | 67.3 | 1.8 |
The cities at the top end are in countries whose Tripadvisor editions appear most in Muse’s search results (.nz, .au, .za). This data shows the two together and cannot say whether one causes the other. The cities with the lowest overlap are also the ones where Muse names the most hotels.
Cost, method and limits
The full run cost $12.34, about 2 cents per answer, on the API’s “contributor” tier: the same model at a lower price, in exchange for Meta being allowed to use the prompts and answers for training. Most of it was search fees, counting every search and every opened page as a billed search, which Meta’s pricing page does not confirm for opened pages.
- This is the API with web search on, not the meta.ai app. The app may search, cite and rank differently.
- Muse was captured on 27 October. The other engines are from the week of 19 October, eight days earlier.
- One answer per prompt. Answers vary between runs, so small differences between cities are within the noise.
- Hotels are matched across engines by name, not resolved to a hotel database. Hotels that only Muse names have not been checked one by one; the companion memory study found that 0.7% of Muse’s remembered hotels are invented (upper bound 5.3%).
- Search results are links only. The API does not show whether the model saw a snippet or any text from them. Opened pages are the only pages known to have been fetched.
- Gemini’s links are not measured from 19 October, when its captures stopped including the rendered answer. AI Mode’s cited links are mostly Google place links.
Disclosure: I work at Lighthouse, which owns The Hotels Network.
Summarize with AI
