Hotel Fan-Out Mapper
When an AI assistant answers a hotel question with web search, it first turns the question into searches of its own. Those searches decide which pages it reads, and so which hotels it can name.
This tool shows the searches it actually ran: 36,227 of them, from 16,913 answers to 700 traveller prompts across 56 destinations. Other fan-out tools generate likely searches from autocomplete. These were read from the answers themselves.
Pick a destination
Each page lists the prompts (best hotels, luxury, budget, families, couples, business, solo, an amenity, a neighbourhood or landmark) and the searches behind every one.
What each engine adds to a hotel prompt
Share of answers whose searches added each topic, counting only words that were not in the traveller’s prompt. All destinations together.
ChatGPT keeps it short. 1.3 searches per answer, and 34% of answers only searched the prompt’s own words. When it adds something, it is most often “best”, a hotel class, the year or the country.
Grok searched wide and restricted. 5.7 searches per answer, a year in 98% of answers, and a site: operator in 41%, which limits the search to a list of booking and travel-media domains.
Method and limits
The prompts are the panel behind the AI Hotel Landscape, asked from the US and repeated every week. A fan-out is a search string the engine reported running before it answered.
- Natural prompts only. Part of the panel’s history was asked with an instruction to return JSON and “share the best link to book my stay”. That instruction leaks into the searches: 53% of ChatGPT’s searches mention booking under it, against 0.1% for the plain prompt. Those answers are left out, which also leaves out Perplexity: the panel never asked it a plain prompt.
- Engines that hide their searches. Google AI Mode, Copilot and Gemini as captured since April 2026 do not expose the searches they run, so they cannot appear here. Gemini’s data comes from its earlier interface (Jan 2026 to Feb 2026); Grok’s ends in Apr 2026. Only ChatGPT is current.
- Answers paired with the wrong prompt are dropped. One day of captures (1,623 ChatGPT answers) was loaded in bulk and some answers were filed under another prompt, so that whole day is excluded. Any other answer whose searches name a different destination and never its own is excluded too.
- ChatGPT’s counts are a minimum. The search list it exposes holds its web searches. It also looks up individual hotels in separate calls that the list leaves out, so it runs more searches per answer than shown here.
- Topics are word rules. A topic counts when an added word matches its list (for example “2026” for the year, “rooftop” or “spa” for amenities). Capitalised names are shown as written; run-together names such as “Ubud Seminyak Canggu” are not split.
- The page advice is editorial. It follows from what the engines search for. It is not a measured effect of changing a page.
Data built 2026-11-03.