AI Hotel Landscape
How leading AI assistants recommend hotels — 616 prompts across 56 destinations, refreshed every Monday.
+ 44 more
The 11 prompts we ask ChatGPT about Florence
Same 11 question patterns asked about every destination — persona (couples, families, solo, business), budget (luxury, mid, budget), and zoom (city-wide, neighborhood, landmark).
Named hotels in Florence — ChatGPT
ChatGPT's top picks specifically in Florence, IT. Pick a different destination above, or clear to see worldwide.
Most-mentioned hotel by ChatGPT this week.
| # | Hotel | City | Chain | Mentions | W/W |
|---|---|---|---|---|---|
| 2 | Caruso, A Belmond Hotel, Amalfi Coast | Ravello | Belmond | 6 | ↑ 1 |
| 9 | Anantara Convento di Amalfi Grand Hotel | Amalfi | Minor | 5 | ↑ 3 |
| 16 | Santa Caterina Hotel | Amalfi | — | 5 | ↑ 3 |
| 27 | Le Sirenuse | Positano | — | 4 | ↑ 2 |
| 70 | Il San Pietro di Positano | Positano | — | 3 | ↓ 1 |
| 68 | Hotel Poseidon | Positano | — | 3 | NEW |
| 66 | Hotel Miramalfi | Amalfi | — | 3 | ↑ 1 |
| 65 | Hotel Lungarno | Florence | Lungarno | 3 | ↑ 1 |
| 62 | Hotel Brunelleschi | Florence | — | 3 | NEW |
| 57 | Four Seasons Hotel Firenze | Florence | Four Seasons | 3 | — |
| 48 | Borgo Santandrea | Amalfi | — | 3 | — |
| 196 | Hotel Villa Franca Positano | Positano | — | 2 | — |
| 191 | Hotel Marina Riviera | Amalfi | — | 2 | — |
| 186 | Hotel Calimala Florence | Florence | — | 2 | NEW |
| 170 | Glance Hotel In Florence | Florence | — | 2 | — |
| 139 | Bvlgari Hotel Roma | Rome | Marriott | 2 | NEW |
| 323 | Villa San Michele, A Belmond Hotel, Florence | Fiesole | Belmond | 2 | — |
| 312 | The St. Regis Rome | Rome | Marriott | 2 | NEW |
| 311 | The St. Regis Florence | Florence | Marriott | 2 | ↓ 1 |
| 303 | The Setai, Miami Beach | FL | — | 2 | ↓ 1 |
When ChatGPT hands out a hotel link, where does it go?
Every URL the AI returned for a named hotel, classified into direct (the hotel's own site), OTA (Booking / Expedia / …), chain page (marriott.com / hilton.com / …), or other. Chain folds into direct in the headline number because chain pages are still brand-controlled.
Week by week. Google is split out of what used to be counted as “other”: a link to google.com from an AI assistant is the assistant keeping the click, which is worth seeing on its own rather than filed under the vaguest available label. Metasearch (trivago, kayak) stays in other — those are third-party marketplaces, not the engine linking to itself. Google is a worldwide figure here; the other lines are scoped to this view.
One note on the other line, because it used to be the least trustworthy series here. “Other” is every link we cannot place as the hotel’s own site, an OTA, a chain page or Google — so it is partly a measure of our own coverage, not only of engine behaviour. It previously showed a band of 45–65% between 08 June and 27 July that we could not explain. That band was a bug on our side, and it is now fixed. Google AI Mode returns hotel websites as bare domains (ritzcarlton.com) rather than full URLs, and the query that derives the domain for classification expects a scheme — so 16,194 perfectly classifiable links, every one of them AI Mode, were counted as “other” purely for lacking an https://. Those URLs have been normalised, the ingest no longer stores them scheme-less, and every week has been rebuilt: other is now 3.5% across all engines and weeks, and the band is gone. Read direct and OTA as behavioural; still read a move in other as measurement first.
Weeks with no classified links are omitted rather than drawn as zero — a missing run and a genuinely zero week look identical on a chart, and only one of them is a finding. Link volume per week: 05-04 27 · 05-11 45 · 05-18 51 · 05-25 56 · 06-01 51 · 06-08 44 · 06-15 46 · 06-22 39 · 06-29 51 · 07-06 56 · 07-13 60 · 07-20 44 · 07-27 64 · 08-03 66 · 08-10 50 · 08-17 46 · 08-24 39 · 08-31 95 · 09-07 50 · 09-14 45 · 09-21 36 · 09-28 42 · 10-05 38.
Sponsored placements in Florence
OpenAI is rolling out paid sponsor placements in the US, Australia, New Zealand, and Canada — testing may expand. We detect real single-advertiser ad units (organic shopping cards excluded) and pull the brand.
Where ChatGPT pulls hotel information from
Every URL ChatGPT retrieved classified into 12 buckets: OTAs, editorial, chain sites, direct hotel pages, social, meta-search, AI tools, community, government, directory consortia, and other.
100% of the URLs we captured were cited in the answer. A ratio that high does not mean ChatGPT uses almost everything it finds — it means we are no longer seeing what it discards. ChatGPT sat at 24.8% on 10 August, before the retrieval layer stopped coming through. Publishing 97% as selectivity would invert the finding, so the number is withheld rather than shown. The counts below are what was captured, and are unaffected.
Editorial cntraveler.com 50% · thehotelguru.com 29% · timeout.com 21% | 31.8% | 41 |
Other florencetourism.org 36% · roomfornight.com 36% · myconciergehotel.com 27% | 24.0% | 31 |
OTA & Aggregator booking.com 94% · hotelscombined.com.au 6% | 13.2% | 17 |
Community hotelierschoice.com 85% · hostelworld.com 8% · mumsnet.com 8% | 10.9% | 14 |
Review Site tripadvisor.com 78% · oyster.com 22% | 7.0% | 9 |
Independent Hotel krafthotel.it 50% · grandhotelminerva.com 25% · lloydsbaiahotel.it 25% | 4.7% | 6 |
Hotel Directory hotelsforkings.com 80% · only5star.com 20% | 3.9% | 5 |
Chain all.accor.com 33% · anantara.com 33% · nh-hotels.com 33% | 2.3% | 3 |
Meta-search hotelscombined.com 50% · momondo.com 50% | 1.6% | 2 |
Search engine google.com 100% | 0.8% | 1 |
Share of citations by source type, week by week, for Florence. Search engine is its own bucket rather than part of metasearch, and covers Google (every regional domain), Bing and the rest. Google AI Mode routes its cited sources through google.com/searchviewer instead of linking the publisher, and those citations were landing under “meta-search” alongside trivago and kayak — which made the mix read as competitive pressure from metasearch when actual metasearch is under 1%. The dashed rule marks the week that started.
Which categories get a line is decided once over the whole window by total citations, not week by week — otherwise a category dropping out of the top N would look like it collapsed to zero. Everything below the cut is summed into “smaller categories” rather than dropped, so the lines always account for the full week. Citations per week: 04-27 108 · 05-04 160 · 05-11 16 · 05-18 9 · 05-25 167 · 06-01 409 · 06-08 389 · 06-15 166 · 06-22 192 · 06-29 270 · 07-06 269 · 07-13 80 · 07-20 120 · 07-27 125 · 08-03 104 · 08-10 78 · 08-17 76 · 08-24 71 · 08-31 90 · 09-07 52 · 09-14 35 · 09-21 37 · 09-28 46 · 10-05 39.
Top hotel parent groups in Florence — ChatGPT
Each ChatGPT mention matched against Google Places, then rolled up to parent group (Marriott = Ritz-Carlton + Westin + Sheraton + EDITION + …, etc.). WoW delta vs last week. Detection via 175+ brand-domain rules.
| # | Brand | Mentions | Hotels | W/W | |
|---|---|---|---|---|---|
| 1 | Lungarno | 5 | 2 | ↑ 25.0% | |
| 2 | Four Seasons | 3 | 1 | · 0% | |
| 3 | B&B Hotels | 3 | 3 | · 0% | |
| 4 | Belmond | 2 | 1 | · 0% | |
| 5 | Marriott | 2 | 1 | ↓ 50.0% | |
| 6 | Rocco Forte | 1 | 1 | ↓ 50.0% |
Of the hotels named this week: how many resolved to a chain (Marriott / Accor / …), an independent property, a vacation rental, or stayed unmatched.
What every metric means
Plain-English definitions for each number on this page. For deeper visuals + examples, see the annual landscape report linked at the bottom.
- Captures
- Number of AI responses we collected this week. Target = 616 per platform (one per prompt). Less when there are upstream errors.
- Web search
- Did the response trigger a live web fetch? Detected from the underlying response stream’s search-result events, not the unreliable top-level flag. We do NOT force web search — this is organic model behavior.
- Map widget
- Did the response render a hotel-card map (the Google-Maps-style widget ChatGPT shows for travel queries)? Each card is a "map entity" with its own provider.
- Sponsored placements
- Paid sponsor placements. Detected from real single-advertiser ad units in the response stream. Excludes ChatGPT’s organic shopping cards (which are unpaid product carousels). US/AU/NZ/CA only as of May 2026.
- Sources / response
- Distinct URLs the model consulted while answering — the URLs that show up in its retrieval log, regardless of whether it cited them inline. Computed only over web-search responses (otherwise the answer is from training data, no sources to count).
- Citations / response
- Subset of sources that the model rendered as inline footnote pills in the answer. A source becomes a citation when the model explicitly references it.
- Fanouts / response
- Number of sub-queries the model spun up internally to answer the prompt (e.g. "best hotels Paris" might fan out to "hotels Paris Marais", "luxury hotels Paris", "rooftop hotels Paris", …).
- Map entities / response
- Hotel cards inside the map widget. Each carries a provider (Google Places, TripAdvisor, Yelp, Foursquare, SERP) and a place ID where applicable.
- OTA
- Online travel agency (Booking, Expedia, Hotels.com, Agoda, Trip.com, Priceline, etc.) — commission-based booking sites.
- Direct
- A hotel’s own website. Computed at request time by matching the cited domain against Google Places.
- Chain
- A hotel chain’s brand website (hilton.com, marriott.com, ihg.com, …). 175+ chain-domain rules.
- Editorial
- Travel media (Condé Nast Traveler, Time Out, Lonely Planet, Forbes Travel Guide, NYT, etc.) and city-specific travel blogs (santorinidave.com, theurbanlist.com, etc.).
- Directory
- Multi-property hotel consortia / brand collectives (Small Luxury Hotels of the World, Design Hotels, Preferred Hotels, Virtuoso, …). Aggregate hotels under one umbrella but aren’t a single OTA.
- Review
- TripAdvisor, Oyster, Yelp, Trustpilot, Holidaycheck — review-first platforms.
Does ChatGPT send people to Reddit and YouTube?
Share of answers citing at least one Reddit thread or YouTube video, week by week. Worldwide.
Share of answers citing at least one Reddit thread or YouTube video. Counted from the citation domains, so this measures what ChatGPT showed as a source — not whether it read them.
Latest week (2026-10-05): 1 of 616 answers cited Reddit (0.2%), 0 cited YouTube, across 2,435 citations.