Trust in AI hotel answersCross-checked across sources — but mostly anchored on Google Maps
TL;DR: AI assistants are pretty good at hotels. They don’t just trust the first website they see — they cross-reference websites, OTAs, reviews, third-party listings. That’s genuinely better than older search. But sometimes they still hallucinate (a Florence query that came back with Mexico City hotels), and they can be deceived (we built Hotel Ranque, a fictional Paris boutique, and got it cited in ChatGPT, Gemini, and Perplexity). The common thread under both: every AI assistant we’ve looked at ends up anchoring heavily on Google Maps. And Google Maps is itself easy to add to. Still work to do — on the assistants, and on the underlying data.
The trust picture, in three pieces
AIs cross-reference. They sometimes hallucinate. They can be deceived. All three share the same anchor.
The good news first: modern AI assistants are no longer trusting a single website to answer a hotel question. ChatGPT pulls from 5+ providers and runs multiple parallel searches (see the anatomy article). Claude calls Google Places and reads reviews verbatim (see how Claude searches hotels). Gemini, Perplexity, and Grok all do their own variants of cross-referencing. That’s genuinely an improvement over older keyword search.
The mixed news: even with cross-referencing, edges still slip through. In one week of scraping we saw real hotels confidently surfaced for the wrong city — a Florence query returning Mexico City hotels, a London query returning New York SoHo hotels. These are illustrations, not statistically validated patterns; one weekly snapshot.
The uncomfortable news: we built Hotel Ranque, a fictional Paris boutique with no rooms, no staff, no booking system — just a structured website — and it got cited by ChatGPT, Gemini, and Perplexity for narrow prompts. One small experiment, not a generalisation. But it shows that the cross-referencing has a floor.
The thread linking all three: the cross-references mostly bottom out on Google Maps. And Google Maps is itself surprisingly easy to add to (see how dirty Google Maps hotel data is — 17% of listings fail basic checks). A trust chain is only as strong as its anchor. Still work to do.
Real hotels showing up for the wrong city
Even with cross-referencing, AIs sometimes still mix things up. Of the 9,509 AI hotel mentions we collected this week, 1,549 didn’t match a hotel in the city the user asked about. Of the most-mentioned 500 of those, around 91 turned out to be real hotels — just in a different city or country. We’re not claiming this is a stable rate; it’s one weekly snapshot, and the matching pipeline has its own noise. But the examples below are striking enough on their own.
A few examples
Florence (IT) → Mexico City (MX)
Asked about Florence, the assistant returned three different Mexico City hotels — Gran Hotel Ciudad de México, Círculo Mexicano, Downtown Mexico. Why is hard to say from one example: maybe the model associated “Florence” with Spanish-language content, maybe a name collided. We just noticed it.
Buenos Aires (AR) → Palermo, Sicily (IT)
Buenos Aires queries returned three Sicilian hotels — Grand Hotel et Des Palmes, Grand Hotel Wagner, Quintocanto Hotel & Spa — all from Palermo, Italy. Buenos Aires has a trendy neighborhood called Palermo, so a name collision feels plausible. Whether the assistant does this consistently we don’t know yet.
London (GB) → New York SoHo (US)
London queries returned The Mercer, 11 Howard, and The Manner — three NYC SoHo boutique hotels. None has a London property. Reads like the kind of mix-up you get when “luxury boutique hotel” signals get clustered together regardless of the city in the prompt.
More of the wrong-city examples
| Mentions | Prompt City | Hotel Name (returned) | Actually located in |
|---|---|---|---|
| 6 | Shanghai (CN) | The Middle House | Mayfield (GB) |
| 5 | Florence (IT) | Gran Hotel Ciudad de México | Mexico City (MX) |
| 4 | London (GB) | The Mercer | New York (US) |
| 4 | Buenos Aires (AR) | Grand Hotel et Des Palmes | Palermo (IT) |
| 4 | Florence (IT) | Hotel Spadari al Duomo | Milan (IT) |
| 3 | Buenos Aires (AR) | Grand Hotel Wagner | Palermo (IT) |
Worth noting
None of these are catastrophic failures — the hotels are real, just in the wrong place. But it does show that cross-referencing isn’t the same as verification. The assistant has multiple sources agreeing the hotel exists; it just doesn’t have anything saying “this hotel is in the city you asked about.”
A fictional hotel that got cited
Hotel Ranque is a boutique hotel in Paris’s Bastille neighborhood with a chess club, a cycling lab, a specialty coffee corner, and a yoga studio. It also doesn’t exist. We built it as a one-off experiment to see whether a fictional hotel with a well-structured website could surface in AI assistants for a few narrow prompts. For these specific prompts, the answer was yes in ChatGPT, Gemini, and Perplexity — one example, not a generalisation about how AI ranking works.
The prompts we ran
Where Hotel Ranque appeared
| Platform | Mentions | Trigger prompt |
|---|---|---|
| ChatGPT | 1 | best boutique hotel with specialty coffee near Paris Bastille |
| Gemini | 3 | multiple Bastille / Ledru Rollin / yoga prompts |
| Perplexity | 2 | find me hotels for chess yoga cycling in paris near ledru rollin / coffee bastille |
What the site has
Hotel Ranque has three things AI assistants seem to treat as signal in this case:
- Schema.org/Hotel markup with address, geo-coordinates, amenityFeature, and FAQPage on every page.
- Topical depth: dedicated pages for each amenity (chess, cycling, coffee, yoga) with FAQ sections that mirror plausible search queries.
- Internal coherence: every page links back to the same canonical address, every FAQ answers a real-feeling question.
What it doesn’t have: a Google Business Profile with verified reviews, a TripAdvisor listing, a Booking.com page, or any third-party validation. For these particular prompts, none of the assistants checked.
One observation
The interesting bit isn’t that AI cites a fake hotel — it’s that the cross-referencing chain accepted it. The structured website + a Google Maps listing + a few plausible third-party signals were enough. None of the assistants visited the property. They believed Google Maps. Which means the trust in AI hotel answers is, ultimately, trust in Google Maps.
The common thread: Google Maps
Both stories above — the wrong-city mix-ups and the Hotel Ranque experiment — share an anchor. Every AI assistant we’ve looked at, when it eventually has to answer “does this hotel exist and where is it,” bottoms out on Google Maps. Not exclusively, but predominantly:
- ChatGPT: ~89% of hotel entity cards come from Google Places via SerpAPI (see the anatomy article and the 100K-entity study).
- Claude: in nine captures we ran, eight of them called
places_search(Google Places) directly — see how Claude searches hotels. - Gemini, Google AI Mode: Google’s own ground truth. Hotel clicks in AI Mode go 79% to Google Business Profiles (study).
- Perplexity, Mistral: web-search-based, but a lot of the surfaced content originates with Google Places listings.
That’s a reasonable strategy — Google Maps has the most complete hotel index in the world. But it has its own data quality problems. We scanned 178,647 Google Maps hotel listings across 11 countries and found 17% fail basic checks: 8,167 OYO vacation rentals listed as hotels, Belgium losing 54% of listings after cleaning, and so on. And getting onto Google Maps is much easier than people think — Hotel Ranque has a Google Maps listing.
Why this matters
The trust chain in AI hotel answers is: multiple sources cross-referenced → mostly resolving to Google Maps → Google Maps trusts whoever fills out the form correctly. Each link is reasonable; stacked together, the chain is only as solid as the anchor. Improvements at the assistant level (better verification, place-ID gates, freshness checks) help. Improvements at the Google Maps level help more.
If you run a hotel
- Get your Google Maps listing right. Address, geo-coordinates, photos, hours, reviews. That’s where the cross-references converge, so that’s where you compete.
- Build the third-party trail too. OTAs, TripAdvisor, review platforms. The cross-referencing is real — the more places agree your hotel exists at this address, the harder it is for an AI to get it wrong.
- Publish structured data on your own site. Schema.org/Hotel with address and geo-coordinates. Cheap, useful, no downside.
How we collected this
This is a snapshot, not a study. The numbers below describe one week of scraping and a one-off experiment. Treat them as context for the examples, not as base rates.
The weekly scrape
We run weekly automated scrapes of six AI assistants — ChatGPT, Perplexity, Gemini, Google AI Mode, Bing Copilot, and Grok — via Bright Data’s SERP and AI scraping API. The current run executes a library of 616 prompts across 56 city/country combinations. Hotel names are extracted from each response.
Matching names to hotels
Each name is matched against a 195,585-row Google Places lodging index with a five-tier fuzzy pipeline (exact normalized match → token-sort → token-set → partial with distinctive-token filter → cached aliases). Names that don’t match in the prompted city are bucketed as unresolved.
Finding the wrong-city examples
For the most-mentioned 500 unresolved names, we ILIKE-match against the global index without a city filter. A match in a different city or country becomes a candidate example. We then eyeballed the top results to pick the ones in this article. None of this is rerun stability-checked; we’d need several weeks of data to say which examples recur and which were one-off.
Caveats worth knowing
- The Google Places index is incomplete in resort and emerging markets (Punta Cana, Maldives, Phuket). Some “unresolved” names are real hotels in the right city — the index just doesn’t have them yet.
- ILIKE matching is conservative; stylised names with diacritics or punctuation differences slip through.
- The other-categories table is hand-built from anecdote, not measurement. The severity column is a gut call, not a number.
- Single-week sample. We don’t know how many of these examples would recur next week.
Frequently asked questions
Summarize with AI
