Ask AI Twice, Get Different Hotels
We asked ChatGPT, Gemini and DeepSeek the same hotel questions three times. Only about 3 in 10 hotels made every list. Here is what that means for hotels.
How much do AI hotel answers change between runs?
Searching "best hotels in Paris" on Google twice gives the same results page, while AI assistants write a fresh answer each time. To see how much that changes the hotels they recommend, we asked ChatGPT, Gemini and DeepSeek the same 64 hotel questions three times each on 21 October 2026, and compared the lists.
How much the lists changed
| Assistant | Hotels from one run that came back in another run | Hotels named in all three runs |
|---|---|---|
| ChatGPT | 54% | 27% |
| Gemini | 53% | 29% |
| DeepSeek | 62% | 33% |
About half the hotels changed when ChatGPT or Gemini answered the same question again. DeepSeek was a little steadier, but even there only about one in three of all the hotels it named for a question appeared in all three runs, and on ChatGPT and Gemini it was closer to 28%.
What it looks like
Here is ChatGPT answering "romantic hotel in Paris for couples" three times:
- Run 1: Shangri-La Paris, Maison Souquet, Hôtel La Comtesse, Paris j'Adore, Esprit Saint Germain, Le Meurice
- Run 2: Shangri-La Paris, Paris j'Adore, Hôtel La Comtesse, Maison Souquet, Relais Christine
- Run 3: Shangri-La Paris, Paris j'Adore, Plaza Athénée, Relais Christine, Hôtel de Londres Eiffel
Shangri-La and Paris j'Adore appeared every time while the other hotels changed.
Gemini changed more on "where to stay in Jakarta for a business trip". Its first answer was a list of five-star hotels (The Ritz-Carlton Pacific Place, Fairmont, Hotel Mulia, Alila SCBD and The Westin), while its next two answers were mostly mid-range and budget hotels such as AONE Hotel and Grand Sahid Jaya, as if it had assumed a different kind of trip.
Why the answers change
An assistant creates each answer anew by searching, choosing a few pages and summarising them. If the search returns different results or it reads different pages, the shortlist changes, so there's no fixed first place for a hotel to hold.
A long tail of hotels
Across the 72 answers for each city, the assistants named roughly 140 to 240 different hotels, and about 35% to 48% of them came up only once. In each city, the five most-named hotels made up only 12% to 28% of all mentions.
No hotel owns a city in these answers, including the famous ones. A smaller hotel can therefore appear for the right question without being a household name.
What this means for hotels
- A single check tells you very little. If you asked ChatGPT about your city once and saw your hotel, it may be gone tomorrow, and if you didn't see it, it may appear next time.
- Measure how often you appear across many runs and many days. Your share of answers over time tells you more than one yes or no.
- Give assistants the same description of your hotel everywhere they look. If your website, booking listings and guides describe you consistently, an assistant has more reasons to choose you each time.
- Follow the questions your guests ask. The lists for "romantic", "business trip" and "family" questions were very different from each other, so a family resort learns more from family questions than from "best hotels".
How SearchSeal helps
SearchSeal asks your questions every day and keeps every answer, so you can follow your share of answers over time instead of relying on one snapshot. It also lists hotels repeatedly appearing instead of you and the sources returned with those answers. See AI visibility for hotels.
Method and limits
- 8 cities, 8 questions each, 3 runs, on ChatGPT, Gemini and DeepSeek: 576 answers on 21 October 2026.
- The first column is the share of one run's hotels that were named again in a second run of the same question by the same assistant. The second column is the hotels named in all three runs divided by every hotel named in any run. Different spellings of the same hotel were grouped under one hotel ID, which is published with the dataset together with a script that recomputes these numbers.
- ChatGPT and Gemini answers came from their apps, collected from the US. DeepSeek answers came from its API with web search (temperature 0.5), not from the DeepSeek app.
- All three runs shared one day, so variation over weeks is likely to be larger.
- Full method: Where ChatGPT, Gemini and DeepSeek find hotels.
Frequently asked questions
How much do AI hotel lists change when the same question is asked again?
About half the hotels changed when ChatGPT or Gemini answered the same question again. DeepSeek was a little steadier, but even there only about one in three of all the hotels it named for a question appeared in all three runs, and on ChatGPT and Gemini it was closer to 28%.
Why do the answers change?
An assistant creates each answer anew by searching, choosing a few pages and summarising them. If the search returns different results or it reads different pages, the shortlist changes, so there's no fixed first place for a hotel to hold.