# AI hotel answers dataset (October 2026)

576 answers from ChatGPT, Gemini and DeepSeek to hotel questions in 8 cities, collected by SearchSeal on 21 October 2026. This is the data behind our hotel research articles.

Licence: CC BY 4.0. You can use, share and adapt it for any purpose, including commercial use, if you credit "SearchSeal" with a link to https://searchseal.com/research/ai-hotel-answers-2026.

## What was asked

8 cities (Bali, Bangkok, Jakarta, London, New York, Paris, Singapore, Tokyo) x 8 questions x 3 runs x 3 assistants. The question types are best, boutique, luxury, budget, family, business, couples and one local-language question, for example "romantic hotel in Paris for couples" or "meilleur hôtel à Paris". The full wording is in answers.csv.

ChatGPT and Gemini answers were collected from their web apps from the United States. DeepSeek answers came from its API with web search on (model deepseek-flash, temperature 0.5). All runs were on one day.

## Files

| File | Rows | One row per |
|---|---|---|
| answers.csv | 576 | answer: assistant, city, question, run, hotels named, number of sources and links |
| hotel_mentions.csv | 5,156 | hotel named in an answer, with its position, hotel_id and chain or independent label |
| hotels.csv | 1,471 | hotel per city (hotel_id), with total mentions and mentions per assistant |
| metrics.py | | recomputes the figures in our articles from these files: `python3 metrics.py` |
| sources.csv | 3,945 | link attached to an answer, with its domain and type |
| domains.csv | 595 | domain and its type |

In sources.csv, "source" means a page the assistant cited or searched for the answer, and "link in answer" means a link shown to the reader inside the answer (for Gemini these are Google Travel hotel cards). DeepSeek's API returns sources only.

## How the labels were made

Hotel names and chain brands were pulled from each answer by a language model. Different spellings of one hotel were grouped under one hotel_id (city plus the distinctive words of its name), and each hotel gets the label it was given most often across all its mentions (ties count as chain). Domain types (booking site, editorial, hotel site, review site, other) were also assigned by a language model, with a handful of manual corrections. Treat the labels as good for totals and check any single hotel before quoting it.

## Not included

The full answer text is not included, and link URLs keep only scheme, host and path (plus the search term on Google links). Hotel names are as the assistant wrote them, so the same hotel can appear under slightly different names.