Reading OpenAI’s Usage Data for Travel and Hotels
OpenAI publishes a running dataset of how people actually use ChatGPT: a sample of 1,000,000 messages a month, from July 2024 through September 2026, classified automatically into a handful of broad categories. We read all 24 files in the October 2026 release and recomputed the travel-relevant numbers ourselves, rather than restating OpenAI’s own write-up.
Travel and hotels never appear as their own category in it. What follows reads the closest proxies the release has: the two topic buckets that hold most trip research, and the handful of O*NET work-activity codes that cover shopping, price comparison and advising. The direction of each finding is solid. The exact percentages are not travel bookings; they are a general labor taxonomy’s codes, applied to messages that happen to look like travel shopping.
What this actually measures
OpenAI’s classifier sorts every sampled message into one of seven coarse topics: Writing, Practical Guidance, Seeking information, Technical help, Multimedia, Self-expression, and Other/Unknown. Two of those would hold most travel research. Practical Guidance covers how-to advice, tutoring, creative ideation, and health, fitness, beauty, or self-care; Seeking information covers specific facts, purchasable products, and cooking and recipes. “Where should we eat in Lisbon” and “what should I pack for a October trip to Japan” both land in one of these two buckets. So does “explain the offside rule,” which has nothing to do with travel. The two buckets are a proxy for search-shaped questions in general, not a travel count.
For work-related messages, OpenAI also classifies each one against the 332 Intermediate Work Activities (IWAs) from the US Department of Labor’s O*NET taxonomy, a general-purpose list built to describe jobs, not AI chat. None of the 332 names a travel or hospitality activity directly. The closest four are “Obtain information about goods or services,” “Determine values or prices of goods or services,” “Purchase goods or services,” and “Provide information to guests, clients, or customers.” We use those because they are the nearest available signal, not because they were built to track travel shopping. A message about comparing flight prices and a message about comparing laptop prices score identically here.
The data itself carries two more limits worth stating up front: differentially private noise is added to every cell, and any cell built from fewer than 100 effective messages (1,000 for the country-by-quarter work-activity tables) is withheld entirely rather than shown with a large error bar. Full detail in Method.
Shopping and price-comparison activity is the fastest-growing thing in the work data
Among US work-related messages, “Obtain information about goods or services” grew from 1.7% to 2.9% of the total between Q1 2025 and Q3 2026 (+73%), and “Determine values or prices of goods or services” grew from 0.3% to 0.7% (+163%). “Purchase goods or services” grew faster still in relative terms, from 0.02% to 0.10% (+352%), though the base is small enough that this one is best read as a direction, not a reliable rate.
Every tourism market we checked shows the same direction, most of them faster than the US
We pulled the same two work activities for 11 more tourism-dependent economies. “Obtain info about goods/services” grew by an average of 67% across the 12 markets, and “Determine prices” by an average of 143%, both well above what a flat market would show. Greece leads on the first measure (+127%), Mexico and Japan tie for the lead on the second (+216%).
A literal hospitality signal, concentrated in tourism economies
One O*NET activity is, by its own description, front-desk and concierge work: “Provide information to guests, clients, or customers.” In Greece, where tourism is a large share of the economy, this activity went from 0.07% to 0.31% of work messages between Q1 2025 and Q3 2026, the highest level of any of the 12 countries we checked in either quarter and a genuine climb, not just growth off a zero base. The United Arab Emirates went from an unmeasurable level to 0.16%. Several other markets (the United Kingdom, France, Thailand, Portugal) show the same activity appearing from a near-zero or zero starting point, which we read as early presence rather than a reliable growth rate.
Advising on products and services is growing even where the guest signal is flat
“Advise others on products or services,” the activity closest to what a travel advisor or concierge does when recommending rather than just answering, grew in 11 of the 12 markets. Spain leads (+328%), followed by Mexico (+248%) and Japan (+222%). Greece, which leads on the guest-facing measure, is nearly flat here (+2%), already near the top of the range in Q1 2025.
A top ChatGPT market is not necessarily a search-heavy one
Overall ChatGPT usage per capita and the search-like share of that usage are different measurements. The UAE ranks 10th of roughly 150 countries for per-capita ChatGPT use in Q3 2026, but only 45.2% of its messages are search-like, the second-lowest share of the 12 markets here. Greece ranks 40th for overall use, mid-table, but has the highest search-like share of the group at 56.3%. Spain (rank 32, 55.3%) and France (rank 22, 53.9%) show the same pattern: high search-share without leading the per-capita table.
Bars in violet (Greece, UAE) mark the highest and one of the lowest search-like shares in the group, to make the point above easier to spot on the chart.
The full panel
| Country | Rank (Q3 '26) | Search-like share | Obtain info growth | Price growth | Guests, Q3 '26 | Advise growth |
|---|---|---|---|---|---|---|
| United States | 49 | 50.4% | +73% | +163% | 0.05% | +211% |
| United Kingdom | 41 | 51.2% | +75% | +118% | 0.05% | +134% |
| France | 22 | 53.9% | +87% | +123% | 0.06% | +104% |
| Spain | 32 | 55.3% | +67% | +170% | 0.06% | +328% |
| Italy | 88 | 51.7% | +48% | +162% | 0.08% | +24% |
| Germany | 19 | 51.2% | +64% | +126% | 0.03% | +59% |
| Japan | 62 | 48.2% | +65% | +216% | 0.02% | +222% |
| Thailand | 106 | 40.8% | +19% | +85% | 0.02% | +47% |
| Mexico | 96 | 46.8% | +20% | +216% | 0.07% | +248% |
| UAE | 10 | 45.2% | +79% | +69% | 0.16% | +40% |
| Portugal | 31 | 51.9% | +85% | +155% | 0.10% | +22% |
| Greece | 40 | 56.3% | +127% | +108% | 0.31% | +2% |
How we did it
- Source. OpenAI Signals v3.0, downloaded from openai.com/signals/data in October 2026: 24 CSV files plus a README, covering July 2024 through September 2026. 1,000,000 ChatGPT consumer messages are sampled each month, from accounts declaring an age of 18 or older, excluding enterprise and Codex traffic. Licensed CC BY 4.0.
- Recomputation. Every number in this article was recomputed directly from the released CSVs, not copied from OpenAI’s own write-up. We summed the two coarse topics (Seeking information, Practical Guidance) for the search-like share, and read specific O*NET IWA codes from the country-by-quarter work-activity table for the shopping, pricing, and hospitality figures.
- Country selection. 12 countries with meaningful inbound or outbound tourism: the United States, United Kingdom, France, Spain, Italy, Germany, Japan, Thailand, Mexico, the UAE, Portugal, and Greece. This is a judgment sample, not the full ~150 countries in the release, chosen to represent a spread of tourism-dependent economies rather than the largest ChatGPT markets as such.
- Privacy and suppression. OpenAI adds calibrated differentially private noise to every released cell (ε=1, δ=1/N per the README’s v3.0 budget) and withholds any cell built from fewer than 100 effective (weighted, post-noise) messages, or 1,000 for the country-by-quarter work-activity table. A handful of our 12-country cells came back at exactly 0.0, which we treat as “at or near the suppression floor,” not literally zero messages.
- Limits. No travel-specific classifier exists in this release; every travel-adjacent number here is read off general-purpose topic and labor-activity codes that happen to catch travel-shaped language. Percentages built on a pre-growth base under 0.05% of messages (several of the hospitality and purchase figures) are volatile by construction and reported as direction, with the underlying levels shown alongside.
Frequently Asked Questions
Summarize with AI
