OpenAI has published its first country-by-country view of how people use ChatGPT. The most useful finding is not which country moved up a ranking. It is that work use is shifting from asking for an answer to producing an output.
Across the countries in the release, people using ChatGPT at work were more than twice as likely to be “doing”—writing, coding, editing, or analyzing—than people using it outside work. That is a meaningful change in ChatGPT country usage, but the dataset needs careful handling. It covers individual Free, Go, Plus, and Pro accounts. It does not cover ChatGPT Business, Enterprise, or Education.
In other words, this is a large window into individual behavior, not a census of every company deploying ChatGPT. Used that way, the data is genuinely useful.
The work split matters more than the country leaderboard
OpenAI divides messages into broad intent groups. “Asking” covers information and clarification. “Doing” covers the production of an output or completion of a task. At work, doing dominates. Outside work, asking remains the largest category.
That distinction gives builders a better adoption measure than weekly active users. Opening ChatGPT does not tell you whether it changed the job. A completed draft, analysis, code change, image, or decision aid does.
| Signal | Verified finding | Do not infer |
|---|---|---|
| Work intent | At work, doing is more than twice as likely as outside work. | That the output was correct, accepted, or valuable. |
| Country rank | Per-capita adoption ranks changed between Q1 and Q2 2026. | That a higher rank means more total users or more economic impact. |
| Multimedia | 7.8% of classified messages globally were multimedia in Q2. | That every multimedia message generated an image or video. |
| Age | The share of messages from self-reported users aged 35+ rose in nearly every country. | That the sample represents everyone over 35 in that country. |
The practical move is simple: if you are measuring an AI rollout, count useful outputs and the human corrections they require. Our guide to using ChatGPT effectively makes the same point at the individual level: a clear deliverable beats a longer prompt.
Multimedia is now roughly one message in thirteen
OpenAI says multimedia generation, analysis, and retrieval reached 7.8% of classified messages globally in Q2 2026. Put another way, that is about one in every 12.8 messages. In Brazil and Colombia, the share exceeded one in ten.
The category is broader than image generation. It can include making media, analyzing an uploaded image, or retrieving information from multimedia. That makes the trend more important for product teams: the interface is becoming a mixed-media workspace rather than only a text box with an image button.
For a builder, this changes the failure modes. Text workflows need citation and reasoning checks. Multimedia workflows also need provenance, consent, rights, visual accuracy, and a way to preserve disclosure when a file leaves the original app. The lesson from the Bold Glamour disclosure problem applies here too: a label attached inside one platform may disappear when the media travels.
The over-35 shift is real, but the denominator is narrow
Among users who self-reported their age, the share of messages from people aged 35 and older increased in almost every country. Globally, their share was five percentage points higher than a year earlier. France and Czechia each recorded increases of more than ten percentage points.
This is not proof that five percent more adults adopted ChatGPT. It is a change in the share of age-classified messages, and age is self-reported. OpenAI’s chart uses 111 countries with complete estimates and compares each country with its own Q2 2025 baseline.
Builder implication: stop designing every AI workflow for an early-adopter stereotype. Test onboarding, terminology, error recovery, and trust cues with the people who will actually do the job.
The age result is a product-design signal, not a demographic victory lap. A workflow that works for a technical enthusiast may still fail a manager, teacher, clinician, shop owner, or parent who approaches the tool with a different vocabulary and a lower tolerance for unexplained errors.
The adoption gap is closing—but rank is not reach
OpenAI compared messages-per-capita ranks across 144 countries from Q1 to Q2 2026. Peru, Uruguay, and Costa Rica rose the most in the global ranking, while parts of Latin America, Oceania, and Africa gained faster than established early-adopter regions.
Rank movement is useful for spotting momentum. It is poor evidence for market size. A country can climb because its own usage rose, because another country slowed, or because the eligible population and classification changed. The public release does not turn a ranking change into a customer-acquisition forecast.
If you are deciding where to localize a product, combine the Signals trend with your own activation, retention, support, payment, and task-completion data. Country enthusiasm cannot tell you whether your product solves a local problem.
Four decisions I would make from this data
- Measure outputs, not chats. Track completed artifacts, acceptance, corrections, and time-to-finish. A busy chat can still be a failed workflow.
- Design multimedia as a full workflow. Add rights, provenance, export, and review controls instead of treating media as decoration around text.
- Test beyond early adopters. Recruit users across age, role, and confidence levels. Watch where they hesitate and what explanations they need.
- Localize around tasks. Translate the job, examples, support, and trust model—not only the interface strings.
Plan choice should follow the same discipline. Our ChatGPT Go versus Plus comparison focuses on what a person actually needs to complete, while our model selection framework keeps the decision tied to failure cost rather than popularity.
My verdict: this is an operational shift
The headline is not that more countries and age groups are using ChatGPT. It is that individual users are increasingly asking it to produce something that leaves the chat and enters a real workflow.
That raises the standard. A useful AI product must do more than attract messages. It needs to help people finish a job, preserve context, catch errors, and hand the result to the next person or system without losing accountability.
Read the primary record
- Read OpenAI’s country-by-country ChatGPT usage analysis.
- Explore the OpenAI Signals data hub and its downloadable releases.
What useful output would you measure if chat volume disappeared from your dashboard tomorrow?
Checked August 9, 2026. Country, intent, multimedia, and age findings come from OpenAI’s primary release. The one-in-12.8 figure is Musthave.ai’s calculation: 100 divided by the reported 7.8% global multimedia share. Limitations are stated alongside the findings.
