The Maze: ChatGPT serves different jobs in different countries. A September 2026 comparison shared by Erik Sarissky, an international marketing and localization specialist at Ahrefs, puts writing at 35.4% of messages in the Philippines and 35.1% in Nigeria, against 21.4% globally. These are shares of messages, not shares of people. The selected-country comparison points to a problem for businesses building with AI: translating the same experience may leave the wrong task at its centre.
Writing takes a much larger share in the Philippines and Nigeria. The comparison puts both roughly 14 percentage points above the global writing share. That means writing accounts for more of the observed conversation mix; it does not mean a third of residents use ChatGPT to write. Türkiye sits at 15.8%, the lowest writing share among the countries shown. A single global task mix would flatten this substantial difference.
Other countries give more space to different tasks. Self-expression accounts for 17.4% of Japanese messages, nearly twice the global 8.9%. Information seeking reaches 25.3% in Spain, 25.0% in France and 24.7% in Poland, against 20.2% globally. Images and media reach 12.9% in India and 11.5% in Brazil, versus 8.8% worldwide. These are several distinct patterns, not one divide between countries that use AI more or less seriously.
The global average conceals competing definitions of a useful answer. Practical guidance is the largest global category shown at 28.1%; writing follows at 21.4%, then information seeking at 20.2%. Technical help represents 4.6%. The US is close to the global baseline on writing at 21.5% and identical on information seeking at 20.2%, while practical guidance is higher at 30.2%. That makes a US task mix a plausible starting hypothesis for some tasks, but it does not make it a proxy for every country. The much larger writing shares in the Philippines and Nigeria would be missed by that shortcut. Similarity on two categories also does not establish similarity across the whole conversation mix. For a business designing an assistant, the task changes the test: a writing request needs a usable draft, while an information request needs a clear, dependable answer. This is an operating implication of the differences, not evidence that any particular local product design already works better.
The sample describes individual use, not every business opportunity. OpenAI's Signals data covers sampled messages on individual Free, Go, Plus and Pro plans through September 2026. The comparison concerns adults aged 18 and over and excludes Enterprise, Edu and Codex use. Business and technical work can therefore be underrepresented. The six displayed categories do not exhaust all messages, and a message-based measure gives frequent message senders more weight than occasional ones. It cannot establish population adoption, cultural causes or shopping conversion.
Why it matters: AI localization has a task problem as well as a language problem. A retailer considering an assistant should treat these differences as hypotheses for local testing: do customers need help composing, finding an answer or completing a practical task? The evidence does not identify retail demand by country. It does show why importing a global prompt menu and changing its language is an incomplete starting point. Measure whether the intended local task gets completed before treating usage as commercial value.
