Skip to content
← Insights
Nov 14, 20264 min read

Your top ChatGPT use case depends on what you count

Message shares and survey responses measure different behaviour. The original usage study and newer 2026 data show why population, unit, and observation period belong beside every percentage.

Three differently sized measuring containers beside a screen with simple unlabeled conversation blocks, with a magnifying glass and blank calendar.

The most common ChatGPT task depends on whether you count messages, people, or the time they spend. Combining those measures produces a ranking with no consistent meaning.

A person might ask dozens of short questions while editing one document over a much longer session. Message count can make the questions look dominant. A survey asking whether that person uses AI for writing would count writing too. Both observations can be accurate while answering different questions.

That is why I would replace a triangulated top-ten list with a smaller set of comparisons whose denominators remain visible. Different studies can inform the same discussion without being combined into one league table.

Start with what the conversation study measured

OpenAI's 2025 account of how people use ChatGPT describes research based on about 1.5 million consumer conversations. In that study, guidance, information seeking, and writing together accounted for nearly four-fifths of the classified conversation topics. Those are findings from a defined sample and period, not a live query board for every ChatGPT product.

The study also used an activity classification that distinguished asking for information from producing an output or expressing something. That is a different axis from topic. A writing conversation can involve a request for advice, a draft, or a reaction to an existing paragraph.

Mixing those axes can create duplicate entries. A top-ten list containing writing, summarisation, and asking may place the same interaction in more than one position. The categories need to be mutually interpretable before their shares can be compared.

The unit also changes how a minority task looks. Coding can represent a modest share of broad consumer usage while being central to a particular developer's day. Neither observation is a contradiction. The consumer population contains many people who have no reason to write software.

Keep the newer data attached to its method

OpenAI now provides Signals data with information about individual ChatGPT usage, including purpose and conversation topics. Its October 2026 update presents overall usage based on September 2026 messages. It covers individual plans and excludes enterprise and Codex usage, so it can underrepresent business or technical work.

A separate 2026 adoption analysis follows a sampled cohort of newer accounts over a defined observation period. It examines how usage broadens after someone starts using the product. A cohort describing those new users cannot automatically represent every long-standing account in every later month.

This is an important limit on comparisons over time. A larger share of one task can reflect a change in how existing people use the product, a different mix of new users, or a classification change. The published method is needed to identify which comparison the data supports.

The same care applies to an activity count. More messages do not directly establish more economic value or better task completion. Someone repeatedly repairing an unsatisfactory answer can generate more messages than someone who receives an adequate response immediately.

For a product team, that difference is useful. A high-volume category can identify a behaviour worth examining. It leaves a separate question about whether the conversation accomplished the user's goal.

A survey tells you about the people asked

Pew Research Center's 2026 U.S. research on AI asks people about their use and views. That is a different method from classifying messages in a platform log. The geography and the survey's question wording remain part of its interpretation.

A percentage of adults who have ever used a chatbot cannot be read as the percentage of chatbot messages devoted to a task. A person can select several uses in a survey, while an automated conversation classifier may assign a single main category. Adding the results together would hide the different counting rules.

The useful comparison is narrower. A survey can show which groups report adopting a tool. A conversation study can describe what appeared in the sampled interactions. Together they can suggest where to investigate further, provided neither is used as a substitute for the other.

I would make a product decision using the population the product is intended to serve. For a writing tool, observe whether the intended users can complete a revision and whether they accept the result. For a support tool, examine the actual support task. A global usage ranking supplies context, but it cannot define success for either product.

A good usage claim can fit in one explicit sentence: whose behaviour was counted, what unit was counted, and during which period. Once those details are present, the statistic becomes easier to use. If they are missing, a precise-looking percentage may only make an unclear comparison harder to question.

AIResearch methods