How people use ChatGPT in 2026: implications for product discovery and AEO/GEO

What OpenAI's ChatGPT usage report measures

OpenAI's October 26 report, ChatGPT's expanding world of work, adds work-activity classifications to its public Signals data. It lets researchers examine which tasks people bring to ChatGPT and how that mix changes over time.^[1]

For AEO, the question is how those tasks relate to the product information a customer needs. Someone asking whether a service supports a particular workflow needs more than a product name: they need the relevant capability, its commercial conditions and any limit that would rule it out. Our recommendations below apply that distinction to content planning and answer review. OpenAI did not test these recommendations or measure their effect on visibility.

Signals version 3.0 samples one million messages per month from July 2024 through September 2026. Eligible users are declared adults on Free, Go, Plus and Pro plans. Enterprise and Codex usage are excluded, as are deleted accounts and users who opted out of training. Automated classifiers label de-identified content; published aggregates use weighting and differential privacy, which adds noise to protect individuals.^[3]

Three boundaries govern the analysis:

• Messages are the unit. A message share is not a share of people or businesses.
• Share and volume differ. A topic can become a smaller share of a growing platform while its message count increases. The published shares do not supply the totals needed to calculate that count.
• The release does not identify search behavior. It does not report whether a message triggered web search, which sources were retrieved or cited, or whether an answer led to a purchase.

The official data download and methodology accompany the source notes. We recalculated the two charts from the same release.

Technical help fell as a share of consumer ChatGPT messages

Technical help accounted for 13.5% of eligible messages in July 2024 and 4.6% in September 2026: a decline of 8.9 percentage points. Seeking information rose from 16.7% to 20.2% over the same period.^[2]

Technical help includes mathematics and data analysis as well as programming. Seeking information includes many questions unrelated to products. The two series therefore cannot tell us that developers stopped using ChatGPT or that product research replaced coding.

They also cannot tell a software company to reduce investment in documentation. Signals excludes enterprise and Codex usage, and it does not measure the contribution of a documentation page to an answer. Decisions about a specific page still need evidence from the questions and problems your customers bring to you.

Product-information research grew within U.S. work messages

OpenAI classifies work messages using the Occupational Information Network (O*NET) taxonomy. An activity describes a task associated with a message, such as obtaining information about goods or services. It does not identify the sender's job title.^[1]

Among U.S. work-related messages, the share classified as obtaining information about goods or services increased from 1.69% to 2.92% between the first quarter of 2025 and the third quarter of 2026. The programming activity's share decreased by 4.29 percentage points. Pricing and purchasing activities also increased, from much smaller starting shares.^[2]

The increase concerns product information broadly. It does not isolate software evaluation, and a pricing activity may involve determining the value of something a business sells. These categories cannot be combined into a measured software-buying funnel.

The AEO opportunity we would investigate is coverage of customer evaluation questions: can someone establish what the product supports, under which plan, at what cost and with which restrictions? Your sales conversations, support requests and documentation searches supply the category-specific evidence that Signals lacks.

Broader work use calls for questions tied to real tasks

OpenAI also reports that work use became less concentrated in the most common activities. Across the 74 countries present in every quarter, the average share represented by each country's ten most common activities fell from 56.0% to 51.1%. This is an average across countries, rather than a global message-weighted share.^[1]

Our recommendation is to organize AEO questions around tasks customers need to accomplish. A generic prompt such as “best webhook service” leaves out the conditions that would make a recommendation useful. “Which webhook service lets our support team replay failed deliveries on a self-serve plan?” gives a reviewer something specific to evaluate.

The report does not establish demand for that question. A sales or support record would have to do that. The example shows why a list of broad category prompts can miss commercially important requirements even when it covers the right product category.

AEO content needs to explain when the product fits

Consider that webhook question. A customer wants support staff to replay failed deliveries without negotiating an enterprise contract. This is a hypothetical example, not a claim about a particular provider.

Suppose the provider's plan reference states that manual replay requires an enterprise contract. Compare these hypothetical answers:

• Misleading: “This service supports replay and offers self-serve pricing.”
• Accurate: “Manual replay requires an enterprise contract, so this service does not meet your requirement to use replay on a self-serve plan.”

Both answers could mention the same product. The first drops the condition that determines whether the customer can use it. The second preserves that condition and explains the exclusion. A citation only helps if the linked page supports the plan requirement, rather than merely confirming that replay exists.

Publish the facts that change the decision

Use the customer's question to identify the evidence the answer needs. For this example, the content audit would cover:

“Supports replay” is only one part of the answer. The plan, permissions, price and retention conditions determine whether the feature solves this customer's problem. That is the level at which we would assess product-information coverage.

A useful comparison page can explain the choice and link to these maintained references. It does not need to duplicate every limit. Keep the qualification beside the claim it modifies: a feature description should link to its plan entitlement, and a pricing explanation should identify the usage that incurs a charge.

This arrangement makes the information easier to inspect and maintain. It does not guarantee that ChatGPT will retrieve the page or cite it.

Let customer evidence set the prompt priorities

Build a prompt universe: the working collection of customer questions and prompt variants chosen for evaluation. Give each prompt candidate a source and a reason it matters. That could be a recurring support question, a requirement in a sales conversation or a documented objection to adoption.

Prioritize questions whose answers change product eligibility or expected cost. An incorrect plan requirement can rule out a suitable product; an omitted charge can create the wrong expectation. These are more consequential errors than a harmless variation in product wording.

Use Signals to frame the investigation, while your customer evidence determines which questions to test. Neither the 2.92% product-information share nor the 4.6% technical-help share supplies search volume for your market. They should not become weights in your prompt universe.

Our AI-search prompt strategy guide explains how to select a fixed set of prompts for repeat review without treating generated questions as observed demand.

Review answer accuracy before scoring visibility

For each selected question, write down the requirements a useful answer must preserve. In the replay example, those include the support-team workflow and the self-serve plan condition. Save the answer, date, model where disclosed, search setting and relevant conversation context so later reviews can compare the same question under recorded conditions.

Check three things:

1. Product fit: does the recommendation satisfy the stated requirements, or does it omit a condition that would disqualify the product?
2. Factual accuracy: are the feature, plan, price and limit statements consistent with the maintained product references?
3. Citation support: do the cited pages support the complete statements attributed to them, including their qualifications?

Report unresolved cases separately. If the customer needs older deliveries replayed but has not specified how old, clarify the retention requirement. If the inspected sources do not establish feature availability, mark it unresolved. Missing evidence does not establish that a capability is unavailable.

Different results need different content decisions

Compare the answer with the customer requirement and the published facts before assigning work:

• A required fact is unpublished or ambiguous. Verify it with the product owner, then publish the scoped explanation on the responsible page.
• The answer contradicts a clear published fact. Inspect its citations and the available search evidence. Correct conflicting or stale documentation where you find it; retain the mismatch for further investigation if the sources are already clear.
• The product does not meet the requirement. Accept a correct exclusion. Make the limitation easy to find and explain the supported alternative where one exists.
• The product appears eligible but is omitted. Record a visibility observation. An open-ended shortlist does not have to name every eligible product. Inspect the sources shown; the answer alone cannot establish whether the cause was indexing, source selection, stale context or generation.

After a content change, compare new answers with the saved results using the same questions and criteria. Report observed changes in accuracy, qualification and citation support. A better answer after an edit does not establish that the edit caused the improvement, and answer visibility remains separate from qualified inquiries or product adoption.

For the replay question, the review should tell the team whether the plan condition was preserved, which page supports it and whether a content change is needed. Report that alongside mention and citation counts so the numbers retain their meaning.

Methodology and source notes

The two quantitative figures were recalculated from Signals version 3.0, downloaded October 27, 2026. Topic values come from the monthly topic table. Work activities use the U.S. rows for the first quarter of 2025 and the third quarter of 2026 in the O*NET intermediate-work-activity table. That table assigns a most-applicable activity and applies publication thresholds to small groups.^[2]

All changes refer to shares of eligible messages. A percentage-point change subtracts the starting percentage from the ending percentage. For example, 4.6% minus 13.5% is −8.9 percentage points. The corresponding relative decline is about 66%; neither calculation supplies an absolute message count.

Chart downloads preserve additional decimal places for reproducibility, rather than to imply measurement certainty. Differential privacy, rounding, classifier error and eligibility restrictions remain limitations. The cross-country concentration finding is attributed to OpenAI's report. The product example, review diagram and AEO recommendations are our editorial analysis, rather than findings tested in Signals.

OpenAI Signals data is licensed under Creative Commons Attribution 4.0. The figures credit OpenAI and identify our selection and calculations.

Sources

1. OpenAI: ChatGPT's expanding world of work, October 26, 2026
2. OpenAI Signals v3.0: official data download, methodology and attribution
3. OpenAI Signals: current data dictionary, version 3.0