Product Questions Gain Share in ChatGPT's U.S. Work Chats

ChatGPT’s U.S. work chats devote a larger share of messages to product information, OpenAI’s October 26 data release shows. B2B marketers have a reason to test how the AI explains what they sell.

Messages seeking information about goods and services rose from 1.69% of U.S. work chats in early 2025 to 2.92% in the third quarter of 2026. A marketing team can use the finding to examine the product questions its own website answers, before buying another service to track appearances in AI answers.

The new release connects ChatGPT messages with 165 work activities across nearly 100 countries. Our reading of the U.S. rows in the published CSV archive finds a growing share devoted to obtaining product or service information. The same rows put developing marketing and promotional materials at roughly 5% in both periods.

Product marketers can investigate how well their information serves the person asking about a product. That question sits alongside how efficiently AI helps their own team produce material about it.

Product information has a precise meaning here

The data dictionary maps messages to standardized work activities. The category behind the increase concerns obtaining information about goods or services. It doesn’t identify the sender’s occupation, establish that they’re shopping, or record a purchase.

A message classified in that category can therefore tell a marketer about the kind of work being brought to ChatGPT. It can’t identify a new lead for the sales team.

Our recent Attest coverage examined a researcher correcting an interpretation of shoppers’ decisions. These message records answer a different question: what activities appear in work-related conversations on individual ChatGPT accounts? Keeping that question intact prevents a usage finding from becoming an unsupported sales claim.

The U.S. increase was 1.23 percentage points between the first and last quarters. Those endpoints also cover changes in the people using ChatGPT and the other activities they bring to it. The data doesn’t follow a fixed group of buyers through a purchasing decision.

Price questions need a careful label

A separate activity involves determining the value or price of goods and services. Its share of U.S. work messages rose from 0.26% to 0.69% over the same period. The latest figure was slightly below the preceding quarter’s 0.70%.

That category includes more than someone asking what a product costs. Setting a selling price or valuing an item can fit it too. The dataset doesn’t isolate customer price comparisons from those other tasks.

For a software marketer, this still supplies a useful prompt for a small test. Ask ChatGPT to explain the product’s commercial terms, then compare the answer with the current pricing page. Check whether it distinguishes the advertised starting price from the conditions attached to it. Treat any error as something to investigate, rather than assuming the model found or ignored a particular page.

This is a suggested check, not an experiment reported by OpenAI. The usage data doesn’t establish that changing a pricing page improves sales.

Individual accounts leave part of work unseen

The Signals individual-data page specifies Free, Go, Plus, and Pro accounts. Enterprise conversations and Codex use are excluded, so the finding doesn’t represent all business uses. A work-related conversation on a personal account doesn’t establish that the person’s employer has adopted ChatGPT.

The global sample contains 1 million messages each month, from July 2024 through September 2026. That total is larger than the U.S. work-message subset used here. The public rows don’t supply its separate sample count or uncertainty intervals.

Accounts must have a declared adult age. People who deleted their accounts or opted out of training are excluded. Automated systems classify de-identified messages, and the published figures include weights and added noise to protect privacy. Smaller groups are less precise.

OpenAI also sells the platform whose usage it measures. Its download page makes the underlying tables available, allowing readers to inspect the categories rather than relying on a broad claim about AI adoption.

Use the finding to choose a question

Start with a product question that already appears in a sales conversation. For a small software company, that might be whether a particular integration is included in the plan being considered. Have someone responsible for the product check the answer against current documentation.

Record what was missing before commissioning more promotional content. The distinction matters because producing marketing materials held about the same share of these work messages while product-information requests gained share. Flat share alone says nothing about the absolute volume or value of content creation.

Keep website results alongside this check. Our coverage of Kevin Indig’s organic-search comparison explains why traffic losses need comparable sites and reporting periods. Signals contains no matching Google-search history, so it can’t explain where a site’s missing visits went.

For the next AI content and brand discussion, bring one checked product answer and the source that supports it. That gives the team a specific information problem to fix before treating a larger share of product questions as proof of new customers.