OpenAI Signals: Information Seeking Accounts for 20.7% of Messages

By Elena Vasquez

Information seeking reached 20.7% of consumer ChatGPT messages in August, nearly matching writing at 20.6%, according to OpenAI Signals data released Oct. 26.

August was the first month in the series to put information seeking above writing; the series begins in July 2024. Writing returned ahead in September, at 21.4% compared with 20.2% for information seeking.

The September reversal also undercuts any claim that information seeking has permanently displaced writing across every month of the public consumer series since July 2024.

Takeaway: Lookups briefly led writing, but the monthly series does not show a permanent replacement.

What does Signals V3.0 measure?

OpenAI published Signals V3.0 and a separate post about how users bring work to ChatGPT.

The monthly dataset samples 1 million messages from account holders with a declared age of 18 or older on the Free, Go, Plus and Pro plans. It excludes Enterprise and Codex. An automated system classifies each sampled message by subject area.

According to the data dictionary, seeking information includes requests for facts, questions about products and services for sale, and recipe or cooking help. OpenAI combines those three subjects, so the release cannot show their separate shares.

The release cautions: “When citing percentages representing the use of ChatGPT, please reference the source and time frame.” Version numbers matter because each changelog entry adds a quarter and revises earlier months.

Takeaway: The sample covers consumer accounts, not Enterprise or Codex accounts.

How did information seeking and writing change?

OpenAI’s monthly series shows three large shifts:

• Writing fell from 34.1% of messages in July 2024 to 21.4% in September 2026.
• Seeking information rose from 16.7% to 20.2% over the same months.
• Practical guidance, including how-to advice and tutoring, remained the largest topic every month since March 2025 and reached 28.1% in September.

Work-related messages made up 32.2% of all messages in September. Within that group, information seeking stayed between 10.7% and 12.1% each month from July 2024 through July 2026. It then rose to 13.9% in August and 15.2% in September.

The category rose inside work-related prompts, but the release does not identify the occupations or industries behind it.

Takeaway: Information requests grew inside work use, even as writing lost share overall.

What does the work activity post add?

OpenAI’s work post organizes messages into 165 activities drawn from O*NET, a standard framework for tasks shared across jobs. It presents quarterly data from Q1 2025 through Q3 2026 across nearly 100 countries.

Across the 74 countries with data in every quarter, the 10 most common activities averaged 51.1% of classified work messages in Q3 2026, down from 56.0% in Q1 2025.

OpenAI described the change as “work-related use becoming less concentrated in the most common activities.”

One of those activities, Getting information, represented 11.6% of work messages in the United States in Q1 2025 and 9.9% in Q3 2026. These figures use a separate classification system and cover one country, so they do not measure the global seeking-information category.

Because those labels come from a separate country-specific classification, they should not be presented as a direct continuation of the global seeking-information series in this release.

Takeaway: Work activity broadened, but the release does not show more information-seeking behavior.

How does Signals compare with the 2025 paper?

A 2025 OpenAI-Harvard paper reported that seeking information grew from 14% to 24% during the year after July 2024, putting the category near one quarter of usage by mid-2025.

According to the data dictionary, Signals uses the same classifier prompts as that study and groups the same three request types under seeking information. It records 16.7% in July 2024 and 18.3% in June 2025, the paper’s final month. That produces an almost six-point gap at mid-2025.

Researchers selected approximately 1.1 million conversations, took one message from each conversation, and averaged each share over 28-day periods. Signals selects 1 million messages per month and revised earlier months to match that higher count.

The paper recorded writing falling from 36% to 24% during the same period, while Signals moved from 34.1% to 25.7%. The writing series shows a smaller mismatch.

That six-point gap matters because classification prompts align while sampling units, weighting periods, historical revisions, and information-seeking estimates differ, with no stated explanation in the release blog or data dictionary.

Takeaway: Treat the two datasets as related measurements, not interchangeable series.

What does this mean for sites and search teams?

Approximately one in five consumer ChatGPT messages now looks something up, roughly matching the writing share. For product-focused sites, the bucket includes product and service questions, but offers no clean measure of commercial intent. Recipe publishers cannot isolate their share because OpenAI combines facts, products and recipes.

The exclusions make these figures unsuitable for claims about enterprise adoption or developer workflows. Likewise, the U.S.-only activity series cannot stand in for the global consumer sample.

Most important for search teams, Signals tracks conversation topics but not how often ChatGPT accessed external websites or cited a site. It therefore cannot establish ChatGPT referral traffic, citation share, or which sites gained visibility.

The practical limit is clear: ChatGPT topic shares show what users ask, but not which external sources ChatGPT accessed, what replies cited, or what traffic any site receives.

Takeaway: Treat ChatGPT topic demand as an interest signal, not a traffic forecast.

What should readers monitor next?

OpenAI’s Signals changelog adds one quarter with each update and revises earlier months. A figure from v3.0 may differ from the same month in a later version.

Future releases may revise this same history again, so version labels and dated comparisons should remain attached to every percentage quoted from Signals in reports.

Analysts should track four unresolved details:

• Whether later releases separate facts, product questions and recipes.
• Whether external source access and citation data become available.
• Whether revisions change the information-seeking trend.
• Whether the top 10 activity share continues to fall across countries.

Takeaway: Monitor later Signals revisions, source reporting and finer category splits before treating ChatGPT information demand as a stable traffic signal.