Is ChatGPT really shifting "from asking to doing"? OpenAI's own CSVs say questions are still half of all messages — and the shopping-question category grew

Where this stands on September 26, 2026: the same bundle, and the table the first version skipped

We re-downloaded the Signals CSV bundle on 2026-09-26. The download page serves the file published on August 6 (the server's last-modified stamp is 2026-08-06 16:44 UTC), every table still ends at June 2026, and every figure quoted below re-read identically from the fresh files: the topic shares, the Seeking-information rise from July 2024 to June 2026, and the asking/doing/expressing split. The short answer to the question in the title has not moved. At work, doing leads. Outside work, asking leads doing two to one, and question-shaped messages are half of all use. What the first version of this post did not use is the bundle's topic-by-work table, which gives the topic mix inside non-work messages on their own, and the work/non-work split itself. Both sharpen the consumer reading.

Inside non-work use, the two asking-shaped categories are 57.0% of messages in June 2026 (Practical Guidance 35.0% plus Seeking information 22.0%), up from 53.3% in July 2024. Seeking information on its own is flat to slightly down within non-work use (23.1% to 22.0%), so the all-message rise from 16.7% to 18.8% quoted below comes from the other table in the bundle: non-work messages grew from 48.7% of all messages in July 2024 to 69.6% in June 2026. The consumer pool, where Seeking information runs at 22%, now carries seven of every ten messages, and it is the pool that grew. A check on the fresh files: weighting each group's topic mix by its share of all messages reproduces the published all-message shares to the third decimal (Seeking information 0.187 against 0.188 published, Practical Guidance 0.321 against 0.321). Every row above is in the committed cut file for this post, read 2026-09-26.

What OpenAI published, and what we did with it

OpenAI's August 6 post is the first country-by-country release of ChatGPT usage data, covering what OpenAI says is "more than 1 billion people" on individual plans (Free, Go, Plus, Pro). The headline framing is a move "beyond answers and into action." The post links the Signals data-download page, which publishes the underlying tables as CSVs under CC BY 4.0 with a data dictionary. We downloaded the refreshed bundle on 2026-08-08 and read two tables: message share by topic, and message share by asking/doing/expressing split by work vs non-work. Every number below is OpenAI's own, from those files — monthly through June 2026.

The topic table: question-shaped use is 50.9% and the biggest slice

Practical Guidance plus Seeking information — the two asking-shaped categories, and the two where "who should I hire?" and "what should I buy?" live — total 50.9% of all consumer messages. And the trend runs against the headline: Seeking information rose from 16.7% in July 2024 to 18.8% in June 2026 in OpenAI's own monthly table, while Practical Guidance has held near a third (32.2% in March, 32.1% in June). The category explicitly defined to include `purchasable_products` queries is growing, on a user base OpenAI counts in ten figures — so the absolute volume of buy-and-hire questions is growing even faster than the share suggests.

"From asking to doing" is a work-side story

The shift OpenAI describes is real where it says it is: at work, doing (45.4%) beats asking (32.3%), matching the post's "more than twice as likely" claim for task-completion at work vs outside it. But outside work — the mode a consumer is in when they ask for a dentist, a lawyer, or an accounting tool — asking is still the largest category at 45.1%, double doing's 22.0%. Asking's non-work share has drifted down from 53.6% two years ago, yet nothing has replaced it at the top; the growth went to expressing (32.9%), not to task delegation.

What this means if you care about AI visibility

• The recommendation surface is growing, not fading. If you read "from asking to doing" as "the era of asking ChatGPT for recommendations is ending," the CSVs behind the headline refute it: the information-seeking share rose over two years and question-shaped use is half of everything. The demand side keeps compounding — 45% of US consumers have already asked AI for a local business recommendation, and OpenAI's own tables say that behavior class is expanding.
• Consumer and work usage are different games; optimize for the consumer mode. Buyer questions live in the non-work half, where asking dominates. A business that shows up in answers to `purchasable_products`-class questions is fishing where 45.1% of non-work messages are — the single biggest pool OpenAI measures.
• This is a rare fully-auditable vendor claim — audit it. The post's framing and its data ship separately, and the data is CC-BY-downloadable. The same check applies to your own AI visibility: the claim "AI recommends businesses like yours" is testable, per engine, per question, in 60 seconds.

Limits

• These are OpenAI's classifiers on OpenAI's data. Topic and asking/doing/expressing are automated message classifications (Signals v2.0 methodology); no external party can re-derive them from raw messages.
• Individual plans only. The tables cover Free, Go, Plus and Pro accounts — not ChatGPT Work/enterprise — so the work-side numbers describe personal accounts used at work.
• Shares, not volumes. OpenAI publishes shares by month; we quote the share trend and note the >1B user base, but no absolute message counts exist in the release, and we computed none.
• Latest month is June 2026, still, on a 2026-09-26 re-read. The August 6 post is the release date, not the data endpoint; seven weeks later the download page serves the same August 6 file and no later month has been added.

The practical reading: OpenAI's own ledger says the question channel — the one where engines name specific businesses — is the largest and still-growing slice of consumer AI use. Whether the engines name your business in it is the part you can check: run the four-engine measurement on your business. If you want that number tracked over time rather than read once, we compared the AI-visibility trackers on pricing model, commitment and how they read the engines; the honest summary is that most of them price for a brand team, not for a business asking a one-off question.