Why ChatGPT Ads Mostly Reach People Who Aren't Working
OpenAI's Signals v3.0 dataset, released October 26, 2026, shows work-related ChatGPT requests fell from 51.6% to 32.2% of consumer usage between July 2024 and September 2026. On the Free plan, where ChatGPT ads actually run, only 24.7% of requests are work-related. B2B advertisers buying ChatGPT ads are mostly bidding into personal conversations.
The headline most people will share is the writing number. PPC Land's breakdown of the Signals files shows writing dropped from 34.1% of consumer requests to 21.4%, and technical help collapsed from 13.5% to 4.6%. Practical guidance (28.1%) and seeking information (20.2%) now lead. Interesting, sure. But I think the more useful number for anyone with an ad budget is buried in the subscription tier table, and it changes who you should be sending to ChatGPT ads at all.
The ad inventory is the least worky part of ChatGPT
ChatGPT ads only serve on the Free and Go plans. Plus, Pro, and Enterprise stay ad-free, which OpenAI repeated when it expanded ads to 89 countries on October 21 and said more than 100,000 advertisers had run campaigns. Now line that up against the work share by plan for September 2026:
- Free: 24.7% work (down from 44.0% in September 2024)
- Go: 42.4% work (down from 55.3% at launch in August 2025)
- Plus: 50.8% work
- Pro: 68.1% work, and it actually went up
So the plans with ads are the plans where people are mostly not working, and the plans where people are mostly working are the ones you can't reach. Roughly three in four Free requests are personal: recipes, trip planning, relationship stuff, homework-ish things, and health questions (which don't get ads anyway).
Think of it like buying a billboard at the train station on a Saturday. Same station, same foot traffic numbers in the media kit, very different people walking past.
One caveat that matters a lot here: these are shares of requests, not users, and the files don't publish plan sizes. OpenAI reported more than 50 million paying consumer subscribers in February out of 900 million weekly users, so Free is almost certainly the overwhelming majority of the audience. A lot of work requests still happen on Free in absolute terms. They're just a minority of what you're bidding into.
Where did all the work go? Mostly somewhere you can't buy
This part surprised me less than it probably should have. The Signals sample covers 1,000,000 consumer messages a month from adult accounts, and it excludes enterprise messages, Codex, and anyone who opted out of training. Technical help dropping from 13.5% to 4.6% doesn't mean people stopped asking AI for code. It probably means they moved to Codex, to company-paid seats, or to other tools entirely, and none of that shows up in this file.
Which is kind of the point for advertisers. The work usage that left the consumer Free tier went to surfaces OpenAI has explicitly kept ad-free. The professional with a company ChatGPT Enterprise seat is, from what I can tell, the exact person a SaaS marketer wants, and they're structurally out of reach.
The trend line is also older than this dataset. The original "How People Use ChatGPT" NBER paper from Chatterji and colleagues already showed non-work messages outgrowing work messages, and Duke's Fuqua summary put non-work usage at 53% rising to 73% between June 2024 and June 2025. Signals v3.0 just extends the same curve, and splits it by plan in a way that maps neatly onto ad eligibility.
The country data points the same direction. In Q3 2024, 44 of 137 countries had work as the majority of requests. By Q3 2026, none did. Singapore is the most work-heavy market at 42.9%.
A quick note on how noisy the recent months are
I'd be a little careful with the exact August 2026 lows (writing at 20.6%, work share at 29.5%). OpenAI adds differential privacy noise, and it tightened that noise over time: epsilon of 5 for December 2025 and earlier, then 3, then 2, then 1 for July through September 2026. Lower epsilon means more noise. With a million messages a month the topline shares should hold up fine, but I wouldn't build a deck around a one-month dip. The 27-month direction is the signal. The monthly wiggle probably isn't.
Also, a language model classifies every request, using the same prompts as the 2025 paper. That's consistent, which is good. It's also a model deciding what counts as "work," and plenty of real work (drafting a LinkedIn post at 10pm, say) can look personal to a classifier.
Splitting your ChatGPT tests by offer intent
None of this means B2B brands should stay out. It means the default "ChatGPT users are knowledge workers" assumption is wrong for the inventory you can actually buy, and your tests should be set up to prove or disprove that quickly.
How I'd run it:
- Separate consumer-intent and work-intent offers into different campaigns. Practical guidance and information seeking are now nearly half of all requests. Offers that answer a personal question (a calculator, a household budget template, a product that fits a life problem) match the inventory. A demo request for procurement software mostly doesn't.
- Judge B2B campaigns on qualified pipeline, not CTR. A click from a personal chat can look great in platform reporting and then never reply to a sales email. Set a simple bar: if ChatGPT leads convert to sales-qualified at less than half your paid search rate after 30 days, cut spend or move the budget to a consumer-intent angle.
- Lean on markets with higher work share if you're B2B. If your product sells in Singapore or similar work-heavy markets on the 89-country list, test there before a broad rollout.
- Write creative for someone who isn't at their desk. Assume a phone, an evening, and a person asking about their own life. "For your team" copy is probably fighting the context.
On paper, filtering for work context would solve all of this. OpenAI doesn't offer that kind of targeting, and given how ads get selected based on the conversation (we looked at how the orchestrator can override the auction last week), you have less control over context than you'd have with a keyword list. Another reason to let the offer do the filtering.
And to be fair, this was always a bit predictable. The paid tiers were sold partly on the promise of no ads, and the people most willing to pay for a tool are usually the ones using it to earn money. It just feels a lot more concrete with a 24.7% figure attached.
Where I think the Free tier goes from here
My guess: Free-tier work share drops below 20% in the first Signals release that covers mid-2027, as more professional use moves into company seats and Codex. If I'm right, expect OpenAI to pitch ChatGPT ads harder as a consumer and commerce channel (product feeds, shopping, personal finance) and less as a place to find business buyers. The Delta buy through Amazon DSP already looks like that kind of advertiser.
I don't think that makes ChatGPT ads a bad channel. It makes it a different channel than the one in a lot of the early pitch decks, and it probably rewards the advertisers who sell things people want on a Tuesday night, more than the ones hoping to catch a VP between meetings.