I’ve spent the past two days reading about an ad format I cannot yet buy.
Ads started as a US pilot in February 2026, expanded through Canada, Australia and New Zealand in March, and then through the UK, Mexico, Brazil, Japan and South Korea over the middle of the year. Late August added 31 European markets.
They’re not running in Singapore, Malaysia, Indonesia, Thailand, Vietnam, the Philippines, Hong Kong or Taiwan… not yet anyway.
So a good chunk of us in this region are in an unusual position: we can watch a new ad channel be built, in markets close enough to ours to be legible, before we’re asked to have an opinion about it. That doesn’t happen often, but hey, at least we can use this time to start prepping.
What to know about ChatGPT ads
Who sees the ads? They appear on the Free and Go tiers only, while Plus, Pro, Business, Enterprise and Education users won’t see any ads. Free-tier users can opt out of ads in exchange for fewer daily messages, and nobody under 18 will see your ads.
How do ad gets matched to a person? Three signals: what they’re talking about in that conversation, what they’ve talked about before, and which ads they’ve clicked in the past. In old money, that’s contextual targeting plus a light behavioural profile — much closer to how search works than how Meta works. There is no interest picker, no lookalike audience, no uploaded customer list. You’re buying a topic and a moment, not a person.
What you get back. Aggregate numbers only — how many people saw it, how many clicked. No advertiser sees any of the conversation itself. If you’re used to Meta’s breakdowns, this will feel like buying media in 2009, and that’s a genuine planning constraint rather than a teething problem.
Where ads won’t appear at all. Health, mental health and politics are excluded. If the conversation is on any of those, no ad runs.
Then, two weeks ago, India. Free and Go tiers, roughly 50 brands to start, an ads manager opening to marketers next month, and WPP and Omnicom brought in on the rollout. India is over 100 million weekly ChatGPT users, overwhelmingly on Free and Go. That’s the market that matters here, and not only because of its size.
India, a market to watch
Just last week, ads went live on Free and Go in India, starting with around 50 brands.
Next month OpenAI opens an ads manager there. That’s the self-serve platform where you build and buy campaigns yourself — the equivalent of Meta Ads Manager or Google Ads. Until one exists in a market, buying isn’t really open: it’s a conversation with a salesperson and a contract. So the ads manager, not the first ad, is the moment a format actually arrives for the rest of us.
And the scale: India is over 100 million weekly ChatGPT users, overwhelmingly on Free and Go — which means almost all of them are ad-eligible. In most Western markets the ad-supported audience is what's left after the paid tiers take the valuable users. In India it's close to the entire market.
That’s why I’d watch India rather than America. It’s the nearest thing to a preview of how this behaves in a market shaped like APAC.
3 things I’d actually do with the head start
None of these need budget or a platform login.
Find out whether your category can even be advertised against. Ads are excluded from health, mental health and politics topics. If you work in regulated financial services, that exclusion list is the single most important document about this format, and it’s going to get longer, not shorter, as it hits markets with stricter ad codes. Worth knowing now whether the interesting half of your funnel is off-limits before you build a plan around it.
Watch the Indian rollout like it’s a case study, because it is one. Which categories show up. Whether the ad sits under the answer or inside it. Whether anyone complains. What the local regulators say. You will get more from three months of watching India than from any launch-day webinar.
Ask what happens to your organic mentions when the space above them gets sold. This is the one I’d genuinely prioritise. Everyone in AEO / GEO / whatever we’re calling it this quarter has spent a year optimising to be mentioned in the answer. Introducing paid placement into that surface changes the value of the organic mention — maybe upward, if paid validates the channel, maybe downward, if it starts looking like every other auction. Nobody knows yet. But you should know how many of your inbound conversations currently start with “ChatGPT said,” and right now most marketers I ask have no idea.
Why this might all be nothing
Here’s the honest counterweight, and it’s stronger than I’d like.
Perplexity killed advertising entirely in February this year. They’d been testing sponsored placements since 2024 and pulled the whole thing. Their stated reason was trust — a user has to believe the answer is the best possible answer, and a commercial placement makes that belief negotiable. Worth noting the financial context: ads were reportedly around $20,000 of revenue in 2024. It’s easy to walk away from a rounding error, and I don’t think Perplexity’s decision proves the format can’t work.
But the trust argument survives the cynicism. The estimated click-through rate on ChatGPT ads is somewhere in the low single digits versus something like 29% of Google searches producing a click to an external site — and that estimate is floating around industry analyses without a hard source, so hold it loosely. The direction is the point. People don’t come to a chat interface to leave it. If the format’s whole promise is intent quality rather than click volume, then it lives or dies on measurement, and measurement in beta is currently a weekly CSV of impressions and clicks.
So the fair summary is: an unproven format, with weak measurement, on a surface where clicking is the exception, in a channel whose main asset is the trust that the ads erode. That can still be enormous. Search looked stupid too. But anyone selling you certainty about this in September 2026 is selling you something.
Is this the future of LLMs?
We spent the last year being told AI would break how people find things, and we responded by trying to get mentioned in the answers.
Now the answer has a price attached, and the region I work in is going to get it later than most, with less coverage in our languages and fewer local benchmarks. In practice that means when it does land in Southeast Asia, someone senior will ask “what’s our ChatGPT ads strategy” on about day four, and the honest answer will have to be assembled from a standing start.
The AI;DR
Elsewhere in the AIverse
OpenAI ships GPT-6 Astra, and says "AGI". GPT-6 Astra landed on 3 September, rolling out to Plus, Pro and Enterprise, with the API and AWS. President Greg Brockman called it “the start of the AGI era”. ARC-AGI-3 went from 7.8% on the previous model to 99.9% with tools. The model is being rolled out in phases to tackle any cyber security issues, but should be generally available in the coming days.
Google puts image editing inside Workspace. Google Pics is an image generation and editing tool built on Nano Banana, sitting inside Workspace and working directly in Docs and Slides. You can edit individual elements and text in an image rather than regenerating the whole thing. It’s live on Business Standard and Plus, Enterprise Standard and Plus, AI Pro and AI Ultra.
Meta returns to the frontier with Muse Spark 1.3. Zuckerberg is back in the game with a new model that wants to put Meta on par with frontier offerings from Anthropic and OpenAI. His own line was “frontier performance almost too cheap to meter”. It’s still in safety testing, but the version developers can actually use scores 61 on Artificial Analysis’s Intelligence Index, behind Claude Fable 5.1 at 66.