The short answer
ChatGPT ads are worth testing if your customers research purchases inside an assistant, your margins survive a premium CPM, and you have the measurement to tell whether anything came back. They are not a replacement for being the source the assistant cites, which remains the more durable asset and the cheaper one.
The timeline matters for context. OpenAI began showing ads to logged-in adults on the free and Go tiers in the United States on 9 February 2026, opened a self-serve Ads Manager in May 2026, and extended availability to more countries including the United Kingdom through August. The market is roughly seven months old.
How buying works, and why it is not Google Ads
The biggest adjustment is that you are not bidding on keywords. Advertisers supply natural-language context hints describing the kinds of conversations where the ad belongs, alongside more familiar phrase signals, and the system matches against the conversation rather than against an exact query string. Relevance is weighted in the auction, so the highest bid does not simply win.
That has two implications. First, your targeting is now a piece of writing, and it is worth the same care as ad copy. Describe the customer's situation, not your product category. Second, you have less control over placement than you are used to, which is exactly the trade you accept when the surface is a conversation rather than a results page.
Ads appear as labelled placements around the assistant's answer rather than inside it. OpenAI's stated position is that advertising does not change the substance of the response.
What it costs and what that implies
Pricing has moved quickly. Early access opened at around a 60 dollar CPM with a six-figure minimum commitment; by the time self-serve buying opened the minimum had gone and reported CPMs sat closer to 25 dollars. Both numbers will be out of date soon, which is itself the useful signal: this is an inventory market in its price discovery phase, not a settled channel with benchmarks.
A 25 dollar CPM is expensive against display and cheap against high-intent search in competitive categories. Whether it works is entirely a function of intent quality, and intent quality in an assistant is genuinely high: people describe their situation in full sentences, with budget, constraints and timeline, because they are talking to something that reads paragraphs.
The catch is who sees it. Ads run on the free and Go tiers; Plus, Pro, Business, Enterprise and Education users see none. In B2B especially, the buyer with budget authority is disproportionately on a paid tier, which means the ad-supported audience skews away from your best prospect. Test accordingly.
The relevance problem
Independent tracking has found ads appearing on roughly a quarter of commercial prompts, with a noticeable share poorly matched to what was being asked. For advertisers that cuts both ways: wasted impressions when your ad shows against an unrelated question, and reputational noise when a category gets crowded with placements that read as intrusive.
It also creates an opening. Assistants are where people go for judgement rather than links, and an ad that reads like an interruption in that setting performs worse than one that reads like a useful next step. The creative that works looks less like a banner and more like the answer to the follow-up question the user was about to ask.
Ads versus citations, honestly compared
The strategic question is not whether to buy ChatGPT ads. It is how to split effort between renting attention and earning it.
| Buying ads | Earning citations | |
|---|---|---|
| Speed | Live in days | Weeks to months |
| Cost shape | Per impression, forever | Front-loaded, then compounding |
| Control | You choose the message | The assistant chooses what to quote |
| Credibility | Labelled as advertising | Reads as the assistant's own recommendation |
| Durability | Stops when spend stops | Survives as long as the page is the best answer |
Citations are the better asset and the slower one. The work behind them is the work described in generative engine optimisation and query fan-out: pages that answer the sub-questions an assistant actually searches, with evidence it can quote. Ads are what you buy while that work matures, or where the category moves faster than content can.
Measurement, which is the real obstacle
Attribution inside assistants is poor and will stay poor for a while. Referrals arrive with thin or missing referrer data, sessions often start on a different device from the conversation, and assistant-driven visits look deceptively small in analytics while influencing decisions upstream. We set out the practical workarounds in tracking AI referral traffic in GA4.
Before spending, put three things in place: a dedicated landing path so the traffic is identifiable, a "how did you hear about us" field on your enquiry form with an assistant option, and a holdout period so you can compare enquiry volume with and without the spend. Without those you will be reading noise, confidently.
Who should test this now
Consumer brands with clear, comparable products and healthy margins, in categories people research conversationally — home, travel, retail, personal electronics — have the most obvious fit, and those are the categories dominating early inventory. Local services can work where the assistant surfaces regional intent. Considered B2B purchases with long cycles are the hardest case, both because of the paid-tier skew and because attribution over a six-month cycle defeats a two-week test.
If you are already spending on paid search and finding performance drifting, the diagnosis in why paid campaigns underperform applies here unchanged: the destination page decides the outcome more often than the channel does. A new surface does not fix an old landing page.
Our advice to clients this quarter is unromantic. Run a small, measured test if you sell something people ask assistants about, keep the budget at a size whose failure teaches you something cheaply, and keep the majority of the effort on becoming the answer rather than the advertisement beside it. If you want help deciding which half of that applies to you, talk to us about your category.