How to launch and optimize ChatGPT Ads
If you run paid media, you've probably had three different people send you the OpenAI Ads Manager announcement in the last few weeks. The headlines all say roughly the same…
If you run paid media, you've probably had three different people send you the OpenAI Ads Manager announcement in the last few weeks. The headlines all say roughly the same thing: ChatGPT now has self-serve ads, it's open to U.S. businesses, CPC bidding is live, the minimum spend is gone.
What the headlines don't tell you is whether you should actually go run a campaign tomorrow.
Here's the short answer: maybe! But not the way you'd run a Google or Meta campaign. Almost everything about ChatGPT Ads is structurally different from what paid media folks are used to, and going in with Google Ads instincts is a great way to spend $3,000 on nothing.
What's actually new
OpenAI's self-serve Ads Manager went live in beta on May 5. The cost barrier dropped dramatically from where it started: the original February pilot wanted $250,000 minimums and CPMs around $60. By April, the minimum was $50,000. Now there's no minimum at all, and CPCs are landing in the $3–$5 range with CPMs reportedly closer to $25.
Translation: any business that's willing to learn the platform can show up. The price of admission is now competence, not budget.
OpenAI says it's chasing $2.5 billion in ad revenue this year. They hit $100 million annualized in the first six weeks of the original pilot. The rapid price drop isn't them panicking — it's the usual playbook for getting enough auction participants in the door to make the marketplace work. The window where being early is genuinely cheaper than being late is open. It won't stay open forever.
The mental shift: no keywords. Stop looking for keywords.
ChatGPT Ads doesn't have keyword targeting. Instead, advertisers write context hints — short descriptions of the kinds of questions or situations a user might bring to ChatGPT that would make your offer relevant. The platform matches via semantic similarity, not exact text.
Practitioners testing the beta have flagged a counterintuitive rule: in Google Ads, adding more keywords expands coverage. In ChatGPT Ads, adding more context hints actually dilutes match precision, because the underlying embeddings start to overlap and step on each other. Fewer, more precise hints outperform broader ones.
If you've spent a decade building 800-keyword Google Ads accounts, that hurts a little. But it's the single most important adjustment to make.
A practical exercise: take five of your strongest Google Ads keywords, then rewrite each one as a one-paragraph description of the situation a user is in when they'd type that keyword. That's a context hint. The keyword is the search; the hint is the human moment behind the search.
Targeting: less than you'd hope, more than you'd think
Here's what you can control today:
- Geography — country, plus U.S. state, DMA, and ZIP code as of late May
- Audience tier — Free or Go users (Plus, Pro, Business, Enterprise users don't see ads)
- Context hints — your intent descriptions
- Daily or lifetime budgets — added late May
- Bidding model — CPC or CPM
- Creative format — sponsored placement below the relevant ChatGPT response
Here's what you can't control:
- Audience age, gender, or interests
- Specific keywords or phrases
- Day-parting (not yet)
- Negative targeting
- The call-to-action label on your own ad — OpenAI picks it automatically based on your creative
- Anything resembling individual user data (chats stay inside ChatGPT, advertisers get aggregated reporting only)
So: less granular than Meta, more privacy-respecting than Google, and you're essentially betting on the quality of your context hints to do most of the targeting work.
One important constraint: eligible advertiser categories are still narrow. Household goods, consumer goods, local services, travel, entertainment, digital products, and education are in. Health, mental health, politics, and most regulated categories are out. If you're a healthcare brand, this isn't your moment yet.
Who should test now, who should wait
A reasonable test framework:
Worth testing: B2B SaaS with research-heavy buying cycles. Professional services. Education and online courses. High-consideration ecommerce where buyers actually compare before they purchase. Anything where customers normally Google for hours before deciding.
The reason these work: ChatGPT users aren't doom-scrolling. They're typing questions and reading answers. They're already in research mode. If your buyer's journey starts with "let me ask a question first," ChatGPT is the closest thing yet to advertising directly inside that question.
Not worth testing yet: Local services that depend on phone calls (no robust local intent signals). Impulse ecommerce. Anything that needs precise demographic targeting. Regulated categories. Businesses whose CFO wants every dollar attributed three ways from Sunday.
Realistic budget for a test: $2,000–$3,000 over 14–21 days is enough to generate meaningful click volume and learn what works without committing real money to a platform whose benchmarks are still being written.
The measurement situation, told straight
OpenAI now has a pixel (the SDK is called OAIQ), a Conversions API, and reporting on impressions, clicks, spend, CTR, CPC, CPM, and conversions. The infrastructure looks like what you'd expect from a modern ad platform.
In practice, it's closer to CTV or podcast advertising than to Google Ads. You can prove a click happened. You can technically prove a conversion happened. What you cannot yet do, with the confidence Google and Meta have earned over fifteen years of measurement infrastructure, is prove incremental revenue, lifetime value attribution, or media mix contribution.
The honest framing: this is currently a brand-awareness channel with a performance price tag. Budget it accordingly. If you can carve out 5–10% of your test budget for learning a new surface, this is exactly the moment to do it. If everything has to be fully attributed before it gets greenlit, sit this one out for six months and check back.
What we'd love OpenAI to add next
A short wishlist, since they're clearly iterating fast:
- Negative targeting. Even a basic exclusion list for context hints would make creative work way more efficient. Right now you can describe what you want, but not what you don't.
- Audience exclusion for converters. Letting advertisers suppress ads to people who've already converted would dramatically improve performance for the long-cycle B2B advertisers most likely to test the platform.
- Real day-parting. Conversation patterns probably differ between Tuesday morning and Saturday night. Advertisers should be able to bid accordingly.
- Advertiser-controlled CTAs. OpenAI picks them automatically right now. Letting brands at least choose from a list would meaningfully change creative testing.
- Category expansion with reasonable guardrails. Health, financial services, and legal are massive paid-media verticals. They'll come eventually; the speed at which they do will tell us a lot about how seriously OpenAI wants to be a top-3 ad platform.
- Third-party measurement partnerships. Native pixel data is fine. Mature measurement is partnerships with the MMM and incrementality vendors advertisers already trust.
The bottom line
ChatGPT Ads is real, the door is open, and the cost to learn the platform is finally reasonable. The skill that compounds — the one that will still matter in 18 months when everyone else shows up — is learning to write context hints that describe user intent with surgical clarity.
That skill doesn't come from a webinar. It comes from running a few hundred dollars through the platform and watching what gets matched to what.
If you want a second set of eyes before launching a test, contact us. We're happy to think through whether your business is one of the ones that should be early on this channel, or one that should wait.