ChatGPT Ads: Setting Up Your First Campaign

OpenAI's advertising inventory has moved from speculation to a line item. In the markets where the Ads Manager beta has opened — Turkey among the more recent additions — you can create an account, add a payment method, and serve ads beneath chat answers within a day. That closes the argument media teams have been having since 2023 about whether AI interfaces would carry advertising. It replaces it with a duller and far more useful question: does this inventory deserve a line in your plan, and if so, how big a line?

What the beta opens, and what it still keeps shut

The constraints of the beta determine most of what you can do with it. The confirmed frame:

• Ads are shown only to users on the Free and Go tiers. Plus, Pro and Enterprise remain ad-free, which puts a slice of the highest-willingness-to-pay audience outside the inventory entirely.
• One placement: the area below the chat answer. There is no format embedded inside the response or dressed up as a citation.
• Targeting is contextual. No cookies, no personal-data audiences, no remarketing lists. Geography breaks down to country, region and city.
• Buying models are CPC and CPM, with conversion optimisation available alongside them.
• Measurement runs on the OpenAI Pixel and a Conversions API. Server-side transmission is on the table from day one.
• The floor is $25 per campaign per day. That is a technical minimum, not a recommendation.

What is missing matters as much as what is there. No historical performance data, no category benchmarks, no audience overlap reporting, and nowhere near the frequency granularity you are used to. The reflexes you built in Google Ads through the 2010s do not transfer cleanly. This behaves more like early Display, or Performance Max in its first year: you send signals, the system interprets them.

Who should test now, and who should wait

Contextual targeting plus a low daily floor narrows the winning profile considerably. Test now if you recognise your business here:

• Research-heavy categories. If people ask ChatGPT "how do I choose between X and Y" or "what should I look for in a provider", the space under the answer is a natural continuation. Software, education, insurance, healthcare services and B2B service purchases all fit.
• Short conversion cycles. For a 30-day test to say anything, the distance between click and conversion needs to be measured in days, not weeks.
• Advertisers spending above roughly $20,000 a month on media. The arithmetic is below; minimum spend takes a disproportionate bite out of smaller budgets.
• Low brand search volume against high category demand. If you need to capture new demand rather than harvest existing intent, a discovery surface is worth more to you than to an established name.

Wait if you are a local service business (city-level geography exists, neighbourhood does not, and impression density is unclear), an e-commerce operation with thin margins and high price sensitivity (you cannot model unit economics before CPC settles), a funnel that leans on remarketing (no cookies, no lower-funnel build), or a team in the middle of a critical season. Paying the learning cost outside peak trading is simply cheaper.

How much of your Google and Meta budget should you move?

The concrete number: 3-5% of combined monthly Google and Meta spend, over a 45-day window. Treat it as a ceiling, not a floor. The logic is that this is roughly the largest slice you can divert to a new channel's learning cost without visibly degrading the performance of the channels that pay the bills. Push to 10% and you will be explaining your quarter if the channel disappoints. Stay at 1% and you will learn nothing you can defend statistically.

At $25 per campaign per day, one campaign costs about $750 a month. A meaningful test needs at least two — different context clusters or different message angles — so the practical entry is around $1,500 a month. On a $30,000 monthly media budget that is 5% and entirely healthy. On $10,000 it becomes 15%, which is too much to commit to a channel that does not yet have a published benchmark. Below that threshold, run a single campaign against a narrow context set over a longer window, or wait a couple of quarters.

Where the money comes from matters too. Not from brand search. The most defensible source is usually broad upper-funnel prospecting on Meta, or discovery spend on Google's Display and Demand Gen side. Both do the same job: reaching someone who is not yet looking for you. Benchmark against those sets, not against the CPA of a brand campaign.

What 30 days can prove, and what it cannot

Writing this distinction down before the test starts prevents about half the arguments that follow it.

What you can read: the real range of CPC and CPM; which context clusters generate clicks; on-site behavioural quality via the pixel — session duration, page depth, bounce; micro-conversion volume such as form starts, pricing page views and demo requests; raw conversion count and raw CPA; the share of new users in the traffic; and city-level performance differences. With adequate volume, all of that is legible within 30 days.

What you cannot read: incrementality. You will not know how much of this traffic would have found you anyway — that needs a geo holdout, and 30 days is not enough. Cross-channel attribution: last-click will not show you how much a ChatGPT ad fed your branded search. Audience overlap: there is no data telling you whether you are paying twice for people Meta already reached. Creative fatigue curves: you need several months of series data to see a pattern. Category benchmarks: no shared CPC average exists yet, so your first month is your own benchmark. And the durable scale of the inventory itself, since the size of the Free and Go tiers and the density of impressions are both still moving.

Send that second list to your CFO or your client before the test, not after. Answering "how much extra revenue did this generate" with "we can't measure that yet" is a death sentence for a channel unless it was flagged in advance.

The order of operations for the first campaign

1. Build measurement first. Deploy the OpenAI Pixel site-wide, then add server-side transmission through the Conversions API. The pixel alone will undercount under browser restrictions. Validate the data flow at least three days before anything goes live.
2. Freeze your UTM scheme. Tag the source with a distinct label so you can separate this traffic from other search traffic in GA4 or your warehouse. Fixing it later costs you the first week of the test.
3. Run two campaigns against two hypotheses. One broad context set around category questions ("how to choose", "which is better"), one narrow set around high purchase intent. Do not split a single budget in half — give each the $25 floor.
4. Spend the first 10-14 days on CPC. Do not switch on conversion optimisation on day one; the system needs accumulated conversion signal. Move to a conversion goal once volume supports it.
5. Write the message against the answer context. The user has just received a response. What sits below it should read as the practical next step, not as a brand slogan.
6. Use a dedicated landing page. Send traffic to a page that extends the answer and asks for one action, not to a generic homepage. It is the only clean way to read this traffic's behaviour.
7. Review on day 14 and day 30. On day 14 look only at delivery and traffic quality; make no kill decision. On day 30 shift budget to the winning context cluster. Decide continue or stop on day 45.

What a realistic expectation looks like

Treat ChatGPT Ads as a third discovery surface sitting between search and social, not as something that reshapes the spine of your media plan. Contextual targeting feels like a step backwards to teams raised on precise audience construction — but given where privacy regulation is heading, learning to work inside that constraint was on the roadmap anyway. The real return on entering early is probably not cheap clicks. It is holding your own data before the rules of the channel settle. When competition arrives, the gap between a team with 45 days of real numbers and a team starting from zero is exactly where it shows.

And an honest caveat: in the first months of a beta, inventory, pricing and placement behaviour all move. Plan this as an experiment you revisit every two weeks, not as a structure you set up and forget for a quarter.