In short: ChatGPT Ads has three buying models: Reach, where you pay per thousand impressions, Clicks, where you pay per click, and Conversions, where the system optimizes toward your conversions. For most advertisers the right order is to start with Clicks, install the pixel in parallel and switch to Conversions once the data is in place. This guide covers the choice of model, what a sensible test budget looks like, who actually sees the ads, and the creative choices that decide delivery. The US reference numbers: Daily budgets from around 25 dollars, click bids of 3 to 5 dollars and CPMs of 25 to 60 dollars.
The three buying models, and when they fit
When you create a campaign in Ads Manager, the objective is the first choice, and it determines how you pay:
| Model | You pay for | Fits when |
|---|---|---|
| Reach (CPM) | A thousand impressions | The goal is awareness and presence in the conversations, and you have budget for frequency |
| Clicks (CPC) | Clicks to the landing page | The goal is qualified traffic, and you want to learn the channel with controlled risk |
| Conversions (oCPC) | Clicks, optimized toward conversion | The pixel is installed and has collected enough conversions for the system to learn |
Our recommendation for almost everyone: Start with Clicks. Reach is hard to evaluate without solid brand measurement, and Conversions without data is just Clicks with extra steps. The sensible path is to run Clicks for the first weeks, set up the OpenAI pixel from day one, and switch the campaign objective to Conversions once conversions are actually coming in. Then the system has real signals to optimize against, and you have your own numbers to compare with.
Budget: What is sensible in an early market?
Local price data barely exists yet in the newly opened European markets, so the reference is the US: Daily budgets have started at around 25 dollars, OpenAI recommends click bids of 3 to 5 dollars, and observed CPM levels have ranged from 25 to 60 dollars. Click bids below roughly 3 dollars trigger a warning that the ad may not deliver, so that is the floor in practice. Three budget rules on top of the numbers:
- Set a test share, not a bet. 10 to 15 percent of the digital budget is enough to learn the channel. Scale when your own numbers say it delivers, not when the enthusiasm does.
- Give the channel a learning period. Experienced advertisers treat the first 60 to 90 days as a learning investment, with enough budget for the numbers to mean something. If you conclude after two weeks, you conclude on noise.
- Calculate cost per conversion, not per click. Results vary widely between industries: Some report conversion rates beating Google Ads, others weaker numbers than search traffic. The only number that decides is what a customer costs you in this channel versus the others.
And remember the context: Competition is still low, and early prices in new channels have historically been favorable. The window resembles search advertising around 2005, and it closes as more budgets move in.
Who actually sees your ads?
Ads are shown only to signed-in adult users on the free tier and the Go subscription. Plus, Pro, Business, Enterprise and Edu are ad-free. That has a strategic consequence which is easy to overlook: Part of your audience never sees the ads. If you sell to developers or executives who happily pay for Plus and Pro, the ads only reach the part of your audience on the free version. For consumer products and broad B2B the coverage is good, for narrow premium segments expectations should be adjusted. Ask yourself before budgeting: Do my customers use ChatGPT, and if so, on which plan?
Also remember that at the European launch the ads are not personalized: Targeting is based on the conversation's topic, approximate location, device, time and language. That makes context hints your most important lever, and that is where the optimization hours should go.
The creative choices that decide
The format is tight: Company name and logo from the account settings, a title of up to 50 characters, a description of up to 100 characters, a square image of at least 256 by 256 pixels and a landing page. There are also product ads pulling price and reviews from a product feed. Within those constraints, four choices separate those who deliver from those who do not:
- Write as a helpful answer. The ad sits below an answer from an advisor, and the conversation's rules apply: "Cut reporting from three days to three hours" beats "Transform your business", every time.
- Build variants for coverage. Several titles and descriptions per offer, each taking a new angle, give the system more chances to find relevant conversations. Repetitions in other words do not count.
- Send traffic to the right page. A relevant product or topic page beats the front page, and UTM parameters on the link give you your own numbers in GA4 independent of the platform's reports.
- Keep the landing page technically open. OpenAI requires that the page does not block OAI-AdsBot and OAI-SearchBot. If you have followed the guide on technical SEO for AI search, this is already in order.
The structure that makes the numbers readable
Buying model and budget only work when the structure underneath is clean: One need per ad group, with its own context hints and its own ad variants. Then you can see which needs deliver and move budget there, instead of staring at an average of three audiences in the same group. Start narrow, with one product and one clear goal, and expand once you know what works. It is the same advice that applies to the whole channel: Learn cheaply now, while competition is low.
Where to go next
The full setup from account to published campaign is in the step by step guide, the craft behind the targeting in the context hints guide, and the measurement that makes the Conversions objective possible in the pixel guide. The big picture, with all the AI surfaces and the interplay between paid and organic visibility, lives in the pillar guide on AI advertising. And do not forget the other half of the equation: The ad buys you a place next to the AI answer, but what the answer actually recommends is built organically.