Running ChatGPT Ads for conversions: the oCPC optimization playbook
what oCPC needs before it works, how long to leave it alone, and the four things I changed that actually moved cost per conversion.
the short version: oCPC works well once it has data, and fails quietly before that. most of the skill is in what you do before launch and in resisting the urge to touch it afterwards.
What oCPC actually is
you still pay per click, but delivery is optimised toward a conversion event you define rather than toward clicks. so the system decides who to show your ad to based on who it thinks will convert, not who it thinks will click.
that distinction is the whole thing. it can only make that judgement using the conversion signals you send it.
The three prerequisites
Working tracking, verified with a real event. not a green pixel, an actual conversion that appeared in Ads Manager. under this objective, missing signals do not just break reporting, they degrade delivery, because the model is learning from data that is not arriving. the setup is in ChatGPT Ads conversion tracking.
A fresh campaign. you cannot convert an existing CPM or CPC campaign to oCPC. create a new one with the objective set to conversions. plan the switch rather than discovering this mid-quarter.
Enough conversion volume to learn from. this is the one people skip. a campaign producing three conversions a week has almost nothing to optimise against. if your conversion event is too rare, optimise toward something earlier in the funnel that correlates with it, like a trial start rather than a paid renewal.
Leave it alone
the hardest discipline is doing nothing for the first stretch.
every meaningful edit, budget, targeting, creative, restarts learning to some degree. the pattern I fell into early was checking daily, seeing a bad first two days, changing something, and thereby guaranteeing it never got past learning. the campaign never had a stable period long enough to be judged.
set a budget you are comfortable losing, set a date you will review it, and do not open it in between.
The four changes that actually moved cost per conversion
after three months, these are the ones that mattered. everything else was noise.
One. Sending a value with the conversion. optimising toward “a conversion” treats every customer as identical. once real values were flowing, delivery shifted toward the users worth more, and the blended number improved without any change to targeting.
Two. Moving the conversion event earlier. my first event was too deep in the funnel and too rare. moving it to a stronger early signal gave the system enough volume to learn, and the deeper metric improved anyway because it was being fed better users.
Three. Fixing the landing page rather than the ads. the largest single improvement came from the page, not the campaign. if people click and leave, no amount of optimisation fixes it, because you are teaching the system to find more people who will also leave.
Four. Killing the losing angle entirely instead of optimising it. a concept that is not working after a fair test is not going to be rescued by a budget change.
The number to judge it by
not cost per conversion on its own. what you can afford to pay depends on what a customer is worth and how fast you get the money back, and the LTV to CAC and payback calculator gives you both. a ratio that looks healthy with a fourteen month payback will still run you out of cash, and that is the failure mode nobody sees coming.
if you are comparing this channel against your others, the channel CAC comparison shows you where moving budget actually buys more users, and warns you when the answer is “not much”.
What I would tell a beginner
start on CPC, get tracking right, build a separate conversion campaign once you have volume, then leave it alone long enough to judge it honestly. that sequence is boring and it is most of the advantage.
the wider testing and creative discipline is in the paid ads playbook, and the full playbook library covers the fifteen other channels, because this one will not stay cheap forever.
Sources: objective behaviour and the no-conversion-of-existing-campaigns rule from OpenAI’s conversion-optimized campaigns documentation.