An instruction, not a guarantee
Target CPA steers delivery: the algorithm bids higher when it judges the probability of conversion to be high, and lower otherwise. The average it aims for is the target CPA, with some spread around it.
Setting it too low is the most common and most costly mistake. Under that constraint, the algorithm restricts delivery to only the auctions it judges very safe. Volume collapses, and with it the signals it needs to improve. You end up with a CPA that looks on target, on a trivial volume.
The sensible starting point is the CPA actually observed over a recent stable period, not the one you are hoping for. You then tighten it gradually.
Respect the learning phase
Any significant change to the target CPA restarts a learning phase. During that period, performance is unstable and unrepresentative: judging it leads to systematically wrong decisions.
The most destructive reflex is to make correction after correction: each adjustment restarts learning, so the campaign never gets out of it and never stabilizes.
The practical rule: one change at a time, in moderate steps, then wait for learning to end completely before judging. A monitoring setup should also exclude campaigns in learning from its alerts, or it will produce a constant stream of false positives.
Frequently asked questions
How do you set a target CPA?
Start from the CPA actually observed over a stable period, not from the one you are hoping for. Then tighten it in small steps, letting each change stabilize.
Why does lowering the target CPA hurt performance?
Because the algorithm restricts delivery to meet the constraint: volume drops, learning signals become scarce, and the campaign loses efficiency. The reported CPA may look on target, on a volume that has become negligible.
How long does the learning phase last?
It depends on conversion volume: the lower it is, the longer the phase. The reliable marker is not a duration but a return to stable performance over several consecutive days.
tCPA or tROAS: which should you choose?
tCPA when all conversions have roughly the same value, tROAS when order values vary widely. Using tCPA on uneven order values amounts to optimizing for volume at the expense of value.