The reported CPA is not your real cost
A platform's CPA only counts its own attributed conversions, over its own attribution window. Adding up the conversions of several platforms invariably gives a total higher than reality: the same sale is claimed more than once.
The honest calculation divides total spend, across all platforms, by the number of conversions actually recorded in your back office. This consolidated CPA is almost always higher than the average of the platform CPAs, and it is the one that determines profitability.
The gap between the two is an indicator in its own right: if it suddenly widens, that is the sign of a measurement failure or new double counting, not of a drop in performance.
Target CPA and the learning phase
Target CPA is an instruction given to the bidding algorithm, not a guarantee. Set too low, it restricts delivery to the point of drying up conversion volume, and deprives the algorithm of the signals it needs to improve.
Every significant change to the target CPA restarts a learning phase, during which performance is unstable and unrepresentative. Judging a campaign during its learning phase leads to systematically wrong decisions.
The practical rule: change it in small steps, let the learning phase run its full course, and only compare stabilized periods.
Frequently asked questions
How do you calculate CPA?
Divide ad spend by the number of conversions over the same period. €2,000 of spend for 50 conversions gives a CPA of €40.
What is the difference between CPA and target CPA?
CPA is observed after the fact. Target CPA is an instruction given to the bidding algorithm. The latter steers delivery; it does not guarantee the former.
Why is my consolidated CPA higher than the platforms' CPA?
Because each platform credits itself with conversions that others also claim. Dividing total spend by the conversions actually recorded in your back office gives the only CPA that determines profitability.
Should you lower the target CPA when performance drops?
Rarely in the heat of the moment: lowering it restarts learning and restricts delivery, which often makes the problem worse in the short term. First check that the drop is not a measurement artifact.