Right, the cost functions could be defined in terms profit maximization (thus, the cost being lost sales, eg).

Either way, these experiments are very difficult to control because you can't fully validate models by predicting sales a posteriori with holdouts once you have the data as you don't know if some algorithm might influence certain sales not reflected in such real-life data. You could try a/b testing, but I'm not sure the Amazons are willing to run someone's models very readily. Either way, this is a good thing.