I think a better approach might be using bayesian optimization, each variation would be a dimension, and the algorithm would find the (almost) optimal possibili

I think a better approach might be using bayesian optimization, each variation would be a dimension, and the algorithm would find the (almost) optimal possibility instead of a/b testing, which finds the local maximus.