The difference with the evolution argument is that evolution is based on random mutations, sometimes drastically different in a single iteration. The reason A/B tests arrive only at local maxima is because we don't often throw big random changes at it -- we make a small change and the numbers move slightly in the direction we want, so we follow that up with successive changes in the same vein, usually stopping when we can't find any (minor) modification that gives better numbers.