When an A/B test is not best