It's actually a more subtle problem: if you test a radical redesign, there's a higher likelihood it will fail even if a subsequent series of small changes to it would have produced something better, just because the incumbent is so well optimized.