This is part of the "unchecked" part of AB testing the headline mentions. You, of course, need to ensure the granular tweaks can be rolled up into something us

This is part of the "unchecked" part of AB testing the headline mentions. You, of course, need to ensure the granular tweaks can be rolled up into something usable as the granular tweaks prove successful. You can't just keep bolting on UI changes while losing sight of the larger experience. Each incremental A/B test is testing against a previously successful variant so eventually the control is radically different from where it started and you're only concerned about beating the control. Using a longer-term holdout group or reseting the control experience during incremental testing can help mitigate this and get you zoomed out a bit from the local maxima.