Classical A/B has confounds. Holdouts are cleaner: half the audience gets the feature/message, half doesn't. Compare.
Can measure incrementality, what wouldn't have happened without the treatment.
Holdout group gets worse experience (no new feature). Minimize duration.