The split unit follows what you generalize to

When splitting data into train/test, take into account what level you generalize to and group the observations to that level prior to splitting. For example, when evaluating customer journeys, the generalization is ‘customer’, so instead of splitting observations, we split customers. This prevents the group leak.