@paul_rietschka What people do: - sprinkle in random items to get data. But this costs because those are almost always bad. - trying to account for position bias by reweighting - maybe quickly changing assortment actually helps, but what to do about cold start problems?