In this paper we argue that an attractive and potentially very general way of achieving generalized arc consistency (GAC) on a constraint is by using unit propagation (UP) over a CNF encoding of the constraint. This approach to GAC offers a number of advantages over traditional constraint specific algorithms (propagators): it is easier to implement, it automatically provides incrementality and decrementality in a backtracking context, and it can provide clausal reasons to support learning and non-chronological backtracking. Although UP on standard CNF encodings of a constraint fails to achieve GAC, we show here that alternate CNF encodings can be used on which UP does achieve GAC. We provide a generic encoding applicable to any constraint. We also give structure specific encodings for the regular, among, and gen-sequence constraints on which UP can achieve GAC with the same run time bounds as previously presented propagators. Finally, we explain how a UP engine can be added to a CSP solver to achieve a seamless integration of constraints encoded in CNF and propagated via UP and those propagated via traditional constraint specific propagators.