We develop a methodology for studying and comparing the expressiveness of computational models which are based on rewriting. We consider a class of rewriting systems, which we call normal, admitting a natural interpretation in terms of event structures; this is done by building up on work of Mazurkiewicz, Nielsen, Plotkin, Winskel, Melliès, and Mimram. Then, we introduce the notion of bisimilar embedding of event structures, which allows us to say when a computational system is "at least as expressive" as another one, as soon as these two are described in terms of event structures. Finally, we prove a few separation results for normal rewriting systems based on this notion, and we give an application of these to interaction nets and their non-deterministic variants, including Ehrhard and Regnier's differential interaction nets.