Effect handlers are a modern and increasingly popular approach to structuring computational effects in functional programming languages. However, while their traditional operational semantics is well-suited to implementation tasks, it is less ideal as a reduction theory. We introduce a fine-grained reduction theory for deep effect handlers, along with a standard reduction strategy. We relate this strategy to the traditional, non-local operational semantics via a simulation argument, and show that the reduction theory preserves observational equivalence with respect to the classical semantics of handlers, thus allowing its use as a rewriting theory for handler-equipped programming languages. In the process we establish theoretical properties of our reduction theory, including confluence and standardisation theorems, adapting and extending existing techniques. Finally, we demonstrate the utility of our semantics by providing the first normalisation-by-evaluation algorithm for effect handlers, and prove its soundness and completeness. Additionally, we establish non-expressibility of the lift operator, found in some effect-handler calculi, by the other constructs.