Many XML-based information systems that must handle highly heterogeneous information require multiple similarity measures. Until now, little guidance exists for the design of application-dependent measures in such systems. This paper contributes a four-step methodology that guides the development of multi-similarity systems, and shows its usefulness with a use case. The experiments show that the methodology can benefit from appropriate tools (test collection generators, query systems, assessment criteria), thus achieving significant reductions in the complexity of the design process.