This paper empirically investigates how humans use reference in space when interacting with a multimodal system able to understand written natural language and pointing with the mouse. We verified that user expertise plays an important role in the use of multimodal systems: experienced users performed 84% multimodal inputs while inexpert only 30%. Moreover experienced are able to efficiently use multimodality shortening the written input and transferring part of the reference meaning on the pointing. Results showed also the importance of the system layout: when very short labels (one character) are available users strongly adopt a redundant reference strategy, i.e. they referred to the object in a linguistic way and use pointing too. Starting from these facts some guidelines for future multimodal systems are suggested.