In many real world applications, autonomous mobile robots are required to observe or retrieve objects in their environment, despite not having accurate estimates of the objects' locations. Finding objects in real-world settings is a non-trivial task, given the complexity and the dynamics of human environments. However, by understanding and exploiting the structure of such environments, e.g. where objects are commonly placed as part of everyday activities, robots can perform search tasks more efficiently and effectively than without such knowledge. In this paper we investigate how probabilistic models of qualitative spatial relations can improve the performance in object search tasks. Specifically, we learn Gaussian Mixture Models of spatial relations between object classes from descriptive statistics of real office environments. Experimental results with a range of sensor models suggest that our model improves overall performance in object search tasks.