Gravity Spatial Interaction Models have been used consistently to model migration, commuting, and trade. However, classic gravity models were developed and have mainly been used to predict among networks where origins can serve as destinations, and vice versa. Without this unipartite structure, gravity model performance is not as clear. Additionally, spatial interaction models are usually assessed using their predictive performance alone, which does not allow evaluate how well the models capture the overall pattern of ows. Both of these problems result from one common source - the concept of “spatial structure” is still not clearly or consistently conceptualized or used appropriately in models. In this work, we explore the concept of spatial structure and analyze its representation in current modelling frameworks. We then explore the potential of a graph structure measure, Page Rank, to provide a general measure of spatial structure. We examine Page Rank by comparing how classical spatial interaction model accessibility terms and Page Rank respond to changes in the interaction network with unipartite and bipartite structure. Lastly, we compare models built with these measures using standard predictive performance methods, as well as comparing their fidelity to the overall spatial structure of observed networks. We find that Page Rank is sensitive to network structure in both unipartite and bipartite graphs, and it accounts for changes in structure in both a local and a global sense. We also find that Page Rank improves upon classic measures of accessibility in spatial interaction modelling, since it does not depend on unipartite structure and it yields better estimates in terms of both predictive performance and pattern replication. Overall, this work encourages us to think more critically about measures of spatial structure in spatial interaction models and widen our ideas of what constitutes “good performance” from a spatial interaction model.