This paper explores the effect of social structures on the beliefs and knowledge of agents who reason in an evidence-based manner. We introduce Evidence Diffusion Models for the formal analysis of multi-agent evidence-based reasoning, focusing on the context of a social network where evidence pieces are being communicated between agents via threshold-limited diffusion. Our models bridge the gap between multi-agent Topological Evidence Models (Partitional Models) and Threshold Models for Diffusion. Firstly, we show that, in our setting, network structures are expressible and known both defeasibly and infallibly by the agents. Then, we prove that defeasible knowledge and (defeasible) ‘evidence-based distributed knowledge’ are easily lost under the diffusion of pieces of evidence in a network, whereas ‘group knowledge with distributed evidence’ is strongly robust. Finally, we obtain so-called Cluster Theorems characterising the evidential and network conditions for evidence cascades to form, and for individuals and groups to obtain knowledge in the diffusion process.