Decision makers often infer population-level social statistics such as risk frequencies or consumer preferences by recalling people from their own social networks from memory—that is, by social sampling. Although people’s social interactions increasingly occur on social media rather than through face-to-face contact, it is unclear which influence online social media contacts have on people’s inferences. In this study, we examine to what extent social sampling is affected by whether one usually interacts with a person online or offline, and which weight online contacts have for people’s inferences. Participants judged the prevalence of different health issues (e.g., anxiety disorders) and recalled people in their personal social networks who had experienced each issue. For each recalled case, participants indicated the primary mode of contact (offline, online, or mixed) and the social category (self, family member, friend, or acquaintance). Based on Bayesian hierarchical mixture modeling, we compared sequential, limited social sampling strategies guided by either contact mode or social category to exhaustive search and guessing. Most participants were best described by a strategy that assumes limited rather than exhaustive search. Social sampling based on contact mode provided the best account for around a third of participants. The estimated model parameters suggested that participants relied less strongly on information from online contacts than on information from other social subgroups. Thus, in addition to demonstrating that the mode of contact is used to guide social sampling, our results suggest that social media contacts are less important for people’s inferences than face-to-face contacts.