Research on social capital routinely relies on survey measures of trust which can be collected in large and heterogeneous samples at low cost. We validate such survey measures in an incentivized public good experiment and show that they are importantly related to cooperation behavior in a large and heterogeneous sample. We provide evidence on the microfoundation of this relation by use of an experimental design that enables us to disentangle preferences for cooperation from beliefs about others' cooperation. Our analysis suggests that the standard trust question used in the World Values Survey is a proxy for cooperation preferences rather than beliefs about others' cooperation. In contrast, the “fairness question”, a recently proposed alternative to the standard trust question, seems to operate through beliefs rather than preferences.