Agents attempting to acquire information often lack exogenous information technologies of their own and thus rely on experts conveying cheap talk messages. To address this gap, I examine a model of costly search in which an uninformed receiver sequentially consults randomly drawn cheap-talking experts. Crucially, the pool of experts is heterogeneous, and the receiver cannot observe the sender’s motives. The dynamic nature of search creates a potential time inconsistency problem for the receiver, so the receiver searches weakly too much. In particular, low search costs make the receiver strictly worse off whenever i) senders’ preferences are sufficiently opposed and ii) there is a high enough incidence of ‘informative’ senders. Applying this model to social media regulation generates a key insight: maximally informative social media algorithms discriminate across users.