Important economic decisions – from a potential buyer appraising a complex product to a policymaker evaluating the potential impact of a novel social program – rely on selective and budgeted exploration of the numerous attributes of the object considered for adoption. I propose a novel and tractable model of attribute discovery. Similarity among uncertain attributes is modeled through an arbitrary Gaussian process, allowing for a rich class of inference patterns. If a single agent decides on both sampling and subsequent adoption, the optimal sample is maximal with respect to a new centrality measure in the inference graph among attributes.  It is moreover unaffected by prior attribute means and the order in which attributes are discovered: sampling is neutral and thorough.  This no longer holds if adoption is decided upon by a principal. The agent’s value from a given sample hinges on its centrality for the principal and its alignment of players’ interests (i.e. cross-centrality) – thus reflecting sampling’s dual purposes of learning and persuasion. The latter is surprisingly a stronger motive when the players agree ex ante on the promise of the object, leading to the optimal sample suppressing valuable information for both players. I analyze the various manifestations such suppression takes, outlining a cautionary tale about the political context underlying small-scale evaluation of social programs.