We describe and test two process models of candidate evaluation. The memory-based model holds that evaluations are dependent on the mix of pro and con information retrieved from memory. The impression-driven model holds that evaluations are formed and updated “on-line” as information is encountered. The results provide evidence for the existence of stereotyping and projection biases that render the mix of evidence available in memory a nonveridical representation of the information to which subjects were exposed. People do not rely on the specific candidate information available in memory. Rather, consistent with the logic of the impression-driven processing model, an “on-line” judgment formed when the information was encountered best predicts candidate evaluation. The results raise both methodological and substantive challenges to how political scientists measure and model the candidate evaluation process.