We tested whether goals during the processing of evaluative information determine the relative sensitivity of automatic evaluation to the valence of co-occurring stimuli versus the relation between the target and the affective stimuli. For example, “Kindness is uncharacteristic of Phil” has Phil co-occurring with kindness, but the relation suggests that he is unkind. In Experiment 1 (N=1248), targets co-occurred with positive or negative behaviors that were characteristic or uncharacteristic of them. In Experiment 2 (N=185), the targets started or ended pleasant or unpleasant sounds. Thus, the valence that co-occurred with targets was sometimes the opposite of the targets' valence inferred from the relation. In both experiments, we found that automatic evaluation was more sensitive to relational than to co-occurrence information when participants were instructed to form impressions than when they were instructed to memorize co-occurrence. This suggests that processing goals moderate the effect of propositional versus associative information on automatic evaluation. •Automatic evaluation can be sensitive to associative or propositional information.•We predicted that processing goals moderate that sensitivity.•We manipulate the goals during learning of co-occurrence and relational information.•A goal to form impression increased sensitivity to propositional information.•A goal to remember co-occurrence increased sensitivity to associative information.