In this paper, the authors consider three theories that have dominated discussions of metaphor. One view is that metaphors make comparisons, the basis for the comparison being the features (or categories) that the terms of the metaphor share. The second view is that metaphors involve an anomaly. The third view is that metaphors are ‘interactive’, producing a new way of seeing the terms. Tourangeau and Sternberg propose a new theory—the domains-interaction view—that draws on elements of all three earlier views, but borrows especially from the interaction view. The authors consider the implications of our theory for three questions: What are metaphors? How are they understood? What makes a good metaphor? They argue that metaphors correlate two systems of concepts from different domains. The best metaphors involve two diverse domains (more distance between domains making for better metaphors) and close correspondence between the terms within those domains. Metaphors are interpreted in several stages: the terms of the metaphor are encoded; the domains involved are inferred; the structures to be seen as parallel are found; the correspondences between these structures are ‘mapped’ or constructed; the terms of the metaphor are compared. If the terms are not seen to match or occupy analogous roles in their different domains, then the metaphor may be reinterpreted. The evidence on all this is tentative but supports our view. We review two studies (Tourangeau and Sternberg, 1981) that support the hypothesis that distance within domain relates negatively to aptness, whereas distance between domains relates positively. Several studies on comprehension tend to disconfirm the comparison theory's notion that the tenor and vehicle necessarily share features. Tenor and vehicle also appear to have asymmetrical roles in the interpretive process.