Conversation analysis (CA), as currently practiced, comprises two approaches-action-oriented and meaning-oriented. I use CA treatments of 'preference' as a case in point. In current discussions of preference, the emphasis is on action, on what interactants do. Action is grounded in psychological mechanisms, which CA is not equipped to handle. So discussions of preference turn toward a more quantified notion of what people usually do. I argue that attempts at quantification raise problems that are not soluble within the confines of CA methodology. I then turn to the broadest and most discussed preference, the supposed preference for agreement, arguing that it is context sensitive in ways that produce multiple exceptions. Using a gross, transcontextual average, even if that were possible, would be unenlightening. I focus, using an extended example, on one of the exceptions, the case of accusations. I suggest that we drop the action-oriented approach and attend instead to meaning. This approach is grounded in a conception of evidence which does not rely on either falsification criteria or statistical measures. Its generalizations pertain not to what interactants normally do but to the resources they have and the methods they employ in producing meaning and social organization. Conversation analysis (CA) is the study of the organization of talk-in-interaction. It is characterized by a set of methodological techniques and perspectives: primarily, recording of ''natural'' data, participant orientation, attention to detail (''order at all points''), sequentiality, next turn proof procedure, and deviant case analysis. 1 CA, as it is currently practiced, comprises two general approaches, which I will refer to as action-oriented CA and meaning-oriented CA. In action-oriented CA, the objective is to specify what conversationalists do, in general or in specific circumstances. The focus is on the production of talk-in-interaction. Meaning-oriented CA concerns itself with the production/expression, potentialization, ______ 3 A recognitional is a form which enables recognition of the person referred to. I have suggested elsewhere (Bilmes, 1988, footnote 13) that the preference for recognitionals is not a preference at all, but a matter of implicature. Use of a recognitional implicates that the recipient knows the person referred to; use of a nonrecognitional implicates the opposite. The literature offers many examples of nonrecognitionals being corrected to recognitionals (thus the preference for recognitionals), but the converse also occurs with some frequency. Here is an example from Sacks (1992:44): A: Corliss, the g-this chick that I'm hanging around with now ((...)