A grammar of conversation was developed using three units of analysis: talk acts, turns at talk, and topic sequences. Statements, questions, agreements, and disagreements were used as categories of talk acts. Rules were stipulated covering the use of the four kinds of talk acts. From the rules and the definitions of the units of analysis, propositions were derived about the probable location of categories of talk acts in turns and sequences. Theorems and hypotheses were deduced from these propositions. Empirical tests of each of the 16 hypotheses were conducted on transcripts of police‐civilian telephone calls, husband‐wife decision‐making sessions, and selected Watergate transcripts. of 48 possible tests, four could not be made due to lack of data. Forty‐three of the remaining 44 were statistically significant. However, 16 tests showed marginal or poor power levels. The data tended to indicate that knowledge of the grammar of conversation permitted significantly better prediction of the occurrence of acts than would be expected by chance. Further, the grammar also predicted adequately for three diverse social and interpersonal sit