OBJECTIVE: To determine the accuracy of patient recall of a question about smoking in a specified consultation in general practice; the accuracy of smokers' recall of advice to stop smoking; and predictors of accurate recall. DESIGN: Analysis of 1075 audiotapes was compared blind with matching patient questionnaires administered after the consultation to calculate sensitivity and specificity for patient recall as a dichotomous variable. Predictors of recall were determined by logistic regression. SETTING: General practitioner training practices, New South Wales, Australia. SUBJECTS: Thirty-four trainees and consecutive samples of their patients 16-65 years of age. RESULTS: Patient recall of a question about smoking had a high false positive rate of 21% but a sensitivity of 93%. Smokers over-reported advice to stop smoking (specificity 82%; sensitivity 92%). Predictors of accuracy of a question about smoking included patient sex (women 1.58 times more likely than men to be accurate); smoking status (smokers 1.7 times as likely as non-smokers to be accurate); and interval since consultation (those who completed their questionnaires within a week were 1.84 times more likely to be accurate). CONCLUSIONS: Patient recall is systematically biased towards over-reporting of a question about smoking status and, among smokers, of advice to quit. Although we recommend its continued application in health services evaluation, findings should be interpreted with caution, particularly if subjects are male or the interval between recall of smoking cessation advice and the occasion of service in which it might have occurred is considerably delayed.