BACKGROUND: Data on smoking held by general practitioners (GPs) may contribute to clinical care and to an assessment of population health. However, these data are prone to several biases and their validity has not been tested. AIM: To examine the accuracy of general practice data as an estimate for population prevalence of smoking and to estimate the accuracy of GP data on individuals' smoking habit compared with self-report. METHODS: A postal questionnaire on smoking habit over the past six years was sent to a random sample of individuals aged 15 to 74 years and registered with five out of seven general practices in one part of Leicester. GP records of those sampled were examined for an entry of smoking status over this period. RESULTS: Response rate to the postal questionnaire was 1906 out of 2490 (76.5%). Reported smoking prevalence was 35.2%. Of those notes sampled, 1784 out of 2432 (73.4%) had an entry about smoking recorded between 1991 and 1996. Patients recorded as smokers were less likely to respond to the postal questionnaire than non-smokers. Using practice data to ascribe smoking status to non-responders produced an estimated prevalence of 38.6%. Using questionnaire data alone as the 'gold standard', the last practice record collected since 1991 overestimated current smoking prevalence by a factor of 1.22; using questionnaire data supplemented by practice data for non-responders as the 'gold standard' meant that the overestimate was by a factor of 1.11. Data from notes and the questionnaire were available for 1398 individuals and 2188 observations. Levels of agreement were high (kappa = 0.83). CONCLUSION: GP-held data are valid for individuals but over-estimate smoking prevalence at a population level.