BACKGROUND: Previous work has demonstrated that the recording of acute health outcomes, such as myocardial infarction (MI), may be suboptimal in primary healthcare databases. AIM: To assess the completeness and accuracy of the recording of stroke in UK primary care. DESIGN & SETTING: A population-based longitudinal cohort study. METHOD: Cases of stroke were identified separately in Clinical Practice Research Datalink (CPRD) primary care records and linked Hospital Episode Statistics (HES). The recording of events in the same patient across the two datasets was compared. The reliability of strategies to identify fatal strokes in primary care and hospital records was also assessed. RESULTS: Of the 75 674 stroke events that were identified in either CPRD or HES data during the period of the study, 54 929 (72.6%) were recorded in CPRD and 51 013 (67.4%) were recorded in HES. Two-fifths (n = 30 268) of all recorded strokes were found in both datasets (allowing for a time window of 120 days). Among these 'matched' strokes the subtype was recorded accurately in approximately 75% of CPRD records (compared with coding in HES); however, 43.5% of ischaemic strokes in HES were coded as 'non-specific' strokes in CPRD data. Furthermore, 48.2% had same-day recordings, and 56.2% were date-matched within ±1 day. CONCLUSION: The completeness and accuracy of stroke recording is improved by the use of linked hospital and primary care records. For studies that have a time-sensitive research question, the use of linked, as opposed to stand-alone, CPRD data is strongly recommended.