The Alcohol Use Disorders Identification Test (AUDIT) is a 10–item questionnaire designed by the World Health Organization to screen for hazardous alcohol intake in primary health care settings. In this longitudinal study we examine its performance in predicting alcohol‐related harm over the full range of its scores using receiver operating characteristic analyses. Three hundred and thirty ambulatory care patients were interviewed using a detailed assessment schedule which included the AUDIT questions. After 2‐3 years, subjects were reviewed and their experience of alcohol‐related medical and social harm assessed by interview and perusal of medical records. A UDIT was a good predictor of both alcohol‐related social and medical problems. Cut‐off points of 7‐8 maximized discrimination in the prediction of trauma and hypertension. Higher cut‐offs (12 and 22) provided better discrimination in the prediction of alcohol‐related social problems and of liver disease or gastrointestinal bleeding, but high specificity was offset by reduced sensitivity. We conclude that the recommended cut‐off score of eight is a reasonable approximation to the optimal for a variety of endpoints.