Objective To assess the accuracy of S100B serum level to detect intracranial injury in children with mild traumatic brain injury. Methods A multicenter prospective cohort study was carried out in the paediatric emergency departments of three tertiary hospitals in Switzerland between January 2009 and December 2011. Participants included children aged <16 years with a mild traumatic brain injury (GCS ≥13) for whom a head CT was requested by the attending physician. Venous blood was obtained within 6 h of the trauma in all children for S100B measurement before a head CT was performed. As the S100B value was not available during the acute care period, the patient's management was not altered. The main measures were protein S100B value and the CT result. Results 20/73 (27.4%) included children had an intracranial injury detected on CT. S100B receiver operating characteristics area under the curve was 0.73 (95% CI 0.60 to 0.86). With a 0.14 µg/L cut-off point, S100B reached an excellent sensitivity of 95% (95% CI 77% to 100%) and 100% (95% CI 81% to 100%) in all children and in children aged >2 years, respectively. The specificity, however, was 34% (95% CI 27% to 36%) and 37% (95% CI 30% to 37%), respectively. Conclusions S100B has an excellent sensitivity but poor specificity. It is therefore an accurate tool to help rule out an intracranial injury but cannot be used as the sole marker owing to its specificity. Used with clinical decision rules, S100B may help to reduce the number of unnecessary CT scans.