A 2004 meta‐analysis reported good validity for the observer attachment Q‐sort (AQS), but poor validity for the parental self‐report version. Despite this, the self‐report AQS is still widely used, with researchers arguing that providing additional training can improve its validity. The aim of this study was to update the 2004 meta‐analysis. Two hundred forty‐five studies from 1987 to 2016 were included (n = 32,426). Separate meta‐analyses were conducted to examine validity and reliability. The observer AQS showed moderate convergent validity with the Strange Situation Procedure (r = .25; r = .39 for long observation periods) and good predictive validity in terms of associations with sensitivity (r = .32). It showed a relatively weak association with infant temperament (r = .21), suggesting some discriminant validity. The self‐report version showed comparable convergent validity with Strange Situation Procedure (r = .18), but significantly weaker correlations with sensitivity (r = .25) and stronger correlations with temperament (r = .33). There was no evidence that providing additional training improved the validity of the self‐report version. This study corroborates the previous finding that the observer AQS is a valid measure of infant attachment, especially after long periods of observation. The self‐report version showed significantly weaker discriminant and predictive validity.