The recent emphasis on identifying an ‘evidence base’ for classroom practice has generated a series of systematic reviews of the best available research evidence in an attempt to identify best practice in promoting positive learning outcomes for pupils. Drawing on our recent systematic review of research on classroom dialogue in mathematics lessons, commissioned by the Department for Children, Schools and Families (DCSF), we have identified a number of epistemological problems which: (i) place limitations on how far such research findings actually generate useable and convincing ‘evidence based knowledge’ regarding classroom practice; and (ii) place limitations on the extent to which ‘evidence based knowledge’ can interface with ‘practitioner knowledge’. We suggest that there are two contrasting versions of the ‘evidence base discourse’: an eclectic version, where the research analysed encompasses findings from all types of research approaches and the synthesis is interpreted by teachers and others in a way that ‘informs’ rather than ‘directs’ their decision-making; and an RCTs-privileged version where studies using a randomised control trials (RCTs) design are privileged over other types of studies as providers of ‘best evidence’ and the synthesis forms the basis for governmental and school level policies concerning which practices teachers should be directed to make use of. In the light of our analysis we conclude that the evidence base discourse needs to align itself close to the eclectic version.