Stem cell therapy holds great promise in medicine, but clinical development should be based on a sound understanding of potential weaknesses in supporting experimental data. The aim of this article was to provide a systematic overview of evidence relating to the efficacy of stem cell-based therapies in animal models of stroke to foster the clinical application of stem cell-based therapies and to inform the design of large-scale clinical trials. We conducted a systematic search for reports of experiments using stem cells in animal models of cerebral ischaemia, and performed DerSimmonian and Laird random effects meta-analysis. We assessed the impact of study characteristics, of publication bias and of measures to reduce bias. We identified 6059 publications, 117 met our prespecified inclusion criteria. One hundred eighty-seven experiments using 2332 animals described changes in structural outcome and 192 experiments using 2704 animals described changes in functional outcome. Median study quality score was 4 (interquartile range 3 to 6) and less than half of studies reported randomization or blinded outcome assessment; only three studies reported a sample size calculation. Nonrandomized studies gave significantly higher estimates of improvement in structural outcome, and there was evidence of a significant publication bias. For structural outcome autologous (i.e. self-derived) stem cells were more effective than allogeneic (donor-derived) cells, but for functional outcome, the reverse was true. A significant dose–response relationship was observed only for structural outcome. For structural outcome, there was an absolute reduction in efficacy of 1·5% (−2·4 to −0·6) for each days delay to treatment; functional outcome was independent of the time of administration. While stem cells appear to be of some benefit in animal models of stroke the internal and external validity of this literature is potentially confounded by poor study quality and by publication bias. The clinical development of stem cell-based therapies, in stroke and elsewhere, should acknowledge these potential weaknesses in the supporting animal data.