The association between the polymorphic CAG repeat in androgen receptor gene (AR) and prostate cancer susceptibility has been studied extensively. However, the results are contradictory. The purpose of our meta-analysis was to investigate whether CAG repeat related to prostate cancer risk and had genetic heterogeneity across different geographic regions and study designs. Random-effects model was performed irrespective of between-study heterogeneity. Data and study quality were assessed in duplicate. Publication bias was assessed by the fail-safe number and Egger’s test. There were 16 (patients/controls: 2972/3792), 19 (3835/4908) and 12 (3372/2631) study groups for comparisons of ≥20, 22 and 23 repeats of CAG sequence, respectively. Compared with CAG repeat <20, 22 or 23, carriers of ≥20, 22 or 23 repeats had 21% (95% CI: 0.61–1.02; P = 0.076), 5% (95% CI: 0.81–1.11; P = 0.508) and 5% (95% CI: 0.76–1.20; P = 0.681) decreased risk of prostate cancer. After classifying studies by geographic areas, carriers of ≥20 repeats had 11% decreased risk in populations from USA, 53% from Europe, and 20% from Asia (P > 0.05), whereas comparison of ≥23 repeats with others generated a significant prediction in European populations (OR = 1.17; P = 0.039). Stratification by study designs revealed no material changes in risk estimation. Meta-regression analysis found no significant sources of between-study heterogeneity for age, study design and geographic region for all comparisons. There was no identified publication bias. Taken together, our results demonstrated that AR CAG repeat polymorphism with ≥20 repeats might confer a protective effect among the prostate cancer patients with 45 years older but not all the prostate cancer patients.