Abstract Objective The Berlin algorithm was developed to help diagnosing axial spondyloarthritis (axSpA), but new studies suggest some features typical of SpA are less specific than previously assumed. Furthermore, evidence is lacking for other SpA subtypes (e.g. peripheral SpA). We aimed to review the evidence on the performance of SpA features for diagnosing each SpA subtype. Methods Systematic literature review of studies reporting the diagnostic performance of ≥ 1 SpA feature in patients with suspected SpA. The external reference was the rheumatologist’s diagnosis of SpA. Meta-analysis was performed, separately for each SpA subtype, to estimate pooled sensitivity, specificity, positive (LR+) and negative (LR-) likelihood ratios. Meta-regression assessed the effect of covariates (e.g. feature’s prevalence) on each feature’s performance. Results Of 13 844 articles screened, 46 were included. Sacroiliitis on magnetic resonance imaging, damage on pelvic radiographs and elevated C-reactive protein (CRP) had the best balance between LR+ and LR- (LR + 3.9–17.0, LR- 0.5–0.7) for diagnosing axSpA. HLA-B27 had an LR+ lower than anticipated (LR + =3.1). Inflammatory back pain (IBP) had low LR + (LR+∼1), but substantially decreased the likelihood of axSpA when absent (LR-=0.3). Conversely, peripheral features and extra-musculoskeletal manifestations showed high LR + (LR+ 1.6–5.0), but were as common in axSpA as no-axSpA (LR-∼1). The specificity of most features was reduced in settings when these were highly prevalent. Limited data precluded a detailed analysis on diagnosing other SpA subtypes. Conclusion Imaging features and CRP have good diagnostic value for axSpA. However, the specificity of other features, especially HLA-B27 and IBP, is lower than previously known.