BACKGROUND Nowadays, very few patients with non-variceal upper gastrointestinal bleeding fail endoscopic hemostasis (refractory NVUGIB). This subset of patients poses a clinical dilemma: should they be operated on or referred to transcatheter arterial embolization (TAE)? OBJECTIVE To carry out a systematic review of the literature and to perform a meta-analysis of studies that directly compare TAE and surgery in patients with refractory NVUGIB. METHODS We searched PubMed, Ovid MEDLINE, and Embase. A combination of the MeSH terms "gastrointestinal bleeding"; "gastrointestinal hemorrhage"; "embolization"; "embolization, therapeutic"; and "surgery" were used (("gastrointestinal bleeding" or "gastrointestinal hemorrhage") and ("embolization" or "embolization, therapeutic") and "surgery")). The search was performed in June 2018. Studies were retrieved and relevant studies were identified after reading the study title and abstract. Bibliographies of the selected studies were also examined. Statistical analysis was performed using RevMan software. Outcomes considered were all-cause mortality, rebleeding rate, complication rate, and the need for further intervention. RESULTS Eight hundred fifty-six abstracts were found. Only 13 studies were included for a total of 1077 patients (TAE group 427, surgery group 650). All selected papers were non-randomized studies: ten were single-center and two were double-center retrospective comparative studies, while only one was a multicenter prospective cohort study. No comparative randomized clinical trial is reported in the literature.Mortality. Pooled data (1077 patients) showed a tendency toward improved mortality rates after TAE, but this trend was not statistically significant (OD = 0.77; 95% CI 0.50, 1.18; P = 0.05; I 2 = 43% [random effects]). Significant heterogeneity was found among the studies.Rebleeding rate. Pooled data (865 patients, 211 events) showed that the incidence of rebleeding was significantly higher for patients undergoing TAE (OD = 2.44; 95% CI 1.77, 3.36; P = 0.41; I 2 = 4% [fixed effects]).Complication rate. Pooling of the data (487 patients, 206 events) showed a sharp reduction of complications after TAE when compared with surgery (OD = 0.45; 95% CI 0.30, 0.47; P = 0.24; I 2 = 26% [fixed effects]).Need for further intervention. Pooled data (698 patients, 165 events) revealed a significant reduction of further intervention in the surgery group (OD = 2.13; 95% CI 1.21, 3.77; P = 0.02; I 2 = 56% [random effects]). A great degree of heterogeneity was found among the studies. CONCLUSIONS The present study shows that TAE is a safe and effective procedure; when compared to surgery, TAE exhibits a higher rebleeding rate, but this tendency does not affect the clinical outcome as shown by the comparison of mortality rates (slight drift toward lower mortality for patients undergoing TAE). The present study suggests that TAE could be a viable option for the first-line therapy of refractory NVUGIB and sets the foundation for the design of future randomized clinical trials. CONCLUSIONS The retrospective nature of the majority of included studies leads to selection bias. Furthermore, the decision of whether to proceed with surgery or refer to TAE was made on a case-by-case basis by each attending surgeon. Thus, external validity is low. Another limitation involves the variability in etiology of the refractory bleeding. TAE techniques and surgical procedure also differ consistently between different studies. Frame time for mortality detection differs between the studies. These limitations do not impair the power of the present study that represents the largest and most recent meta-analysis currently available.