The prognostic impact of serum α-fetoprotein (AFP) level in patients with hepatocellular carcinoma (HCC) is controversial. This study aimed to investigate the predictive ability of serum AFP in HCC patients. A total of 543 patients undergoing surgical resection (258 patients) and non-surgical treatment (285 patients) including transarterial chemoembolization and percutaneous injection therapy were retrospectively studied. Overall, AFP level> 400 ng/ml was an independent poor prognostic predictor [relative risk (RR): 1.4, 95% confidence interval (CI): 1.0-1.9, p= 0.049]. Stratified analysis showed that there was a sharp contrast of predictive power of AFP level in treatment strategy and tumor size. In surgical patients, serum AFP> 400 ng/ml was a tumor size-independent predictor of tumor recurrence (RR: 1.7, 95% CI: 1.2-2.5, p= 0.006) and survival (RR: 2.3, 95% CI: 1.3-3.8, p= 0.002). However, there was no association between AFP level and survival in the non-surgical group (p= 0.597). Alternatively, among the 157 patients with large (> 5 cm) HCCs, AFP> 400 ng/ml independently predicted a poor survival (RR: 1.9, 95% CI: 1.2-2.5, p= 0.012), whereas no clear relationship between AFP level and survival was found among the 386 patients with small (≤ 5 cm) HCCs (p= 0.685). There was no differential prognostic impact of serum AFP levels in other variables. In conclusion, serum AFP level is a weak prognostic predictor in HCC patients. Its predictive ability is highly selective and dependent on treatment strategy and tumor size. Incorporation of serum AFP level into any prognostic prediction model should be based on its distinctive selective prognostic power.