The purpose of this study is to assess the extent and serial pattern of setup error of conventional fractionated whole pelvic irradiation using a kilovoltage on‐board imager. The daily on‐board images of 69 patients were matched with the digitally reconstructed radiographs of simulation on the basis of pelvic bony structure. The shifts along x‐ (lateral), y‐ (longitudinal), and z‐ (vertical) axes, and the 3D vector, were measured. The shift between an origin of the first fraction and each fraction (Δshift1st) and the shift between an isocenter of simulation and each fraction (ΔshiftSim) were calculated. To evaluate serial changes, the shifts of each fraction were classified into four consecutive sessions, and an ANOVA and chi‐square test were used. The systematic error of the ΔshiftSim and Δshift1st were 2.72 and 1.43 mm along the x‐axis, 2.98 and 1.28 mm along the y‐axis, and 4.26 and 2.39 mm along the z‐axis, respectively. The ΔshiftSim and Δshift1st≥5mm of the 3D vector occurred in 54.3% and 23.1%, respectively. The recommended margins to cover setup error in case of using Δshift1st were 3.81, 3.54, and 6.01 mm along x‐, y‐, and z‐axes, whereas those using ΔshiftSim were 6.39, 6.95, and 9.95 mm, respectively. With the passage of time, the Δshift1st≥5mm of 3D vector and along any axis in supine setup increased from 14.1% for first session to 22.5% for fourth session (p=0.027) and from 10.8% to 18.5% (p=0.034), respectively. In prone setup, first session was better than others in the Δshift1st≥5mm of 3D vector and along any axis. It is expected that the correction using the on‐board images on the first fraction improves geometrical uncertainties and reduces the margin for target coverage. Daily continuous OBI follow‐up during conventional fractionated pelvic irradiation can increase the reproducibility and be more effective in the late period.PACS number: 87.55.km