•Online MR-guided soft-tissue registration protocols exhibited modest improvements of the geometric target coverage probability as compared to online CBCT-guided bone match protocols.•Therefore, highly conformal target irradiation using online MR-guidance can only be achieved by implementing on-table adaptive workflows where new treatment plans are daily generated based on the anatomy of the day. This study aimed to quantify the coverage probability for esophageal cancer radiotherapy as a function of a preset margin for online MR-guided and (CB)CT-guided radiotherapy. Thirty esophageal cancer patients underwent six T2-weighted MRI scans, 1 prior to treatment and 5 during neoadjuvant chemoradiotherapy at weekly intervals. Gross tumor volume (GTV) and clinical target volume (CTV) were delineated on each individual scan. Follow-up scans were rigidly aligned to the bony anatomy and to the clinical target volume itself, mimicking two online set-up correction strategies: a conventional CBCT-guided set-up and a MR-guided set-up, respectively. Geometric coverage probability of the propagated CTVs was assessed for both set-up strategies by expanding the reference CTV with an isotropic margin varying from 0 mm to 15 mm with an increment of 1 mm. A margin of 10 mm could resolve the interfractional changes for 118 out of the 132 (89%) analyzed fractions when applying a bone-match registration, whereas the CTV was adequately covered in 123 (93%) fractions when the registration was directly performed at the CTV itself (soft-tissue registration). Closer analyses revealed that target coverage violation predominantly occurred for distal tumors near the junction and into the cardia. Online MR-guided soft-tissue registration protocols exhibited modest improvements of the geometric target coverage probability as compared to online CBCT-guided bone match protocols. Therefore, highly conformal target irradiation using online MR-guidance can only be achieved by implementing on-table adaptive workflows where new treatment plans are daily generated based on the anatomy of the day.