Objectives Response evaluation criteria in solid tumors (RECIST) often fail to identify clinically meaningful response to bevacizumab-containing therapy in colorectal liver metastasis (CRLM). This study aimed to develop RECIST by combining contrast-enhanced and diffusion-weighted magnetic resonance imaging (MRI). Methods A total of 126 patients with CRLM who underwent hepatic resection after bevacizumab-containing chemotherapy were split into initial analyses cohort (N = 42, with 76 indexed liver metastases) and validation cohort (N = 84). In lesion-based analyses, percentage decrease of arterial enhancement area and percentage increase of apparent diffusion coefficient (ADC) value from baseline to post-chemotherapy were measured. Their optimal cutoff values for distinguishing pathology-confirmed major and minor response were determined. Then, the developed RECIST (D-RECIST) was established by combining functional and size-based items. Survival relevance of D-RECIST and RECIST was examined in the validation cohort. Results Percentage decrease of arterial enhancement area and increase of ADC value significantly differed between lesions of pathologic major or minor response, with optimal cutoffs of approximately 33% and 19%, respectively. Patients defined as responders by D-RECIST had a significantly longer median disease-free survival (DFS) than non-responders (p = 0.021; 12.9 versus 8.6 months). No significant difference was observed with RECIST (p = 0.524). In a Cox regression model, D-RECIST- but not RECIST-defined responses independently predicted the DFS (p = 0.034 and 0.811). Conclusions D-RECIST-defined responses provided significant prognostic information, and thus may serve as a better response evaluation approach than RECIST in CRLM treated with bevacizumab-containing therapy. Key Points Changes in arterial enhancement area and apparent diffusion coefficient value are associated with pathological response in colorectal liver metastases treated with bevacizumab. The MRI-based response criteria developed by combining size-based and functional features can provide significant prognostic information.