Aim To investigate simultaneously the effect of voxel size and fracture width on the accuracy of detecting vertical root fractures (VRFs) in non‐root filled teeth when using cone beam computed tomography. Methodology Fifty‐one of 161 extracted human permanent teeth (16 anterior teeth, 132 premolars and 13 mandibular molars) were selected randomly for VRF induction with two fracture widths. All teeth were scanned with four CBCT units at different voxel sizes provided by the units. Three observers classified the presence or absence of VRF using a 5‐point scale. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and area under the ROC curve (AUC) were calculated. AUCs amongst voxel sizes and between the fracture widths were compared using the Z test. Intra‐ and inter‐observer agreement was assessed using weighted Cohen kappa. Results For the NewTom VGi and ProMax 3D Mid CBCT unit, no significant differences were found amongst voxel sizes for the AUCs, irrespective of the fracture width (P > 0.05). There were significant differences between images scanned with voxel size 250 and 160 μm (P = 0.02), and images scanned with voxel size 250 and 80 μm for AUCs in the narrow VRF group for the 3D Accuitomo 170 unit (P = 0.03). For i‐CAT FLX, significant differences were found between the voxel protocols of 300 μm and of the other three voxel sizes for AUC, sensitivity and NPV (P < 0.05). Significant differences between the wide and the narrow VRF groups for AUCs were found for 3D Accuitomo 170 (P = 0.01) and ProMax 3D Mid (P < 0.01). Conclusions Cone beam computed tomography was accurate for detecting VRF in non‐root filled teeth. Fracture width had an effect on the detection of VRF. The effect of the voxel size on the detection of VRF depended on the CBCT unit used.