The image-guided CyberKnife radiosurgery system is apable of tracking spinal targets without fiducial implants. Recently, a new version of this fiducial-free image guidance modality ("enhanced Xsight spine tracking") has been introduced. We assessed the accuracy of this novel technique versus its precursor in a comparative phantom study. The CyberKnife consists of a 6 MV linac on a six-axis robot and a stereoscopic kV image guidance system. An anthropomorphic head-and-neck phantom with a cervical spine section was mounted on the linac nozzle. The robotic manipulator was used to precisely move the phantom to defined positions in the CyberKnife workspace. Multiple stereoscopic images were acquired at different translational and rotational positions. The enhanced Xsight spine tracking readouts were recorded and compared to the nominal phantom position. These tests were repeated with the original Xsight spine tracking version to analyze potential differences. Enhanced Xsight spine tracking correctly reported translational offsets with an RMS error of less than 0.4 mm. Yaw and roll rotations were detected with an accuracy of 0.2°, 0.25°. Pitch offsets were slightly underestimated, with up to 0.3° for an offset of ± 2°. Nominal X (left-right) translational offsets were partially misinterpreted as roll (0.2° at a 10 mm offset). Apart from this, no correlation between rotational and translational directions was found. In comparison, the original Xsight spine tracking showed identical results for translations, but larger systematic and statistical errors for rotations. Enhanced Xsight spine tracking measurably improves precision in fiducial-free spinal radiosurgery with the CyberKnife.