The purpose of this study is to assess the accuracy of day-to-day predictions of liver tumour position using implanted gold markers as surrogates and to compare the method with alternative set-up strategies, i.e. no correction, vertebrae and 3D diaphragm-based set-up. Twenty patients undergoing stereotactic body radiation therapy (SBRT) with abdominal compression for primary or metastatic liver cancer were analysed. We determined the day-to-day correlation between gold marker and tumour positions in contrast-enhanced CT scans acquired at treatment preparation and before each treatment session. The influence of marker-tumour distance on the accuracy of prediction was estimated by introducing a method extension of the set-up error paradigm. The distance between gold markers and the centre of the tumour varied between 5 and 96 mm. Marker-guidance was superior to guiding treatment using other surrogates, although both the random and systematic components of the prediction error SD depended on the tumour-marker distance. For a marker-tumour distance of 4 cm, we observed σ = 1.3 mm and Σ = 1.6 mm. The 3D position of the diaphragm dome was the second best predictor. In conclusion, the tumour position can be predicted accurately using implanted markers, but marker-guided set-up accuracy decreases with increasing distance between implanted markers and the tumour.