<p>Susceptibility modeling is valuable in combating landslides, one of the natural disasters that cause significant adverse impacts worldwide. The most important data for modeling landslide susceptible areas are inventories containing records of past landslides. How to sample these inventory data and include them as dependent variables in the susceptibility model is controversial. In this study, we aimed to evaluate the effects of different sampling strategies on landslide susceptibility models. Eight different grid-based sampling strategies were presented in assessing shallow landslides in a mountainous area. We modeled landslide susceptibility using these sampling strategies. Extreme gradient boosting, a machine learning method, and ten factors (elevation, slope, plan curvature, profile curvature, topographic position index, topographic wetness index, stream power index, topographic roughness index, distance to streams and roads) were used for modeling in the Çiftekavak basin (Turkey). We found that sampling from the highest 25% grids of landslide areas was the most successful strategy. The comparison of the produced models revealed that the change in sampling strategies significantly affects the susceptibility models. It is strongly recommended that the most accurate sampling strategy be preferred when modeling landslide susceptibility.</p>