<p>Landslides are natural disasters that cannot be prevented. However, identifying susceptible areas and implementing early warning mechanisms in these areas can help reduce the damage they can cause. In this context, landslide susceptibility maps play a critical role in determining the spatial distribution of risky areas. Artificial intelligence (AI) approaches, such as artificial neural networks (ANN), are increasingly being used to produce these maps due to their superior ability to learn nonlinear relationships between multidimensional data. However, their performance strongly depends on the selection of conditioning factors. This study developed a systematic framework integrating multiple feature selection algorithms with the ANN to rigorously evaluate the impact of factors on landslide susceptibility maps. Twenty commonly used conditioning factors were ranked by each feature selection method and successive subsets derived from these rankings were used to train and test the ANN models. The performance metrics indicated that the choice of factors significantly affects model behavior. The highest values were obtained using random forest (RF) rankings. Relief-F produced competitive accuracy even with relatively small subsets, whereas mutual information (MI) and information gain ratio (IGR) produced more variable outcomes across different subset sizes. This study demonstrates that dimensionality reduction not only increases accuracy but also improves model interpretability. Furthermore, it moved beyond a single case study by providing reproducibility within ANN-based landslide susceptibility modelling. Therefore, the study’s findings provide scientifically based guidance for developing land use decision-making processes and disaster risk reduction policies.</p>