To address the issues of high labor and time costs, neglect of geographic spatial interactions, and coarse handling of high-dimensional unstructured features in existing retail chain store site selection methods, this paper proposes a smart site selection model (SmartSite) that integrates multi‑source spatiotemporal knowledge graphs and graph neural networks (GNNs). Specifically, a multi‑source spatiotemporal database (MSSTDB) is first established to collect and standardize multi‑source spatiotemporal data. Subsequently, a multi‑source spatiotemporal knowledge graph (MSSTKG) is constructed to accurately extract spatiotemporal entities and their relationships, and a graph convolutional network (GCN) is employed to derive deep feature representations from geographic data. Finally, an attention mechanism is introduced to dynamically assign feature weights, thereby building an intelligent site selection model. The performance of the proposed model is validated through comparative experiments on benchmark datasets and ablation studies, and the model is further applied to a leading new retail chain enterprise's store location scenario. Experimental results show that, compared with traditional manual approaches and conventional data‑driven methods, the proposed model can effectively exploit the correlations inherent in geographic spatial data, significantly reduce site selection costs, and deliver substantial economic benefits, offering a reliable decision‑making reference for retail chain enterprises in store location planning.