Recommender systems utilize Graph Neural Networks (GNNs) to learn vectorized representations of users and items from user-item interactions for predicting recommendations. Recent methods improve recommendations by incorporating item-related entities through a technique known as the Collaborative Knowledge Graph (CKG). However, the theoretical foundation of entity integration remains underexplored, leading to unresolved challenges in maintaining two critical properties for GNN-based recommender models: Local Consistency and the Inclusion of Indispensable Entities. This paper addresses two key research questions: (1) Do CKG-based models align well with these requirements? (2) Can an alternative graph structure better integrate entities into recommender systems? To answer these questions, we analyze CKG-based models and prove their fundamental limitation: they fail to simultaneously satisfy both properties. To resolve this issue, we propose a novel graph structure, the Fusion Graph (FG). We prove a theorem that demonstrates FG-based models meet the requirements of recommender systems. The source code is available at https://github.com/wangyifeibeijing/FGN .