Unsupervised Cross-domain Image Retrieval (UCIR) aims to learn discriminative and domain-invariant feature representations from unlabeled image data across different domains, which remains a challenging task due to the absence of annotations and the presence of domain discrepancies. In this paper, we propose a novel framework called Consistency Aware Representation Learning (CARL) to address these challenges. CARL integrates a self-distillation alignment module and a contrastive learning module enhanced by optimal transport. The self-distillation alignment module extracts feature embeddings from unlabeled data and achieves both intra-domain and cross-domain alignment, effectively narrowing the domain gap. Furthermore, we introduce ClusterOT, an improved optimal transport strategy that leverages K-means clustering to address spatial distribution imbalance, thereby enhancing the global discriminative capability of contrastive learning. Extensive experiments on three public UCIR benchmarks validate the effectiveness and superiority of our proposed method over existing state-of-the-art approaches.