Click-through rate (CTR) prediction has been widely used in online recommendation systems. Most CTR prediction methods enhance performance by modeling complex feature interactions using well-designed network architectures. However, these methods still encounter challenges related to feature distribution imbalance, parameter distribution imbalance, and parameter optimization imbalance. Such imbalances result in suboptimal feature representation, ultimately degrading prediction performance. To address these issues, we propose a novel framework based on contrastive learning, termed Alignment-Uniformity Aware Representation Learning for CTR Prediction (AU4CTR). AU4CTR is compatible with existing CTR prediction models as a model-agnostic representation learning framework. It directly improves feature representation quality by optimizing two core properties, alignment and uniformity, with three auxiliary contrastive losses: the intra-field alignment loss for assigning similar representations to the features in the same field, the inter-field uniformity loss for maximizing the representation distance across different fields, and the interaction alignment loss for aligning the original interaction representation with its perturbed counterpart. Finally, extensive qualitative and quantitative analyses on six public datasets demonstrate the proposed AU4CTR can enhance the feature representation and show remarkable effectiveness and compatibility when integrated with various representative CTR methods.