In the context of the long-tail scenario, models exhibit a strong demand for high-quality data. Data-centric approaches aim to enhance both the quantity and quality of data to improve model performance. Among these approaches, information augmentation has been progressively introduced as a crucial category. It achieves a balance in model performance by augmenting the richness and quantity of samples in the tail classes. However, there is currently a lack of research into the underlying mechanisms explaining the effectiveness of information augmentation methods. Consequently, the utilization of information augmentation in long-tail recognition tasks relies heavily on empirical and intricate fine-tuning. This work makes two primary contributions. Firstly, we approach the problem from the perspectives of feature diversity and distribution shift, introducing the concept of Feature Diversity Gain (FDG) to elucidate why information augmentation is effective. We find that the performance of information augmentation can be explained by FDG, and its performance peaks when FDG achieves an appropriate balance. Experimental results demonstrate that by using FDG to select augmented data, we can further enhance model performance without the need for any modifications to the model's architecture. Thus, data-centric approaches hold significant potential in the field of long-tail recognition, beyond the development of new model structures. Furthermore, we systematically introduce the core components and fundamental tasks of a data-centric long-tail learning framework for the first time. These core components guide the implementation and deployment of the system, while the corresponding fundamental tasks refine and expand the research area.