Unlearning in large language models (LLMs) is intended to remove the influence of specific data, yet current evaluations predominantly rely on token-level metrics such as accuracy and perplexity. This chapter demonstrates that such metrics can be misleading: models often appear to forget, but their original behavior can be rapidly restored with minimal fine-tuning, indicating that unlearning may conceal information rather than truly erase it. To address this issue, this chapter introduces a representation-level evaluation framework that incorporates PCA-based similarity and shift, centered kernel alignment, and Fisher information. Applying this toolkit across six unlearning methods, three domains (text, code, and mathematics), and two open-source LLMs reveals a critical distinction between reversible and irreversible forgetting. In reversible cases, models experience token-level collapse yet preserve latent representations; in irreversible cases, deeper representational degradation emerges. The chapter further provides a theoretical explanation linking shallow weight perturbations near output layers to misleading unlearning signals, and shows that reversibility is influenced by task type and hyperparameter choices. These findings expose a fundamental gap in current evaluation practices and establish a representation-level diagnostic foundation for more trustworthy unlearning in LLMs.