This paper presents a scalable methodology for extracting actionable insights from skincare product reviews by combining topic modeling, semantic clustering, and generative language models. Using a dataset of more than one million reviews from Sephora’s catalog, trigram-based TF-IDF vectorization with Latent Dirichlet Allocation (LDA) uncovered latent topics, while HDBSCAN clustering identified coherent subgroups reflecting concerns such as hydration performance, packaging usability, scent, and texture. To synthesize these findings, Google’s Flan-T5 Large model generated structured recommendations classified as strengths, areas for improvement, and best practices. Quantitative evaluation included pseudo-perplexity scores obtained with RoBERTa, achieving a mean of 14.33 (SD = 14.77) and a median of 8.01, indicating high syntactic fluency and semantic coherence in most outputs. Additional metrics, such as lexical diversity and redundancy, confirmed the quality of the generated insights. Results demonstrate that the trigram + HDBSCAN configuration produced the most coherent and diverse clusters, outperforming alternative setups. In conclusion, the proposed framework effectively transforms large-scale unstructured feedback into fine-grained, interpretable recommendations, supporting decision-making in product development, marketing, and customer experience design within the skincare industry.