Short-form comments on TikTok's Beauty & Personal Care videos are rich yet noisy signals of consumer sentiment. This study deploys a transformer-based pipeline that couples GPT-derived sentence embeddings with BERTopic to surface salient discussion themes. We scraped 3,912 videos from the 20 highest-revenue beauty influencers and retrieved 34,597 engagement-ranked comments posted between January and August 2024. After a GPT -powered normalisation step that expands slang, translates Malay abbreviations, and converts emoji to text, comments were embedded with OpenAI's text-embedding-3-small model, reduced via UMAP, clustered using HDBSCAN, and distilled into topics with BERTopic. Topic quality was evaluated with four coherence metrics (c<inf>v</inf>, u<inf>mass</inf>, c<inf>uci</inf>, c<inf>npmi</inf>) to ensure semantic consistency. The model revealed two coherent, non-overlapping themes: product-usage experiences (e.g., frequency of use, perceived results) and product-feature commentary (e.g., scent, packaging, variants). Ac<inf>v</inf> score of 0.620 indicates strong interpretability despite the brevity and informality of TikTok discourse. These findings show that embedding-based topic modeling can unearth actionable insights: influencers should foreground authentic usage outcomes, while brands can boost engagement by highlighting sensory cues in visual storytelling. The study demonstrates the viability of LLM -enhanced analytics for short social media text and provides a replicable framework for future TikTok commerce research, noting limitations and opportunities for multi-category, cross-regional, and multimodal extensions. © 2025 IEEE.