10,000 queries were buried in a bucket. I used density clustering and LLM labelling so real user intents surfaced instead of noisy clusters. The result: clear categories and surprising, actionable pain points.