Letting users build the menu instead of guessing at it.

Card sorting and tree testing are two ends of the same question, and I needed both, not just one. Card sorting comes first: you hand people a stack of labels, the real category and subcategory names, and ask them to group the ones that belong together and name each group themselves. No page, no visual design, just how they'd organize it in their own heads. That gave me a rough shape for a new structure, built from actual mental models instead of the taxonomy the CMS had accumulated over the years.

Tree testing comes after, and it asks a different question: given a structure, can someone actually find a specific thing inside it? You strip the candidate structure down to plain text labels, no visual design at all, and give people a task. This wasn't a single pass either. Card sort, tree test, adjust, and repeat until the structure held up. Both agreed on the same underlying problem: the CMS's folder logic and the user's mental model had drifted apart over time.

There was a prior measurement of visits per category, and after launch we compared traffic per category. The less visible categories got more clicks, and the main categories kept priority positions to protect their traffic. The structure was essentially solved at launch and only needed minor adjustments afterwards.