Mytheresa, 2024

Better filters from messy product data

Customers weren't using filters on large category pages, and many who did left soon after. The obvious fix was a new filter design. The real problem was the data behind it.

Role: Senior Product Manager
Timeline: 2023 to 2024
Scope: 300K SKUs, 1,000+ categories
Result: +2.3M filter interactions in 30 days, bounce 25% to 22.5%

Filters on a category page

The problem

Filters existed on every category page, but customers didn't trust them.

Filters customers ignored

On large category pages filter use was low, and a quarter of visitors on filtered pages bounced.

One attribute, three spellings

ERP, CMS and the shop system each defined the same values differently.

Filters that returned nothing

Some filters had almost no products behind them, which broke trust in the rest.

What I did

Trace every value to its source

I mapped where each filter value came from across ERP, CMS and the shop system, then agreed one definition per attribute with the teams that owned them.

A bar before a filter appears

Each category got a minimum coverage for every filter. Filters that couldn't meet it were removed, which was the hardest conversation: it looked like shipping less.

Pushed back on manual tagging

The proposal was to tag around 70K items by hand, about 6,000 hours of work, for filters with 2 to 5% click rates. I argued for AI tagging with human review instead.

Three paths we considered

Redesign the filter UI
Fastest to show, but a better interface on top of the same gaps.
Clean the whole catalog first
Right in principle, but months before any customer saw a change.
What we did
Fix category by category
Ship each category once its data meets the bar, starting with the highest traffic.

Results

More customers filtered, fewer left, and the coverage bar became the standard.

+2.3Mfilter interactions in 30 days, 15M to 17.3M (+15%)

25% to 22.5%bounce rate on filtered pages

300KSKUs across 1,000+ categories

What worked

Fixing the data before the design, so every release was something customers could rely on.
Removing weak filters, even though it looked like shipping less.

What didn't

No control group, so seasonality is hard to rule out.
Merchandising joined late, although they knew which attributes customers shopped by.

What I'd do differently

I'd set up a control group on day one and publish the coverage bar as a shared rule straight away, so other teams could apply it without us.