Rebuilding Search for a 20-Million-Listing Marketplace

Search relevance on a 20-million-listing marketplace had been tuned incrementally, by different people, for years — the kind of system where nobody fully understands why any given ranking weight exists anymore.

Starting from what "relevant" means to a buyer

Rather than re-tune the existing black box, we rebuilt ranking around three explicit signals — exact-match strength, seller responsiveness, and listing freshness — and made each one visible in an internal debug view so future tuning wouldn’t repeat the same drift.

Click-through on the first three results improved by double digits, and — the metric that mattered most to the business — time-to-first-message from buyers dropped, meaning people were finding what they wanted faster, not just clicking more.