At Kindly, a grocery delivery service, people were taking eleven minutes to place a first order. Most of those minutes were spent looking for things.

The work

• Search ranking rebuilt with the data team: purchase history first, then popularity, then text match
• Saved lists and one-tap reorder for the eighty percent of baskets that repeat
• A discovery playbook, so the next feature did not start from zero

The number

Time to first order fell from eleven minutes to five. Reorder became the most used feature in the app within a quarter.

People do not want to search. They want to have found.

What I would do differently

Build reorder first. It was the smallest change and it moved the most.

How it ran

The data team owned the ranking model and I owned the question it answered. We met every Tuesday with the same three charts: time to first order, searches with no result, and reorders as a share of baskets. Every change to the model was shipped behind a flag to ten percent of people first, and most of the changes that felt obvious in the meeting lost in the flag. The ones that won were small and boring, which is how we knew they were real.