Rebuilding checkout for 900 stores

The storefront was a 2016 monolith turning in a 4.7-second LCP on the mid-range Android that most of the traffic actually ran on. Stock was cached for six hours, so customers regularly reserved a sofa that had left the warehouse that morning and found out two days later by email. Mobile carried 71% of sessions and converted at under half the desktop rate. Every previous attempt at fixing it had been another A/B test on button copy.

We rebuilt the storefront edge-rendered with catalogue data cached at the CDN on a five-minute stale-while-revalidate, and pulled stock out into a separate streaming request behind a Suspense boundary so a slow inventory lookup can no longer hold the page hostage. Checkout went from five steps to two, with address, delivery slot and payment resolved on one screen — the delivery slot is where people were abandoning, and it needed to be visible before payment rather than after it. The whole thing shipped behind a 50/50 split in four markets for six weeks. Two of the changes we were most confident about did nothing measurable, so we cut them before the rollout.

Mobile LCP came in at 0.9s at p75 and checkout conversion rose 23% against the held-out control. Reserve-and-fail incidents dropped 88% once stock resolved live instead of from a six-hour cache. The design system that fell out of the work now covers the storefront, the in-store kiosk and the delivery tracking app — which was not in scope, and is the part the client brings up first.