Northwind cut checkout drop-off by 18%
A retailer with 2,400 employees moved every checkout change behind a flag and a test, and found the fix was a smaller form, not a redesign.

The problem
Northwind sells home goods online and in 140 stores. Checkout was the most valuable page on the site and the most feared. Every change went through a two week release train, and nobody could say with confidence whether the last redesign had helped or hurt.
The team had one analyst who built readouts by hand from a warehouse export. A test took six weeks from idea to answer, so the team ran two a quarter and argued about the rest.
What they set up
They started with the SDK, not the dashboard. The web and mobile apps both evaluate flags locally, so a decision costs under 1 ms and does not add a network call to the checkout path.
- Every checkout change shipped behind a flag, off by default.
- Metrics read straight from their warehouse, so finance and product used the same numbers.
- A guardrail on payment success rate and page load time on every test.
- A sample ratio check that blocks a readout when bucketing looks wrong.
The test that mattered
The team assumed the drop-off was caused by the shipping step. They tested a redesigned shipping page against the control for three weeks. It lost. A second test removed three optional fields from the address form, with nothing else changed.
We spent a quarter redesigning the wrong page. The test that won took an afternoon to build.
Product lead, Northwind
Results
The shorter address form cut abandoned carts by 18% against control, with payment success unchanged. Kerf's sequential testing let the team stop early once the result was clear, and the AI readout pointed to replay clips of mobile users stalling on the optional apartment field.
The team now runs about nine tests a quarter. The analyst reviews readouts instead of building them.
What changed in how they work
Release is no longer a calendar event. Engineers merge behind a flag, product turns it on for 5% of traffic, and the guardrails decide whether it goes further. The release train still exists for infrastructure work, but checkout moved off it.
More customer stories

Halcyon ships weekly instead of monthly
A health scheduling platform replaced a manual release process with gradual rollouts and automatic rollback.

Brightpath replaced three tools with one
An education platform consolidated flags, experiments and session replay, and the arguments about whose numbers were right stopped.

Copperline tested pricing without breaking trust
A B2B payments company ran its first pricing experiments with guardrails on churn and support volume.