The challenge

An outdoor apparel direct-to-consumer brand had grown its top line for six straight years on the back of an increasingly aggressive promotional calendar. Quarter over quarter, more SKUs were spending more days at a discount. Full-price sell-through had drifted from healthy to anemic, and the executive team could feel the margin compression in the P&L without being able to point at the specific lever causing it.

The merchandising team’s instinct on any given pricing decision — drop a promo, widen the free-shipping threshold, test a bundle — was sharp, but the feedback loop was broken. Decisions shipped catalog-wide. There was no holdout. The readout, when it came, was a YoY revenue comparison muddied by seasonality, weather, paid-media spend, and a competitor’s promo cycle. Nobody could say with a straight face which moves had earned their margin and which had bled it.

The deeper problem was cultural. The brand’s price had become indistinguishable from its discounted price. New customers were trained to wait for a promo. Loyal customers had figured out the cadence. Competing on price was eating the thing the brand had spent a decade building.

The approach

We treated the work as a pricing-experiment program, not a one-time analysis. Three things had to land together for the program to compound:

• A registry. Every proposed price move — list change, promo depth, promo duration, free-shipping threshold, bundle composition, threshold-discount mechanic — gets logged with a hypothesis, a primary metric, a guardrail metric, and a pre-declared duration. No registry entry, no rollout. The discipline alone surfaced about a third of “we always do it this way” moves that nobody could defend on second look.
• Randomization at the right grain. Instead of A/B at the visitor level (which leaks across sessions and devices) we randomized at the SKU-region grain, with holdouts sized for the merchandising team’s smallest-effect-of-interest. The randomization layer sits between the storefront and the warehouse, so the experimental price is the price the customer sees, the price the order ships at, and the price the warehouse settles on — no reconciliation drift between the test and the books.
• A margin-impact readout the merchandising team actually owns. We built one report, refreshed daily, that takes a registered experiment and emits contribution margin lift, full-price sell-through delta, units-per-order delta, and the guardrail metric, with confidence bands. The merchandising team reads it themselves; they do not file a ticket and wait three weeks for an analyst.

We seeded the program with twelve experiments in the first six weeks: four list-price tests on hero SKUs, four promo-depth tests across categories the team had been promoting on autopilot, three free-shipping-threshold tests, and a bundle test on a slow-moving accessory category.

The results

The first quarter was the noisy one. By the end of the second quarter the program had shipped 22 experiments — versus the historical baseline of three meaningful pricing decisions per quarter — and the catalog-wide contribution margin had moved 4.2 points. The mechanics of that move were not subtle:

• Promo depth on the tested categories came down 18 percent on average, because shallower discounts beat deeper ones almost everywhere we tested.
• Full-price sell-through climbed 11 percent across the catalog as the calendar tightened around moves that had earned their place.
• The free-shipping threshold tests confirmed the team’s quiet suspicion that the existing threshold was below the margin-optimal point on three of the four tested cohorts.
• The bundle test produced the only clear loser of the quarter — and the discipline to kill it before it shipped was, the VP of Merchandising told us, the moment the team trusted the platform.

Time from hypothesis to readout collapsed from roughly six weeks to nine days, which is the change that makes the program self-sustaining. Experiments per quarter is now a metric the team manages to.

What we would do again

We would build the registry before we built anything else. The temptation in a pricing-program engagement is to start with the randomization layer, because it is the technically interesting piece. The registry is what makes the program survive the first executive who wants to ship a promo without one. It is also the artifact that, six months in, lets a new analyst understand why every active price is the price it is. We would also size holdouts honestly on day one; the team’s instinct is always that the smallest effect worth knowing about is bigger than it actually is, and oversizing tests early would have cost real money.