Rebuilt the merchandising engine for a D2C consumer brand, lifting AOV by $14.20

The client's merchandising function had ossified. Product recommendations on the PDP, cart, and post-purchase email were all driven by a hand-maintained spreadsheet of 'goes-well-with' pairings that their head of merch had been updating personally for three years. Return rate on recommended products was 22% — six points above the store average — suggesting the rules were aging poorly as the catalog grew from 400 to 2,100 SKUs.

Leadership had been pitched four different 'AI personalization' platforms, each quoting $240k-$480k annually plus 90+ days of integration. The platforms all looked similar in demos and none could explain how they would handle the client's actual data: a catalog where size curves varied wildly by subcategory and where 34% of revenue came from 'drops' that had no historical signal at all. We were brought in to sanity-check the procurement decision.

Our recommendation was to not buy a platform. The client had three things those platforms needed to be useful — a Shopify Plus data stack, 26 months of clean clickstream, and a merchandising lead willing to be the domain expert — and building internally would leave them with the underlying system as an asset rather than a recurring license.

We built in three layers. First, a two-tower retrieval model on clickstream + catalog metadata, which handled the 66% of revenue that had meaningful historical signal. Second, a cold-start layer for new drops that used catalog embeddings (image + description + price band) to bootstrap from the closest historical analogs. Third, a business-rules overlay the merchandising lead could edit directly — because some decisions (this is the hero SKU for the season, this sold-out size is not coming back) are facts the model has no way to learn.

The A/B framework ran against a 15% holdout from week 5 through week 12. We instrumented not just conversion and AOV but return rate, because recommending the wrong product to the wrong person is worse than recommending nothing.

Over the 8-week A/B window, the treatment arm posted $14.20 higher AOV and 11.3% higher revenue per session against the holdout. More importantly, return rate on recommended items fell from 22% to 12.9% — the old rules had been pushing volume at the cost of fit. Projected annual revenue lift at current traffic levels is $3.8M against a total engagement cost under 8% of one year's savings. The merchandising lead was freed from roughly 40 hours per month of spreadsheet maintenance and is now using that time to stand up two new product categories.