A good merchandiser does not sort a category page arbitrarily. They have a mental model, developed over years of watching what customers buy and what they bring back. They know which products are gateway items: things that first-time buyers try and then come back for more of. They know which items cluster together in customer carts. They know which products look good in photos but generate returns because the in-hand quality does not match expectations.
That knowledge is not written down anywhere. It lives in the merchandiser's judgment. And most personalization systems, as currently designed, cannot encode it.
I want to explain why, because understanding the gap is the first step toward building a model that actually captures what an experienced buying team knows.
What a Merchandiser's Decision Actually Contains
When a senior merchandiser decides to front a specific product on a category page, they are integrating multiple signals simultaneously. Some of them are quantitative: this item sold at 3x the rate of comparable SKUs last quarter. Some are qualitative: this item is new but it fits the purchasing pattern of our most consistent buyer segment. Some are negatives: this item has an 18% return rate, which means it is not as good as its click metrics suggest.
The most experienced merchandisers also think about trajectory, not just current state. A product that is selling moderately today but is bought repeatedly by high-lifetime-value customers looks different in their analysis than a product with a high initial spike and low repeat purchase rate. The spike product is attractive. The repeat-purchase product is valuable. Those are different categories of commercial outcome.
When I started thinking about what the CFY model needed to encode, the merchandiser's decision framework was the reference architecture. Not "what is clicking well right now" but "what has this shopper proven they will pay for, and what will the next logical thing in their trajectory be."
Why Standard Recommendation Models Miss This
Collaborative filtering, the foundational approach behind most recommendation engines, says: "Shoppers who bought A also bought B, therefore recommend B to anyone who bought A." At scale, with millions of transactions, this produces useful results. At the product-catalog level of a mid-size retailer, it tends to surface obvious pairs and miss the trajectory logic that a good merchandiser would apply.
The collaborative filter does not distinguish between a customer who bought A once and a customer who has bought A five times in two years. Both get the same "also bought B" recommendation. But the experienced merchandiser would treat these shoppers very differently. The first-time buyer of A might be exploring. The five-time buyer of A has demonstrated a preference pattern that should inform a different set of recommendations.
Purchase frequency and return behavior are two signals that standard collaborative filtering either ignores or underweights. They are also two of the signals that experienced merchandisers use most heavily. Return behavior in particular is underutilized. A shopper who returned Item X after buying it is giving you explicit negative signal: this item did not meet their actual need, regardless of how well it fit the purchase-pattern prediction. Building that negative feedback into the model matters.
The Three Inputs That Encode Merchandiser Logic
After working through several iterations of the CFY model design, we landed on three inputs that most closely encode what experienced merchandisers are actually tracking.
The first is purchase trajectory: what has this shopper bought, in what order, over time. The order matters. A shopper who started with entry-level gear and progressively bought more technical items is in a different place in their relationship with a product category than a shopper who bought high-end items first. The trajectory tells you where they started and where they are heading.
The second is return-adjusted purchase history. Every item a shopper returned should be treated as a negative data point, with proportional weight based on how recently it happened. A return from two years ago has decayed significance. A return from last month is a strong current signal. This is what a merchandiser does intuitively: they remember what came back and they weight that information in their category placements.
The third is co-purchase density: how strongly an item clusters with the items this shopper has bought before. Not generic collaborative filtering, but purchase-dense co-occurrence within this shopper's specific purchase history and the history of shoppers whose trajectory most closely matches theirs. This is narrower and more precise than site-wide co-occurrence patterns.
An experienced merchandiser looking at a customer file would synthesize these three signals in about thirty seconds. The model does the same thing, just at a per-session scale that no human buying team could match.
What the Model Cannot Replace
There are things an experienced merchandiser does that no personalization model currently replicates, and it is important to be honest about that boundary.
Merchandisers make forward-looking brand decisions. They pull a product not because the data says to, but because it does not fit where the brand is heading. They introduce new products into positions that data would not warrant because they believe in the item and want to give it visibility. These are editorial choices that require an understanding of brand strategy, supplier relationships, and long-term category direction that is not contained in transaction data.
A personalization model like CFY should augment these decisions, not override them. The merchandiser's editorial judgment about which products belong in the catalog and what the brand should stand for is upstream of anything the model does. The model operates within the catalog the buying team has curated. It optimizes placement within that catalog for individual shoppers. It does not define the catalog.
We built the CFY configuration to honor this distinction. Merchandisers can pin specific products to featured positions that the model will not displace. They can create boost rules for new arrivals or strategic items they want to surface regardless of purchase-signal ranking. The model then operates around those editorial anchors, personalizing everything else. The result is a system where the merchandiser's strategic judgment sets the frame and the model fills in the per-shopper details.
Building the Training Set from Merchandiser Intuition
One of the more interesting challenges in building a personalization model for retail is that the ground truth for "good recommendation" is often held in a merchandiser's head, not in any training dataset. How do you encode a standard that is not written down?
The approach that worked best for us was indirect. Rather than trying to extract and formalize merchandiser rules, we look at the outcomes that experienced merchandisers have already produced: which items they chose to feature, which items they chose to suppress, which products they have repeatedly promoted versus buried over multiple seasons. These historical placement decisions contain the accumulated judgment of the buying team even if the reasoning was never documented.
This means the model is partly trained on your own merchandising history, not just on customer purchase behavior. The buying team's category decisions over the past 18 months are implicit labels for which products have performed according to merchandiser expectations and which have underperformed relative to placement. Incorporating that signal into the model means the personalization layer is not starting blind. It is starting with an encoding of the institutional knowledge your team has already accumulated.
The Signal That Most Correlates with Long-Term Value
After working through the data from our early-access retailers, the single signal that most reliably correlated with long-term shopper value was repeat purchase of items in the same sub-category within 90 days. Not just repeat purchase, but quick-repeat purchase within a specific category area. This pattern appeared to indicate a shopper who had found a product that genuinely met a recurring need, not just a one-time purchase of something that looked interesting.
Experienced merchandisers know this intuitively. They watch for replenishment signals. They know that a customer who comes back for the same type of product three times is more valuable than a customer who bought three different things once each. The model can detect this pattern systematically across your full customer base, not just the accounts a merchandiser happens to remember.
That is the real value of encoding merchandiser logic into a model: not replacing the judgment, but applying it at a scale that no team can match manually. The buying team's intuitions are the design specification. The model is the implementation that makes those intuitions work for every shopper, every session, every page view.