PersonalizeIQ answers

Your Recommendations Only Show 4 Percent of the Catalog. That Is a Merchandising Problem.

Pull the logs for a month of recommendation impressions on a mid size catalog and a pattern appears that surprises teams the first time they look. A few dozen products take most of the impressions. The rest of the catalog, including products with stock, margin, and good product page conversion, is effectively invisible to the recommendation system. The model is not broken. It is doing what it was trained to do: show what has sold, which ensures that what has sold is what keeps selling. Left alone, the loop narrows the store a little more every quarter.

Why models concentrate

Popularity is the strongest signal in most training data, and it is self reinforcing. A product shown often gets clicked and bought, which generates the data that gets it shown again. New products enter with no history and lose every comparison against incumbents with thousands of sessions behind them. Seasonal and niche products suffer the same way: their demand is real but sparse, and sparse data reads as weak demand to a model that cannot tell the difference. The result is a catalog that looks, to a returning visitor, smaller and more repetitive than the store the merchant thinks they run.

Coverage is a metric, so track it

Report the share of in stock products that received at least one recommendation impression in the last 30 days, and the share of impressions taken by the top 50 products. Those two numbers, reviewed monthly, turn a vague worry into a managed quantity. Set a floor for exploration: a small, fixed slice of rail slots reserved for new arrivals, high margin lines the merchant wants to push, or products with strong product page conversion but low traffic. Accept that the exploration slice will click worse this week. Its job is to generate the data that lets good products compete next month, and to stop the catalog quietly shrinking.

Diversity inside a single rail

Coverage across a month is one problem. Sameness inside one rail is another, and visitors see that one. Eight near identical products in a row, same category, same colour family, same price, read as a system that ran out of ideas. Cap repeats by category and attribute, mix a complementary item into a similar items rail where the template allows it, and suppress products the visitor has already bought in categories nobody rebuys weekly. These are rule level fixes a merchandiser can own without a data science roadmap, and they change the feel of the rails immediately.

Let merchandising set the guardrails

The durable arrangement is a division of labour. The model ranks within the space it is given. Merchandising sets the space: which categories may appear together, which products are excluded and why, what share of slots is reserved for new and strategic lines, and which campaigns temporarily override the model and for how long. When that division is missing, every trading priority becomes a request to retrain the model, and every model update becomes a surprise to the trading team. Write the guardrails down, review them with the coverage numbers, and the argument about why the rails look the way they do gets much shorter.

Keep reading

Frequently asked questions

What is catalog coverage in recommendations?

The proportion of sellable products that actually appear in recommendation impressions over a period. Low coverage means the system is concentrating on a small set of proven sellers.

Why do new products never appear in recommendations?

They have no click or purchase history, so models trained on past behaviour rank them below established products. An exploration allocation, a reserved slice of slots, is the usual fix.

Does adding diversity hurt conversion?

In the short term an exploration slice usually clicks less than proven sellers. Over time it surfaces products that convert well and widens the effective catalog, which protects revenue when hero products stock out or fade.