Cross-Sell Operations for Cart and Product Pages — Curated Recommendations with an Admin UI

Introduction

"Goes well with this" suggestions have a real effect on average order value. Yet a common story goes like this: you switch on an automatic recommendation feature, the products it surfaces don't quite fit, and before long the slot is sitting there collecting almost no clicks.

On an apparel and gear EC site I worked on, we placed suggestion slots on both the cart page and the product detail page (PDP, the page that presents a single product), and built a system where a person decides what gets suggested, right from an admin screen. This article covers the overall picture and why we chose an operations-first design over full automation.

Why "People Pick" the Cross-Sell

There are places automatic recommendations can't reach

Recommendations derived from purchase history are powerful, but they have blind spots. A product that just launched has no purchase data behind it, so it never becomes a candidate in the first place. The moment you most want to push it is exactly the moment the data isn't there yet.

Specialized combinations are the same story. Deciding that "this jacket goes with this repair part and this inner layer" takes someone who knows how the products are actually used. Looking at data alone, you can end up with unrelated items that simply happened to be bought at the same time.

Curation: letting people who know the products choose

So we went with curation based on staff judgment. Outfit pairings, repair parts, related accessories — a person picks the combinations, reasoning that anyone buying this item will probably need that one too.

In one line, the difference is whether suggestions are decided by past data or by current product knowledge.

None of this means automatic recommendations are a bad idea. Let automation handle staple products and well-stocked categories, and have people override only where intent matters — that split is the practical one. What this topic covers is the operational machinery behind the "people pick" half.

Two Suggestion Slots and How They Fit Together

Shoppers are in different situations on the cart page and the product page, so the two slots serve suggestions differently.

The point of the table: the cart slot decides by rules about cart contents, while the product page slot decides by links set on each individual product.

Drawn out, it's a single path: suggestion data set by staff in the admin screen flows into two slots, and users add products to the cart straight from those slots.

Keeping the Operation Running

Even good suggestions go stale once they stop being updated, because what you want to recommend changes with the season, with sales, and with stock levels. That's why an admin UI that lets staff change settings themselves, without an engineer, is essential. When settings live inside the code, every change needs a development request — and eventually nobody updates anything.

Product page suggestions, though, run to large numbers, and editing them one at a time on screen isn't realistic. So we made it possible to export the current settings as CSV (a list-format file you can open in a spreadsheet), edit them in bulk, and import them back.

For building the admin screen itself, we used an AI agent (AI-assisted development). Honestly, the range of what a non-engineer can get built by describing the workflow they want in plain words is wider than I expected.

Summary

Cross-sell doesn't pay off through algorithmic accuracy alone. What matters is whether you can put human judgment to work and keep it running as a routine. Three points defined this design:

• Cover what automatic recommendations miss — new products and specialized pairings — with human curation
• Serve suggestions suited to each situation: the cart page and the product page
• Use an admin UI and CSV so staff can keep updating without strain

Each part is covered in detail in the three articles below.