People Pick the Suggestions — Admin UI and CSV Bulk Operations

Introduction

You build a cross-sell slot, and six months later it's still showing the same products. This is a familiar state on EC sites. The cause is usually not a lack of motivation but the setup itself: updating requires a development request, or the sheer number of entries makes it impossible to keep up.

This article introduces the admin UI (the screen where staff edit settings) and the CSV-based bulk workflow we built for an apparel and gear EC site. CSV here means a list-format file you can open in a spreadsheet. One screen for careful one-at-a-time fixes, and CSV for handling thousands at once — how we split the work between them is the main topic. The underlying question is how to design an operation that doesn't stop getting updated.

What Changes When People Pick the Suggestions

Product knowledge goes straight into the suggestions

A judgment like "this jacket goes with this repair part and this inner layer" can only come from someone who knows how the products get used. Even pairings that look unrelated in the data can be chosen accurately by that person. The main value of a human-picked approach is that the knowledge in a staff member's head goes directly into the suggestion slot.

Pairings that automatic extraction can't reach — accessories experienced users favor, or a new product you deliberately want to push — can be added freely once a person is choosing.

You can reflect what you want to push right now

Products you need to move quickly, high-margin products, the season's headline item. Selling intent like this shifts constantly. With a human-picked setup, that intent can go into the slot immediately. Getting an algorithm to do the same tends to turn into a large design exercise in conditions and weights.

The difference in one line: whether suggestions rest on past data or on current selling intent.

Combining automatic and manual

That said, not everything needs a person. On sites with large catalogs, the practical approach is to let automatic recommendations handle staple categories and have people override only the products where intent matters. What this article covers is making that override something staff can sustain.

What the Admin Screen Does

Linking suggestions to each product

Product page cross-sell is configured product by product: for each item you set the products to suggest alongside it. In the admin screen you search for the target product, then add the products you want suggested. Order is configurable too, so the item you most want seen goes first. The first one or two slots get looked at far more than the rest, which makes that ordering control worth having.

Keep it fixable the moment you think of it

The purpose of the admin UI is to hand operations back to the staff running the store. When settings live inside the code, every seasonal changeover and every sale needs a development request. Once a few days pass between request and release, updating drops down the priority list, and eventually nobody touches it.

Being able to fix it yourself, right when you think of it — that's what keeps suggestions current. For what it's worth, the admin screen itself was built with the help of an AI agent (AI-assisted development). The range of things you can work out just by describing the workflow you want, right down to the shape of the screen, turned out to be wider than expected.

Handling Thousands of Entries with CSV

Editing one at a time hits a wall

Once a catalog passes a few hundred products, the total number of links runs into the thousands. Editing those one at a time on screen simply takes more hours than anyone has. A task like "add the standard repair part as a suggestion to every product in a newly added category" can eat a full day of clicking.

So we set up a workflow where the current settings are exported as CSV, edited in bulk in a spreadsheet, and imported back. Bulk replacing and copying across many rows while seeing the whole picture is something spreadsheets do better than a web screen. That same full-day task finishes in a few minutes.

Four stages: export, edit, import, confirm. What happens in the middle is entirely up to the spreadsheet.

Validation before import prevents accidents

Bulk import is fast, which also means mistakes spread fast. A product ID that doesn't exist, or a product that's already been discontinued, will produce an empty suggestion slot on the live site if it passes through unchecked.

So on import we validate whether each product ID actually exists and whether the format is correct, rejecting problematic rows. Before committing, you see how many entries will be updated and how many are errors. Error rows are reported by line number, so they can be fixed in the spreadsheet and re-imported.

Making it a routine

CSV operations only deliver value once they're established as a routine the staff actually follow. Keep the export, edit, import cycle as simple as possible, and keep things in a state where re-importing the original CSV puts everything back. With "you can undo a mistake" as a given, staff start running updates on their own judgment.

The table comes down to one split: the admin UI for a few entries right now, CSV for many entries at once.

Summary

Cross-sell quality comes down to the people picking the suggestions and the tools supporting them. Three points defined this system:

• Center on human curation that carries product knowledge and selling intent
• Keep suggestions current with an admin UI that works without relying on engineers
• Update large volumes of links via CSV, with validation before import to prevent accidents

How the slots are presented, and the UI for adding items, are covered in the articles below.