Humans Pick the Suggestions — Admin UI and CSV Bulk Operations

Editing per-product suggestions in an admin UI, with CSV import/export for bulk maintenance

管理画面CSV運用キュレーション一括登録運用
5 min read

Introduction

The quality of cross-sell suggestions changes dramatically with "who decides them, and how."
If humans are going to pick the part algorithms can't fully handle, they need tools that let them work comfortably.

This article introduces the admin UI and CSV-based bulk operations we built for an apparel and gear EC site. The theme is how we use two things — a screen for editing carefully one at a time, and CSV for handling thousands at once — for their respective strengths.

The Value of Humans Picking

Reflecting product knowledge directly into suggestions

Judgments like "this jacket goes with this repair part and this inner layer" can only be made by someone who knows the products. Even combinations that look unrelated in the data can be chosen accurately by someone who knows how they're actually used. The greatest value of human curation is this ability to carry tacit knowledge into the suggestions.

"Connoisseur" combinations that auto-extraction can't pick up, or a new product you deliberately want to push, can be reflected freely once a human is choosing.

Being able to control intent

Products you want to move quickly, high-margin products, the star item of the season — what you "want to push right now" changes day to day. A system where humans choose lets you reflect that moment-to-moment intent in the suggestions — a flexibility that's hard to bake into an algorithm on purpose.

Differences by How Suggestions Are Decided
BEFORE
Auto-extraction only

Based on past data, so intent is hard to embed. New products and stock adjustments are hard to reflect.

AFTER
Human-picked operation

Product knowledge and sales intent reflected directly. Deliberate suggestions, changeable when you want.

Designing the Admin UI

Linking suggestions per product

For PDP cross-sell, you set the "products to suggest together" for each product. In the admin screen, you search for the target product, then pick and arrange the products you want to suggest. You can specify the order too, so the thing you most want seen goes first.

It's crucial that staff can edit this directly, without going through an engineer. When configuration is buried in code, updates stall and the suggestions go stale.

Updatable without relying on engineers

The point of the admin UI is to return operations to the hands of staff. Filing a development request for every season change or sale slows things down, and the settings tend to get neglected in the end. Keeping things in a "fix it yourself the moment you think of it" state is the condition for keeping suggestions fresh.

Bulk Operations with CSV

Why CSV

With many products, the total number of links becomes enormous. Editing them one at a time in the screen isn't realistic. So we made an operation where you export the current settings as CSV, edit them in bulk in a spreadsheet, and import them back.

The strength of a spreadsheet is being able to bulk-replace and copy across many rows while keeping the whole picture in view. A task like "add the standard repair part as a suggestion to every product in a new category" finishes in minutes.

Bulk Maintenance with CSV
Export

Write out the current suggestion settings as CSV

Edit in a spreadsheet

Add, delete, and reorder in bulk

Import

Take in the edited CSV and overwrite the settings

Validate and apply

Check for non-existent product IDs, etc., before committing

Validation on import

Bulk import is powerful, but that also makes it prone to accidents. Specifying a non-existent product ID, or an already-discontinued product, results in an empty suggestion slot being shown. So on import we validate whether product IDs actually exist and whether the format is correct, rejecting problematic rows. Before taking anything in, you can confirm "how many will be updated, and how many are errors."

Establishing it as an operational flow

CSV operations only deliver value once they're established as a routine for staff. Keep the flow of export, edit, and import simple, and keep things in a state where you can revert from the original CSV if something goes wrong. This "safe to touch" design supports continuous maintenance.

Conclusion

The quality of cross-sell comes down to the people picking the suggestions and the tools that support them. The key points of this system were:

  • Center on human curation that reflects product knowledge and sales intent
  • Keep suggestions fresh with an admin UI that's updatable without relying on engineers
  • Run large volumes of links in bulk via CSV, preventing accidents with validation on import

For how suggestions are presented on the cart page, see Designing "one more item" cart cross-sell; for size selection inside the suggestion slot, see Variant-aware cross-sell UX. For the big picture, see the cross-sell operations hub article.