How Does a Store Choose Which Products to Recommend?

Follow a recommendation from candidate selection through stock checks, ranking and a test of what shoppers actually found useful.

A bookseller and reader compare two books beside a row of titles.
A recommendation has to suit the reader and be available when the reader needs it.

A recommendation AI model can learn that readers who viewed one field guide often went on to buy another. It can still recommend a book the reader already owns or a title the store cannot ship. The path from that historical association to a useful suggestion has several distinct steps.

Build a short list before ranking it

A laptop displaying a graph with multiple lines on its screen sits on a wooden desk in front of an office window.
A catalog item left out at retrieval has no opportunity to win at ranking.

A large catalog is expensive to score item by item on every visit. One retrieval method finds items close to the shopper or query in a learned numerical representation. Another can supply recent bestsellers when browsing history is thin. Google's engineering guide describes candidate generation and the trade-offs between different ways of measuring similarity. Google’s recommendation candidate-generation guide

Check the offer that can actually be fulfilled

An Apple computer monitor displays a graph related to inventory management on a wooden desk in front of a window.
A catalog match, a local stock count and a delivery promise are three separate records.

Apply the shopper's explicit constraints to the actual variant and destination. A paperback and a hardcover may share reviews while having different prices and shipping dates. At serving time, check availability and keep a fallback list. Record whether an item disappeared because it was unavailable or because the ranker gave it a low score.

Count successes among people who saw the list

A computer monitor displaying a stock market graph on a wooden desk in an office setting.
Positions shape what people click and what the next training set contains.

A previously popular item gets more chances to earn another click. For an offline test, hold back later visits and compare with a popularity list. In a live test, assign the whole eligible visit, count fulfilled orders and returns, and watch catalog coverage. A better click rate among surviving results can coexist with fewer visits that receive any useful suggestion.

Try it with the evidence

Repair a recommendation list

Source notes. Fictional request for a field guide under $25, delivered by Friday. A costs $18 and arrives Monday. B costs $24 and arrives Thursday. C costs $27 and arrives Wednesday. The system ranked them A, C, B. Shipping is included.

Prompt. Use only the supplied offers. Return the eligible list, give one reason for each exclusion and name one outcome worth checking after showing the result.

Check your reading

B alone satisfies both constraints. A misses the deadline and C exceeds the budget. Check whether B was still available and arrived by Friday. A click alone leaves both questions open.

Illustrated field notes

Finding relevant candidates

Identify which candidates a search method might miss before ranking starts.

A woman is browsing through an online flower shop or nursery app on her iPad, surrounded by various items in a room with white walls and a window.
An item a shopper clicked once may have been a gift.
Customer Segmentation Infographic.
The customer's current request can outweigh an old segment assignment.
A hand holds a smartphone displaying an AR app showing a white Nike shoe.
Visual resemblance does not establish the same function or size.
A woman sitting on a couch or chair in a cozy living room, looking at her phone.
Search terms can reveal a constraint the browsing record missed.
A miniature shopping cart and a black calculator sit on a wooden table or bench in an outdoor setting.
Separate a plausible substitute from a bundle that requires both purchases.
Availability and ranking feedback

Ask whether an item was displayed, available, chosen and ultimately kept.

An illustration of a fruit stand with various fruits and prices displayed on signs.
A merchandising override should be visible in the ranking record.
A woman sits at her desk, focused on her computer screen and working with a pen in hand.
Track which useful products never receive a chance to be seen.