Amazon Abandoned-Cart Reminder Emails: ML Signals and Levers to Lift Purchases

Amazon wants to use machine learning to decide how to email shoppers who have left items sitting in their shopping cart. The email reminds the shopper about the cart, and the point of the model is to target these reminders better so that more carts end in a purchase. As the product manager working with the machine learning team, answer two questions: what would you take into account when designing this algorithm, and which variables would you adjust to raise the rate at which carts turn into purchases? ### Clarifying Questions - How is the cart purchase rate defined: the share of reminded carts that check out, measured over what window, and does a purchase that would have happened without the email count? - Is the target more completed orders, more revenue or more margin, and may the email carry an incentive such as a discount? - How many of these emails may one customer receive, and does the reminder share a frequency budget with Amazon's other marketing email? - Does the algorithm cover every customer with items in the cart in every country, or a subset to start? ### Part 1 — What the algorithm should take into account List the factors that should shape the design: the inputs the model would use, what it should predict, and any rules that sit outside the model. For each factor, explain why it bears on whether a reminder leads to a purchase. ```hint Not every abandoned cart is the same Carts are left behind for very different reasons, and a reminder helps with some of them far more than others. What data would let the model tell those situations apart? ``` #### What This Part Should Cover - Signals about the cart, the customer, the moment the cart was left and past email response, each tied to a reason it predicts a purchase - The likely reasons a cart stalls and how each changes whether and how to remind - What the model predicts and what data it is trained on - Eligibility and exclusion rules enforced outside the model ### Part 2 — Variables to adjust for a higher cart purchase rate Once a first version is live, describe the variables you would tune to raise the cart purchase rate and how you would decide whether each change worked. ```hint Each email is a bundle of choices Break one reminder into its separate decisions: who receives it, when, how many times and what it contains. For each, ask what it costs when you turn it up too far. ``` #### What This Part Should Cover - Levers across targeting, timing, frequency, content and incentives - How each lever is tested and what it risks - The metric that decides whether a change helped, with its guardrails ### What a Strong Answer Covers - A precise objective that counts purchases the email actually caused, not clicks or the raw conversion of emailed carts - A coherent framing that connects the send decision, the label and the training data - A measurement plan built on a randomized holdout, with guardrails for unsubscribes, spam complaints, margin and long-term engagement - Customer trust: honest content, consent and frequency limits, and care with what a cart can reveal ### Follow-up Questions - The model sends to the customers most likely to buy, and the purchase rate among emailed carts climbs, yet total orders barely move. What went wrong and how would you fix it? - Would you ever include a discount in the reminder? How would you decide who gets one without teaching customers to leave carts on purpose? - How would you extend the system to push notifications and in-app messages without flooding customers?