Scaling an E-commerce
Marketplace
Sole designer shaping long-term vision and growth features, driving $45M in net new & retained sales in one quarter.
ML is The Product
They help the model quickly adapt and personalize recommendations or content for new users when there is limited historical data, crucial for cold-start problems.
Referral source
These are immediate signals captured from a user's initial interaction with the product, such as first clicks, searches, or behaviors within the first session.
Weather-based suggestions
e.g. Rainy season → AWD carsHot climates → cars with cooling features like ventilated seats
Geolocation
Popular cars near you (e.g. Texas <> Trucks).
Short-term signals capture trending behaviors, preferences, and contextual data, allowing the model to react dynamically to immediate shifts in user behavior, like changes in preferences or engagement patterns.
Abandoned sessions
Re-engage users who viewed cars but didn’t proceed further
Session Frequency
e.g. "You’ve been browsing daily – here are today’s fresh arrivals"
Saved @ favorited
e.g. “Here’s a similar deal".
Viewed listing
Re-engage users who viewed cars but didn’t proceed further
Long-term signals help in maintaining personalization and accuracy for returning users, ensuring that the model incorporates stable, ongoing trends and habits.
FAQs
Addressing protection plans, financing, and trade-in processes
Continue purchase
Modify the homepage to focus exclusively on the user’s in-progress purchase
Support prompts
Offer proactive support with FAQs or live chat to prevent cancellation
Vehicle resale value
e.g. "Your car’s value has increased by 8% in the last year – consider trading in for an upgrade"
Make the homepage dynamic, as it serves as the entry point for 92% of our customers, optimizing engagement and tailoring content based on user behavior.
Specific
Design a new “new user” experience focused more on the account/sign in.
Measurable
Redesign a user workflow each spring.
Relevant
Yes, ladders up to a company wide goal.
Achievable
Yes, the resources are available, however, true high fidelity is best done on Adobe products which is not used amongst the team.
Time-bound
By the end of Q3.
Zero-day signals
Before users created an account, we had limited data about their activity, so we relied on broad signals like geo-location to enhance the experience.
This allowed us to show cars popular in their region or highlight nearby dealerships, giving relevant suggestions even with minimal interaction.
These broad signals helped make the browsing experience feel locally tailored and engaging, even before the user logged in.
Short-term signals
Long-term signals op & book
After a user made a purchase, long-term signals kicked in.
We used machine learning to anticipate what they might need post-purchase by dynamically presenting FAQs tailored to their recent transaction, helping them with things like registration, insurance, or maintenance.
We also adjusted the homepage to focus on actions that would help them finalize or continue their purchase, ensuring they wouldn’t cancel or be distracted by browsing other cars.
The goal here was to guide them to finish their purchase and build loyalty by offering the right support at the right time.
.avif)
Deposit made
3x More Purchase Completion
2x Less Exits
3x More Visits
Attributed to release
Support Calls
Unique new acts.
High intent feature usage
.webp)
