2023
Verdant
Engagement up 3×. Time-on-platform up 2.4×. Churn down 18%.
The Challenge
Verdant is a B2C platform for independent creators. They had 180,000 monthly active users and strong initial retention — but engagement dropped sharply after the first two weeks. Users would sign up, explore the platform for a while, and then stop coming back. The hypothesis was that users weren't finding content relevant to them quickly enough. The platform had a lot of good content, but discovery was driven by simple chronological feeds and manual category browsing. Verdant needed a way to surface the right content to the right person at the right time.
The Solution
We scoped the problem carefully before writing any code. AI recommendations are easy to build poorly — systems that optimise for clicks rather than satisfaction, or that create filter bubbles that make the platform feel smaller over time. We started by defining what "good" looked like: not just engagement metrics, but diversity of content consumed, creator discovery (finding new creators, not just favorites), and explicit user satisfaction signals. Then we designed a recommendation model that balanced these objectives rather than optimising for a single number. The system we built ran on the existing infrastructure with minimal added cost. It used a lightweight collaborative filtering model for content recommendations, combined with a re-ranking layer that injected diversity and creator discovery signals. We A/B tested it against the control for six weeks before a full rollout. The results were better than the original hypothesis: engagement up 3×, time on platform up 2.4×, and churn in the 30-day window down 18%. Crucially, creator discovery was up significantly — a sign the system was growing the platform rather than just deepening existing usage.
