Building a Home Listing Recommendation Engine

MHVillage had the largest inventory of manufactured home listings in the country, but their email marketing was one-size-fits-all. Every subscriber received the same weekly digest of new listings, regardless of their search preferences, location, or budget.

The result was predictable: low engagement rates, high unsubscribe rates, and a significant missed opportunity to connect buyers with the homes they were most likely to be interested in.

Matt was brought in to build a recommendation engine that would transform their email program from broadcast to personalized, driving higher engagement and ultimately more listing inquiries.

The core challenge was data fragmentation. User preference signals were scattered across multiple systems — search queries in the platform, saved listings in user profiles, geographic IP data, and email click behavior. None of these systems talked to each other in a way that could power real-time recommendations.

Additionally, the manufactured housing market has unique dynamics. Unlike traditional real estate where location is often the primary filter, manufactured home buyers frequently prioritize home features, community amenities, and price point — often across multiple geographic areas.

The technical infrastructure also presented challenges. The email system, listing database, and user profile system were all separate platforms with limited integration capabilities.

The recommendation engine was built on a preference scoring model that synthesized signals from multiple sources. Each user’s implicit and explicit preferences were combined into a weighted profile that updated in real time as they interacted with the platform.

The system used collaborative filtering — identifying patterns in what similar users had engaged with — combined with content-based filtering that matched listing attributes to user preference profiles. This hybrid approach handled both the cold-start problem (new users with limited history) and the long-tail problem (niche listings that didn’t have enough engagement data for pure collaborative filtering).

On the email side, the one-size-fits-all digest was replaced with dynamically assembled emails where every listing was selected specifically for that recipient. Subject lines, featured listings, and even send times were personalized based on user behavior patterns.

Email engagement lifted 3.4x within the first two months. Open rates, click-through rates, and time-in-email all increased dramatically as users received listings that actually matched their interests.

Listing inquiries — the metric that mattered most to the business — increased by 28%. Users who received personalized recommendations were significantly more likely to take the next step and contact a seller or community.

Unsubscribe rates dropped by over 40%, indicating that the personalization made the emails genuinely valuable rather than another inbox burden. The email channel went from a declining asset to one of the strongest drivers of marketplace engagement.

Personalization doesn’t require perfect data — it requires the right data architecture. Starting with the signals you have and building systems that improve as more data flows in is more effective than waiting for a perfect data foundation that never arrives.

The hybrid recommendation approach (collaborative + content-based filtering) proved essential for a marketplace with diverse inventory and user intent. Neither approach alone would have captured the full picture of user preferences.

Email remains one of the highest-ROI channels when done right. The difference between “email is dead” and “email is our strongest channel” often comes down to relevance — sending the right content to the right person at the right time.