66.7% Add-to-Cart Surge: HeritageHunt's Vintage Recommendation Revolution
The Silent Killer: When Unique Items Sell Out
Context: For HeritageHunt, scaling to $2.5M monthly GMV exposed a fatal flaw: No alternative recommendations when unique vintage items sell out, leading to lost sales opportunities.
Their vintage fashion enthusiasts aged 25-45 craved serendipitous discovery. When a rare 1940s gabardine coat sold out, the self-developed system offered irrelevant contemporary alternatives – essentially telling collectors “you missed out, now browse generic jackets.” This wasn’t just a UX issue; it directly cratered sell-through rates and AOV as cross-sell opportunities vanished into thin air.
From Manual Curation to Vector-Powered Intelligence
To solve this, HeritageHunt deployed WooRec Private Deployment.
Note: Given their need for algorithmic sovereignty and complex vintage logic, Private Deployment provided the control required to out-innovate generic SaaS solutions.
The transformation followed a deliberate, three-phase architecture:
Phase 1: Beyond Keyword Matching
Powered by WooRec Strategy Module
We replaced rigid category filters with Semantic Vector Retrieval:
- Foundation: We seeded the system with Hot Retrieval to surface trending eras (e.g., “Y2K revival”), solving cold-start for new arrivals.
- Advanced: We deployed FAISS-based Vector Embeddings trained on:
- Historical item descriptions (e.g., “1920s flapper beading” vs. “1980s power shoulder pads”)
- User behavior sequences (view → cart → purchase patterns)
- Condition metadata (e.g., “mint,” “patina,” “structural wear”) The Logic: By mapping users/items to a 128-dimensional vector space, we could recommend “similar-but-different” pieces when originals sold out – like suggesting a 1930s cloche hat when a rare 1920s version was unavailable.
Phase 2: The Model Evolution (LR → DeepFM → ESSM)
Powered by WooRec Model Serving
To achieve the target 66.7% Add-to-cart Rate, we iterated through three architectural shifts:
- The Baseline (Logistic Regression): Initial models used linear features (price, era, material) but failed to capture interactions like “customers who buy distressed denim also seek vintage band tees.”
- The Upgrade (DeepFM): We introduced Deep Factorization Machines to model high-order interactions (e.g., how “1970s + leather + size medium” signals distinct preferences). This reduced sparsity issues with rare items.
- The Final State (ESSM):
- Why ESSM?: Vintage demand fluctuates wildly. ESSM (Entire Space Multi-Task Model) jointly optimized CTR (engagement) and CVR (conversion) to avoid recommending beautiful-but-unpurchasable items due to sizing/condition concerns.
Phase 3: Vintage-Aware Traffic Control
Powered by WooRec Rule Engine
We baked vintage-specific constraints into re-ranking:
- Diversity (Scatter/Shuffle): Implemented a “no consecutive same-decade” rule to prevent visual fatigue (e.g., blocking 3 1960s mini dresses in a row).
- Business Injection (Hard Insertion): Reserved slots 4 and 10 for high-margin “Consignment Curator’s Picks” to drive AOV.
- Dynamic Weighting: Boosted items with:
- Low Inventory Depth: Prioritizing slow-moving unique pieces
- High Condition Scores: Reducing returns by favoring items with detailed flaw documentation
The Vintage Discovery Experience
Here’s how these algorithms manifest on HeritageHunt’s storefront:
*Figure 1: Context-aware recommendations blending historical relevance, scarcity, and cross-sell logic.*The Impact: 66.7% Engagement Explosion
Within 6 months, HeritageHunt achieved:
> *Interactive Chart: Metrics pre/post-WooRec implementation showing steep positive curves for Add-to-cart, Conversion, and Retention.*- Add-to-cart Rate: 66.7% increase (12% → 20%)
- Return Rate: 40% reduction (25% → 15%)
- Customer Retention: 50% improvement (30% → 45%)
Customer Voice
“Moving from manual rules to ESSM was revolutionary. The system now understands that a customer seeking a 1960s mod dress might also want matching accessories – even if they’ve never viewed them. The 66.7% add-to-cart surge proves our collectors feel truly understood.” — Elena Rossi, Chief Technology Officer at HeritageHunt
Ready to Engineer Your Recommendation Revolution?
You don’t need a PhD in vector math to deploy enterprise-grade AI. With WooRec Private Deployment, HeritageHunt’s team configured a bespoke engine in weeks – not years.