Geo-Personalized Ranking: Unlocking Regional Conversion Gains

Retailers operating in multiple regions often default to one ranking algorithm for all markets. The upside is obvious: more data to train on, simpler infrastructure, and consistent experiences across geographies.

The downside is subtle but important: it's possible for regional shopping behaviors to get averaged out. If one market dominates the data set, its preferences tend to overshadow smaller markets, pushing down items that might otherwise perform well locally.

Monica Vinader, a British luxury jewelry brand with a strong UK base and a growing US audience, faced exactly this tension. Together with Constructor, they tested whether region-specific ranking models could unlock more revenue without hurting performance in their core market.

The Problem: One Global Model, Two Distinct Markets

Monica Vinader sells contemporary fine jewelry in both the UK and the US, but shoppers in each region have notably different tastes and price sensitivities.

Their existing product-ranking algorithm combined data from both markets, weighted toward the UK (their larger market). That boosted the data volume for training but masked US-specific patterns, particularly around higher-value items that US shoppers favored but were often pushed down in rankings.

The Hypothesis: Dedicated Regional Models Could Surface Local Winners

Together, the Monica Vinader and Constructor teams suspected that running a dedicated model per region — rather than a single global model with tweaks — would let the algorithm learn each market’s product affinities independently through reinforcement learning, thereby surfacing more relevant items to each audience. (With reinforcement learning, user feedback improves the models by reinforcing the results that shoppers click on and buy in every given context, such as where they’re geographically located.) So, they designed an experiment to close the gap.

Key bet: a dedicated US model would lift conversion by showcasing higher-value pieces more prominently, without harming UK performance.

The Test: A 4-Week 50/50 Geo Split

• Control (A): Existing global model trained on pooled UK + US data — personalizing 1:1 to each user, but learning across all geographic data pooled together
  
  
• Variant (B): Geo-personalized approach (two separate models trained on region-tagged data, automatically serving the UK model to UK traffic and the US model to US traffic)
  
  
• Goal metric: Revenue per browse user and purchase conversion
  
  

The Results: US Conversion Up, UK Steady

• +5% lift in US conversion rate vs. control
  
  
• No change in UK conversion (steady performance, as intended)

The results confirmed that the new approach (together with reinforcement learning) improved US performance without disrupting what worked well for the UK. The uplift grew even stronger during Monica Vinader’s seasonal sale, when the new ranking model surfaced higher-value items that matched local preferences.

Takeaways for Retailers

1. When markets diverge, consider giving them their own brain. Tweaks to a single global model may not be able to capture fundamentally different shopper preferences.
   
   
2. Validate no-harm scenarios. Keeping the UK flat was as important as lifting the US, proving that the new approach scaled responsibly.
   
   
3. Expect seasonal amplification. Localized ranking helped the model react better to spikes in higher-ticket items during peak sales events.
   
   
4. Let models learn by doing. With reinforcement learning, each iteration enables the system to adapt more quickly and deliver increasingly relevant results with each cycle.

Want to learn more? Read the full case study.