When Does Cross-Market Transfer Help? Evidence from E-Commerce Markets

Multi-country platforms can share interaction data across their stores. Past studies report average gains, not what transfers. This study separates product popularity from co-interaction patterns, using XMarket ratings from 68 Amazon country–category stores, full-catalog ranking, and tuned closed-form recommenders. Duplicated variant reviews, which inflate accuracy, are merged. Foreign co-interaction patterns agree with local ones beyond a degree-preserving null. They lose less cross-country agreement than popularity in 95% of store pairs (difference scale) and 90% (chance-corrected scale), but only partly for one popularity-adjusted measure. In retrospective splits, pooling foreign data into stores with complete local data gives item-based models a small gain (mean +0.017 normalized discounted cumulative gain, NDCG@10, over a median of 0.105). Validation-tuned weighting does not improve it. Under a calendar cutoff, the estimate, +0.006 [−0.008, +0.014], is inconclusive (weakly positive at earlier cutoffs). With scarce local pattern data, foreign data close a median of 0.55–0.71 of the accuracy gap retrospectively. Under the calendar cutoff, they close 0.60 [0.22, 0.80] to 0.79 [0.34, 1.14] (less with more local data). With only user histories missing, they close almost none. Launched products gain only retrospectively. Exploratory tests link geographic distance to one popularity measure, not to transfer gains.