<![CDATA[E-commerce recommendation cards must explain why an item is related while maintaining a compact, consistent visual hierarchy. This study presents an evidence-grounded card framework and evaluates its retrieval inputs with the final TREC Product Recommendation 2025 topics, a 48,616-product corpus, 7,633 NIST judgments for 47 topics, and the SQID supplementary image-URL table. Fixed lexical methods generated separate top-10 complement and substitute rankings over the full corpus before assessor judgments were applied. Under the official scoring procedure, the relation-aware lexical method achieved an average nDCG@10 of 0.1840, with a complement nDCG@10 of 0.0562 and a substitute nDCG@10 of 0.3119. Its average difference from full-text TF–IDF was not statistically reliable (difference = 0.0111, 95% CI [-0.0090, 0.0321]). The complement-minus-substitute gap was -0.2557 (95% CI [-0.3151, -0.1937]), confirming that complementary recommendation remains the weaker relation. Requiring a confirmed supplementary image URL increased confirmed URL coverage among displayed items to 100% but reduced average nDCG@10 to 0.0138. The findings support relation-qualified reasons, confirmed-image and text-led fallback states, and conservative language for complementary suggestions. The framework specifies evidence and layout behavior; shopper responses and different verbalization methods require direct comparative evaluation.]]>
Copyrights © 2026
