Recommendation diversity: avoiding the filter bubble in product carousels
A recommender that only shows more of the same shrinks the shopper's world and your average order value. Pure relevance optimization converges on near-duplicates of what the shopper already viewed. Building diversity into the carousel keeps discovery alive, and discovery is where the margin lives.
How the filter bubble forms
Most recommendation engines rank by predicted click probability. The items most similar to what the shopper just viewed win, because similarity is the safest predictor. Run that logic across every slot and the carousel becomes ten variations of the same black sneaker. Clicks look fine in the short term, but the shopper never encounters the complementary products, the upgrades, or the categories they would have bought if they had seen them.
What diversity actually means here
- Category spread. Reserve slots for adjacent categories, not just the viewed one. A shopper looking at running shoes can use socks, insoles, and a gym bag; none of them will ever outrank another shoe on pure similarity.
- Price spread. Similarity ranking clusters around the viewed price point. Deliberately include one step up (the upgrade) and one accessible entry point (the add-on) in every carousel.
- Novelty injection. New arrivals and trending items deserve guaranteed exposure independent of their predicted score. Without it, the engine cannot learn what it never shows.
- Serendipity budget. A small fraction of slots, even five percent, allocated to high-quality surprises. Measure their long-term effect on repeat visits, not just immediate clicks.
Keeping diversity from hurting relevance
Diversity without guardrails feels random, and random carousels get ignored. The fix is constrained diversification: pick the top candidates by relevance, then re-rank for spread within that qualified pool. Every shown item still clears a relevance bar; the ordering just stops being a similarity contest. Set the diversity strength as a tunable parameter and test it against revenue per session, not click-through rate, because the whole point is that the diverse picks convert differently.
What not to do
- Do not diversify by showing the shopper things they already rejected. Diversity means new directions, not the items they scrolled past twice.
- Do not apply the same diversity level everywhere. The cart page can handle more adventurous picks than the product page, where the shopper's intent is narrowest.
- Do not measure diversity with clicks alone. A diverse carousel can win on revenue while losing on CTR, because the items it surfaces are higher value. Judge it on the money.
Reviewed
Published Sep 27, 2026.