Recommendation diversity: preventing the more-of-the-same trap in product carousels

Pure relevance optimization converges on sameness: ten near-identical products in a row. Why diversity is a ranking objective, not a nice-to-have, and how to build it in.

Why relevance alone converges on sameness

Collaborative filtering and embedding similarity both optimize for closeness: items near the viewed item in the latent space. The nearest neighbors of a product are, by definition, the products most like it. Rank purely by similarity and the carousel becomes an echo of the current page, which adds no information and no reason to click.

The business cost is real. Carousels full of near-duplicates get ignored, which trains the team to distrust recommendations entirely. The engine is not failing at relevance; it is failing at usefulness. Those are different objectives, and only one of them drives incremental revenue.

What diversity actually means in a carousel

Diversity is not randomness. It is coverage of the shopper's plausible intents within a small number of slots. For a shopper viewing a tent, a diverse carousel might include a footprint, a warmer sleeping bag, a camp stove, and one alternative tent, because the intent behind viewing a tent spans the whole trip. Each slot should earn its place by representing a different facet of what the shopper might want next.

Think of it as portfolio construction. A financial advisor would not put a client entirely in one stock no matter how good it looks; the recommendation equivalent is not filling every slot with the highest-scoring item. The marginal value of the fifth similar item is near zero, while the marginal value of the first complementary item is high.

How to build it into ranking

The standard approach is maximal marginal relevance: rank items by a blend of relevance score and dissimilarity to the items already selected. As each slot fills, the remaining candidates are penalized for resembling the chosen ones. The result is a carousel that stays relevant while spreading across categories, price points, and use cases.

Category and attribute caps are the simpler version and often enough. Limit any single category, brand, or price band to two slots in a ten-slot carousel, and the sameness problem mostly disappears. Caps are crude but transparent, easy to tune, and easy to explain to merchandisers who want to understand why the carousel shows what it shows.

Personalizing the diversity itself

Different shoppers want different amounts of variety. A mission-driven shopper comparing options wants depth in one category; an exploratory browser wants breadth across many. Session signals can distinguish them: repeated views within one category suggest depth, while bouncing across categories suggests breadth. The diversity parameter does not have to be global.

New visitors generally benefit from more diversity because you know less about their intent; the carousel has to cover more ground. Returning shoppers with rich history can handle tighter, more focused recommendations. Letting the exploration-exploitation balance shift with what you know about the shopper is personalization applied to the ranking strategy itself.

Measuring whether diversity helps

Judge diversity changes on carousel click-through and downstream conversion, not on relevance scores. A common failure mode is a diversity tweak that looks better to the team but performs worse with shoppers, usually because it diversified into irrelevance. The metric that matters is whether more shoppers find something worth clicking.

Also track category coverage in purchased recommendations. If 90 percent of recommendation-driven purchases come from the same category as the viewed item, your carousel is still an echo chamber no matter what the diversity parameter says. The purchase mix is the ground truth of whether the carousel is expanding the shopper's consideration set or just reflecting it.

Frequently asked questions

What is the more-of-the-same trap in recommendations?

It is what happens when a relevance-only ranker fills a carousel with near-identical products: five black dresses in a row, each barely different from the last. Click-through collapses because there is nothing new to discover. The carousel looks personalized but feels repetitive.

How do you measure diversity in a product carousel?

Common approaches include category or attribute coverage across the shown items and pairwise dissimilarity scores. Track diversity alongside click-through: the goal is the efficient frontier where relevance stays high and variety increases discovery. Most teams find their carousels are far less diverse than they assumed.

Does adding diversity hurt relevance and conversion?

Done badly, yes: random variety is just noise. Done well, diversity is constrained: candidates must clear a relevance bar before diversity re-ranking applies. The result is a carousel that is both relevant and interesting, which typically lifts engagement without sacrificing conversion.