# What are hybrid recommenders?

> A hybrid recommender combines more than one technique, usually collaborative filtering, content, context, and rules, arranged in stages. Retrieval narrows a large catalog to a shortlist, a heavier model ranks it, and re-ranking applies diversity, inventory, and business rules. Almost every production recommender and search ranker works this way.

_Last updated 2026-11-01. From The Personalization Guide, https://personalization.guide_

A hybrid recommender is a recommendation system that combines more than one technique, most often
collaborative filtering, content-based filtering, context, and hand-written rules, arranged in stages.
Almost every production recommender and search ranker is a hybrid, because no single method is strong
enough on its own.

## Why hybrids win

Each pure technique has a built-in weakness. [Collaborative filtering](/library/techniques/collaborative-filtering)
needs no understanding of the content, but it cannot say much about a new user or a new item, the
cold-start problem, and it leans toward popular items because that is where the data is.
[Content-based filtering](/library/techniques/content-based-filtering) handles new items and is easy to
explain, but its results tend to be too similar to each other. A hybrid lets one method cover the
other's blind spot: content features carry a brand-new item until it has enough behavior for
collaborative filtering to take over, and context keeps the results from collapsing into sameness.

## The three stages

Production systems almost never score a whole catalog in one pass. They run a funnel, cheap and broad
at the top, expensive and precise at the bottom.

### Retrieval

Retrieval pulls a few thousand plausible candidates from a catalog that may hold millions. It is built
to be fast and forgiving rather than exact, so it uses cheap signals: recent behavior, item
similarity, embeddings, and simple filters. Its only job is to make sure the good answers are somewhere
in the shortlist.

### Ranking

Ranking scores that shortlist with a heavier model that can afford to be slow, because it only ever
sees a few thousand items rather than the whole catalog. This is where most of the accuracy comes from,
and where features about the person, the item, and the context are combined.

### Re-ranking

Re-ranking adjusts the top of the ranked list for things the ranking model does not optimize directly:
diversity, so the results are not ten versions of the same thing; inventory and freshness limits; and
business rules such as margin or a promotion. It is the last pass before the results are shown.

## Why staging matters

Staging buys two things at once. It makes the system both fast and accurate, because cheap retrieval
narrows the field and the expensive model only ever scores a shortlist. And it gives each concern a
place to live. A merchant can enforce a margin rule in re-ranking without retraining the ranker, and a
new diversity requirement does not mean rebuilding retrieval. The parts stay loosely coupled, which is
what makes a large recommender maintainable.

## Where you see them

Product recommendations and site search are the clearest examples, but the same funnel shapes feed
ranking, home-page module selection, and related-item rails. The pattern is worth remembering: a
recommender is rarely one algorithm. It is a pipeline of them, each doing the part it is good at. Which
combination works for a given catalog and audience is a question only
[measurement](/measurement) can answer, and every stage depends on
[the data and identity layer](/data) underneath.