# What is content-based filtering?

> Content-based filtering recommends items whose attributes or embeddings resemble what a person has already engaged with. It reads the content rather than the crowd, so it handles new items and explains itself, but its results drift toward sameness.

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

Content-based filtering recommends items whose attributes or embeddings resemble the ones a person
has already engaged with. It reads the content itself, a product's category and brand, an article's
topic and entities, a song's genre and tempo, rather than the behavior of other people, and it matches
new candidates to a profile built from what this person has clicked, bought, or read.

## How it works

Every item is described as a set of features: structured attributes such as category, price band,
author, and tags, or a learned embedding, a vector that places the item in a space where similar
things sit close together. The system builds a profile of the person from the items they have engaged
with, often by averaging or weighting those item features, and then recommends the unseen items
closest to that profile. Because the match is item-to-profile, it can score a brand new item the moment
it has a description, with no waiting for anyone to interact with it first.

## What it is good at

Two strengths follow directly from reading the content rather than the crowd.

- It handles new items well. A product added this morning can be recommended this afternoon, because
  its attributes are known even though no one has bought it yet. This is the opposite of collaborative
  filtering's cold-start problem for new items.
- It is easy to explain. The system can point to the shared attribute: "because you read this article
  about interest rates" or "more from this author." Explanations like these are cheap to generate and
  tend to build trust.

It also needs no data about other users, so it can personalize for a person in isolation, which matters
when the audience is small or when privacy constraints limit cross-user data.

## Where it falls short

Its central weakness is the mirror image of collaborative filtering. Because it recommends what
resembles the past, results tend to be too similar to each other, a problem usually called
overspecialization. A person who reads one article about interest rates gets ten more just like it, and
never discovers the adjacent topic they would have loved. It cannot surface the pleasant surprise that
comes from "people like you also enjoyed this," because it does not look at other people at all.

It is also only as good as its features. If the item descriptions are thin, inconsistent, or missing,
the similarity it computes is weak. Rich embeddings help, but they can encode biases present in the
content, and they are harder to inspect than a plain tag.

## How it is used in practice

Content-based filtering is rarely the whole system. Most production recommenders are hybrids that use
it alongside collaborative filtering, so that content similarity covers the new and long-tail items
where behavioral data is thin, while collaborative signals add the serendipity that content alone
misses. A common pattern is to let content-based similarity generate candidates and a behavioral model
rank them, or to feed content features into a single ranking model. Which combination works for a given
catalog is a question only measurement can answer, and it depends on the data and identity layer
underneath.