Ask three vendors which recommendation algorithm is best and you will get three confident, contradictory answers. The honest truth is that there is no best algorithm — only the right fit for a given placement, catalogue and stage of data maturity. This article compares the three families you will actually encounter: collaborative filtering, content-based, and hybrid. The aim is to leave you able to ask the right questions of any vendor or engineer.
No maths, just the trade-offs that matter to the business.
The three families at a glance
- Collaborative filtering learns from behaviour: who looked at, bought, or rated what. It finds patterns across shoppers and products without knowing anything about the products themselves.
- Content-based learns from the products: their attributes, descriptions and images. It recommends items similar to ones a shopper has shown interest in.
- Hybrid blends both, so each covers the other’s blind spots.
Each makes a different bet about where the useful signal lives — in behaviour or in the products — and each fails in a characteristic way when that bet is wrong.
Collaborative filtering
How it works in plain terms
“People who interacted with X also interacted with Y.” The system spots that shoppers who bought a particular tent often also bought a specific stove, and recommends accordingly — without understanding what a tent or stove is.
Where it shines
- Discovery. It surfaces non-obvious pairings a human merchandiser would never think to connect.
- Basket building. “Frequently bought together” is collaborative filtering at its most effective, lifting average order value. We cover the wider tactic in Guided Selling and AOV.
- Scale. With enough traffic, it improves on its own as behaviour accumulates.
Where it struggles
- The cold-start problem. New products with no interaction history are invisible to it. New shoppers with no history get generic results.
- Sparse data. Long-tail products and lower-traffic stores do not generate enough signal for reliable patterns.
- Popularity bias. Left unchecked, it over-recommends best sellers and starves the long tail.
Collaborative filtering is the strongest choice once you have real behavioural volume and want discovery and basket lift.
Content-based recommendations
How it works in plain terms
“This product is similar to one you liked.” Similarity is computed from product attributes — category, colour, material, price band, description, sometimes the image itself.
Where it shines
- New products. A brand-new item with good attributes can be recommended on day one, no history needed. This directly addresses collaborative filtering’s cold-start weakness.
- “Similar items” placements. On a product page, “you might also like” is content-based at heart.
- Smaller catalogues and lower traffic, where behavioural signal is thin but product data is solid.
Where it struggles
- Over-similarity. It can trap shoppers in a narrow bubble — endless near-identical products with no genuine discovery.
- Data dependency. It is only as good as your product attributes. Thin or inconsistent data produces weak recommendations. This is why eCommerce Data Foundations matters so much here.
- No sense of taste. It knows two coats are similar; it does not know which one shoppers actually prefer.
Modern content-based systems often use embeddings to compute similarity by meaning rather than exact attribute matching — the same idea behind vector search.
Hybrid approaches
In practice, most mature retailers run a hybrid, because the weaknesses of one family are the strengths of the other.
Common ways to combine them
- Switching. Use content-based for new products and shoppers, then switch to collaborative filtering once enough behaviour exists. The simplest, most robust pattern.
- Weighted blending. Combine scores from both, weighting toward whichever has stronger signal for that product or shopper.
- Layered ranking. Generate candidates with one method, then re-rank with another plus business rules.
Why hybrid usually wins
- It handles cold start gracefully — content-based fills the gap until behaviour arrives.
- It balances discovery and similarity, avoiding both random noise and narrow bubbles.
- It leaves room for a business layer on top — margin, stock, strategy — which no pure algorithm provides.
The trade-off is complexity: more to build, tune and monitor. For most mid-market retailers the right path is a sensible hybrid, not a research-grade system.
Do not forget the baseline
Before any of this, popularity-based recommendations — best sellers, trending, new arrivals — are a strong, cheap baseline. They require no personalisation, work for anonymous visitors, and make an excellent fallback. Many teams overreach for personalisation when a well-placed best-sellers rail would already capture most of the value. Always keep popularity as your safety net.
How to choose for your situation
Work through these questions honestly:
- How much behavioural data do you have? Low volume favours content-based and popularity; high volume unlocks collaborative filtering.
- How fast does your catalogue turn over? Frequent new products make content-based handling of cold start essential.
- How clean are your product attributes? Content-based and embedding methods live or die on this.
- What is the placement’s job? Discovery and basket building lean collaborative; “similar items” leans content-based. We map placements to jobs in A Practical Guide to Personalised Recommendations.
- Can you maintain a hybrid? Be realistic about your team’s capacity to tune and monitor.
Measuring which actually performs
Whatever you choose, prove it the same way: A/B test against a holdout and measure incremental revenue, not attributed clicks. Different algorithms will win in different placements, and the only way to know is to test. Be wary of any comparison that relies on attributed revenue alone — it flatters every algorithm equally.
The bottom line
Collaborative filtering excels at discovery and basket building once you have traffic but stumbles on cold start; content-based handles new products and similarity but depends on clean data and risks narrow bubbles; hybrid combines them and is where most mature retailers land — over a dependable popularity baseline. The right choice follows from your data, your catalogue turnover and the job of each placement, validated by proper testing.
If you want help choosing and testing the right recommendation approach for your store, our AI Search and Recommendations team can guide the decision and measure the lift. Get in touch for a straight assessment.