TL;DR
- Recommendations lift engagement by surfacing the next relevant item before users go looking
- Three core approaches: collaborative filtering, content-based, and embeddings/hybrid
- A solid architecture has four layers: events, features, model, and serving
- Cold-start is solved with content signals and sensible fallbacks, not magic
- Measure CTR, dwell time, and retention: not just offline accuracy
Why Recommendations Boost Engagement
Every content product faces the same problem: users see a fraction of what is available and bounce when nothing obvious is next. A recommendation engine closes that gap by predicting the most relevant item for each user in context. This turns a single visit into a session. We work with publishers, marketplaces, and SaaS teams across the USA, UK, Canada and Europe. For them, well-tuned recommendations are one of the highest-leverage features available. They raise pages per session, time on site, and return visits without acquiring a single new user. The mechanics differ by product. But the goal is constant: reduce the effort between a user and the next thing worth their attention.
The Three Core Approaches
Collaborative filtering Recommends what similar users engaged with. Powerful once you have interaction volume, but blind to brand-new items and users. Content-based Recommends items similar to what a user already liked, using metadata and text. Works from day one and handles new items well. Embeddings / hybrid Represents users and items as vectors and blends both signals. The modern default: flexible, scalable, and discovery-friendly.In practice, embeddings have become the connective tissue. You encode each item, an article, video, or product, into a vector that captures its meaning. Do the same for user behaviour, then retrieve nearest neighbours with vector search. A hybrid model then combines that semantic relevance with collaborative signals, so popular-but-relevant items rise without burying fresh content. This is core machine learning work, and it is where most of the tuning effort lives.
A Reference Architecture
Layer 1Events
- Capture views, clicks, dwell time, saves, and conversions
- Stream through a log or queue with a stable event schema
- Store raw events for replay and offline training
Features
- Build item embeddings from text and metadata
- Aggregate user histories into profile vectors
- Serve features from a store that is consistent online and offline
Model
- Two stages: candidate retrieval, then precise ranking
- Retrieve with vector search; rank with a learned scorer
- Retrain on a schedule and validate before promotion
Serving
- Return ranked results within a tight latency budget
- Apply business rules: dedupe, diversity, freshness, filters
- Log impressions so today's serving trains tomorrow's model
The Cold-Start Problem
Cold-start is the recurring headache of every recommender: you cannot recommend based on behaviour you do not yet have. It shows up in two forms, and each has a different fix.
New items
A just-published article has no clicks yet. Content-based signals, text embeddings, category, tags, let it be recommended immediately, before any behaviour exists.
New users
A first-time visitor has no history. Fall back to trending and popular items, and capture a few onboarding preferences. Then use context such as device, location, and referrer until personal signals build up.
Build vs Buy
| Path | Best when | Trade-off |
|---|---|---|
| Buy / managed | Recs are not your differentiator; speed matters | Less control over relevance and data |
| Managed vector search + custom logic | You want ownership without running infra | You still build ranking and serving |
| Fully custom | Personalisation is core; data is proprietary | Highest cost and ongoing maintenance |
A pragmatic path is to start with a managed vector search service to prove value fast. Then grow into a custom hybrid model once relevance becomes a real differentiator. Explore how this fits a broader build on our services overview.
Measuring Success
Click-through rate (CTR)
The fastest signal that a recommendation is relevant. Useful, but optimise it alone and you invite clickbait: always pair it with a quality metric.
Dwell time
How long users stay with the recommended item. A strong proxy for genuine value that guards against shallow clicks.
Retention
The metric that pays the bills. Good recommendations bring users back. Measure whether exposed cohorts return more often than a holdout.
Validate offline with ranking metrics, but trust online A/B tests with a holdout group for the final call. Offline accuracy and real engagement frequently disagree. Across deployments in the USA, UK, Canada and Europe, the live test is always the tie-breaker.