Content Recommendations
Why recommendations matter
Section titled “Why recommendations matter”Visitors do not browse your entire site. They see a handful of pages, make a decision, and leave. If the most relevant content or product is three clicks away, most visitors will never find it. They bounce, and you lose the conversion.
Recommendations solve this by surfacing the right content or product at the right moment. Instead of relying on visitors to navigate your site structure, you bring relevant items to them based on what they are doing, what similar visitors have done, and what your data says they are likely to want next.
Done well, recommendations increase pages per session, reduce bounce rates, lift average order value, and keep visitors engaged longer. Done poorly — irrelevant suggestions, stale data, no fallback for new visitors — they become noise that visitors learn to ignore.
Content recommendations vs product recommendations
Section titled “Content recommendations vs product recommendations”Optimizely provides two recommendation flavors, each designed for a different use case.
| Content recommendations | Product recommendations | |
|---|---|---|
| What they recommend | Articles, guides, videos, pages | Products, variants, SKUs |
| Data source | Page views, reading behavior, content metadata | Purchase history, browsing, cart data, catalog attributes |
| Primary goal | Engagement and education | Revenue and conversion |
| Typical placement | Blog sidebars, article footers, knowledge bases | Product pages, cart, homepage, email |
| Managed by | Content teams and marketers | Commerce and merchandising teams |
Both types share the same underlying recommendation engine. The difference is the item catalog they draw from and the behavioral signals they use.
How the recommendation engine works
Section titled “How the recommendation engine works”The engine follows a three-stage pipeline: data collection, model training, and prediction serving.
Stage 1: Data collection
Section titled “Stage 1: Data collection”The engine collects behavioral signals from every visitor interaction:
- Page views — Which pages or products a visitor views
- Click events — What a visitor clicks on within a page
- Cart and purchase events — Items added to cart, purchased, or abandoned
- Session data — How long a visitor spends on each item, scroll depth, return visits
- Content metadata — Categories, tags, attributes, and descriptions of each item
This data flows into Optimizely Data Platform (ODP) in real time. The more data the engine has, the better its predictions become.
Stage 2: Model training
Section titled “Stage 2: Model training”The engine trains recommendation models on your collected data. Models are retrained regularly (typically daily) to reflect new content, changing visitor behavior, and seasonal patterns.
Three core algorithm families power the models:
Collaborative filtering finds patterns across visitors. If visitors A and B both read articles 1, 2, and 3, and visitor A also read article 4, the engine recommends article 4 to visitor B.
- Strengths: Discovers non-obvious connections, works across content types, improves with scale
- Weaknesses: Cold-start problem (new items with no interactions get no recommendations), popularity bias (popular items get recommended more)
Content-based filtering matches items by their attributes. If a visitor reads an article about JavaScript testing frameworks, the engine recommends other articles with similar topics, tags, or content structure.
- Strengths: Works for new items (no interaction data needed), explainable (“recommended because you read about testing”)
- Weaknesses: Limited to item similarity, cannot discover cross-topic interests
Hybrid models combine both approaches. They use collaborative signals when available and fall back to content-based matching for new items or visitors with limited history. Most production deployments use hybrid models.
Stage 3: Prediction serving
Section titled “Stage 3: Prediction serving”When a page loads and requests recommendations, the engine:
- Identifies the visitor (via cookie, login, or ODP profile)
- Retrieves the visitor’s behavioral profile
- Queries the trained model for the best items given the visitor’s context
- Applies business rules (exclude already-viewed items, enforce category diversity, respect editorial overrides)
- Returns a ranked list of recommended items
This happens in milliseconds. The API returns item IDs, titles, images, and relevance scores that your front-end renders as a recommendation widget.
Recommendation strategies
Section titled “Recommendation strategies”Different placements call for different strategies.
| Strategy | Algorithm basis | Use case | Example |
|---|---|---|---|
| Similar items | Content-based | ”More like this” on a product or article page | Articles with matching topics |
| Frequently bought together | Collaborative (co-purchase) | Cart page cross-sell | Accessories for the item in cart |
| Personalized for you | Hybrid (visitor profile) | Homepage or landing page | Items based on browsing history |
| Trending | Popularity-based | Homepage, category pages | Most-viewed items this week |
| Recently viewed | Session history | Persistent sidebar or footer | Items the visitor looked at today |
| Post-purchase | Collaborative (post-purchase patterns) | Order confirmation, email | Complementary products |
Each strategy can be configured with rules that override or boost algorithmic output — for example, always include a new product launch in trending recommendations, or exclude items the visitor has already purchased.
The cold-start problem
Section titled “The cold-start problem”Every recommendation system faces cold start: what do you recommend when you have no data?
New visitors (no browsing history): Show trending items, editorially curated picks, or content-based recommendations tied to the current page context. As the visitor interacts, switch to personalized recommendations.
New items (no interaction data): Use content-based attributes (category, tags, description) to place the item into existing recommendation flows. Boost new items in trending lists to gather initial interaction data quickly.
New sites (no historical data at all): Start with editorial curation and trending-based strategies. The engine needs a few weeks of traffic data before collaborative models produce meaningful results. Content-based models can work immediately if your content has good metadata.
How recommendations connect to the rest of Optimizely One
Section titled “How recommendations connect to the rest of Optimizely One”Recommendations do not operate in isolation. They integrate with several platform products:
- ODP provides the behavioral data that powers models and the visitor profiles that enable personalization
- CMS hosts the recommendation widgets on your pages, either through built-in blocks or API integration
- Commerce supplies product catalog data, pricing, and inventory status to product recommendations
- Experimentation lets you A/B test recommendation strategies (does “similar items” or “trending” perform better on this page?)
- Visitor groups can control which recommendation widget appears for which audience segment
Measuring recommendation effectiveness
Section titled “Measuring recommendation effectiveness”Track these metrics to evaluate whether your recommendations are working:
| Metric | What it measures | Healthy range |
|---|---|---|
| Click-through rate (CTR) | % of visitors who click a recommendation | 5-15% for content, 2-8% for products |
| Revenue per recommendation | Revenue attributed to recommendation clicks | Compare against baseline without recommendations |
| Coverage | % of your catalog that appears in recommendations | Above 60% (low coverage means recommendations are biased toward popular items) |
| Diversity | How varied recommendations are across categories | Depends on strategy — “similar items” should be focused, “personalized” should be diverse |