Configure Recommendation Models

Why model configuration matters

A recommendation engine is only as good as its configuration. The default settings will produce results, but they will not reflect your business priorities. A fashion retailer wants recommendations that respect seasonality. A B2B knowledge base wants recommendations scoped to the visitor’s product line. A media site wants recommendations that balance recency with relevance.

Model configuration is where you translate business intent into algorithmic behavior. This guide shows you how to set up models, choose algorithms, and apply the rules that make recommendations work for your specific scenario.

What you will do

1. Create a recommendation model
2. Select the right algorithm for your use case
3. Configure business rules and content filters
4. Test and validate the model output

Create a recommendation model

A model defines a recommendation strategy — what to recommend, using which algorithm, subject to which rules.

1. Navigate to Optimizely > Recommendations > Models
2. Click Create Model
3. Give the model a descriptive name that reflects its purpose (e.g., “Product Page - Similar Items” or “Blog Footer - Related Articles”)
4. Select the item catalog the model draws from:
   • Product catalog for product recommendations
   • Content catalog for article/page recommendations
5. Save the model to proceed to algorithm selection

Naming convention tip: Use the format [Placement] - [Strategy]. This makes models easy to identify when you have several running across your site.

Select the right algorithm

Each model uses one algorithm. Choose based on what you want to optimize.

Algorithm	How it works	Best for	Data requirement
Collaborative filtering	Recommends items that similar visitors engaged with	Established sites with high traffic	10,000+ interactions
Content-based	Recommends items with similar attributes (tags, categories, text)	New sites, long-tail content, cold-start scenarios	Good content metadata
Hybrid	Combines collaborative and content-based signals	Most production use cases	Some interaction data + good metadata
Trending	Ranks items by recent popularity	Homepages, category pages, new visitor fallback	1 week of traffic data
Co-purchase	Recommends items frequently bought together	Cart pages, post-purchase emails	Purchase event data

To set the algorithm:

1. Open your model
2. Navigate to Algorithm Settings
3. Select the algorithm type
4. Adjust the algorithm-specific parameters:
   • Collaborative filtering: Set the similarity threshold (higher = more relevant but fewer results) and the lookback window (how far back to consider visitor behavior)
   • Content-based: Select which content attributes to match on (categories, tags, description text)
   • Hybrid: Set the weight balance between collaborative and content-based signals (e.g., 70/30)
   • Trending: Set the trending window (24 hours, 7 days, 30 days)

Configure business rules

Business rules override or adjust algorithmic output to reflect business priorities.

Inclusion and exclusion rules

Control which items can appear in recommendations:

Rule	Purpose	Example
Exclude categories	Prevent items from certain categories	Exclude “clearance” from premium product pages
Include only	Restrict to specific categories or tags	Only recommend from the same product line
Exclude viewed	Skip items the visitor already saw	Prevent showing the current article
Exclude purchased	Skip items the visitor already bought	Avoid recommending a product they just ordered
Date filter	Only recommend items published or updated within a window	For news sites, only show articles from the last 30 days

To add rules:

1. Open your model
2. Navigate to Business Rules
3. Click Add Rule
4. Select the rule type and configure its parameters
5. Set the rule priority (rules are evaluated in order; the first matching rule wins)

Boosting and pinning

Sometimes you want the algorithm to consider items but give certain ones a nudge:

• Boost — Increase the relevance score of items matching a criterion (e.g., boost new arrivals by 20%)
• Pin — Force a specific item into a specific position (e.g., always show the latest product launch as the first recommendation)
• Diversify — Ensure recommendations span multiple categories rather than clustering in one

Cross-Product Integration

Recommendations

Boosting and pinning rules are evaluated after the algorithm generates its initial ranking. This means the algorithm still determines relevance, but your business rules adjust the final output. Use pinning sparingly — too many pins defeat the purpose of algorithmic recommendations.

Fallback strategy

Configure what happens when the model cannot generate enough recommendations (new visitor, insufficient data, or too many exclusions):

1. Trending fallback — Fill remaining slots with trending items
2. Editorial fallback — Show a curated list you maintain manually
3. Empty — Show nothing (useful when irrelevant recommendations are worse than none)

Set the fallback in the model’s Fallback Settings section.

Test and validate the model

Before deploying a model to your live site, verify it produces sensible results.

Preview mode:

1. Open your model
2. Click Preview
3. Select a test item (a product page or article) as context
4. Review the recommended items the model returns
5. Try different test items across categories to check diversity

Validation checklist:

Check	What to look for	Action if it fails
Relevance	Do recommended items relate to the context item?	Adjust algorithm weights or attribute matching
Diversity	Are recommendations spread across categories?	Add a diversity rule
Freshness	Are stale or outdated items appearing?	Add a date filter rule
Exclusions	Are excluded items properly filtered?	Verify exclusion rule configuration
Fallback	What shows for a brand-new visitor?	Configure trending or editorial fallback
Coverage	Are niche items ever recommended, or only popular ones?	Lower the popularity bias in algorithm settings

After validation:

• Publish the model to make it available for embedding
• Monitor performance metrics (CTR, revenue per recommendation) for the first two weeks
• Iterate on rules based on real-world data — no model is perfect on the first configuration

Common issues

Issue	Cause	Fix
Model returns no results	Insufficient behavioral data	Wait for more traffic, or switch to content-based algorithm
Same items always recommended	Popularity bias in collaborative filtering	Add diversity rules or boost underrepresented categories
Irrelevant items appearing	Metadata quality issues	Improve content tagging and categorization
Results not updating	Model retraining not running	Check the retraining schedule in model settings

Related content

Continue your journey with these related pages.

📖Understand the concepts

• Content Recommendations

🔧Follow a guide

• Embed Recommendations in Your Site
• Configure Visitor Groups