Prompt
How do I recommend content on a homepage without manual curation?
Latest observation
To recommend content on a homepage without manual curation, use an automated ranking system that scores items per visitor and assembles the page dynamically.
Core approach
-
Collect signals
- User behavior: clicks, dwell time, likes, saves, shares, purchases, scroll depth
- Content metadata: topic, author, freshness, popularity, length, format
- Context: device, time of day, location, referral source
- Aggregate trends: globally trending, category trending
-
Build candidate pools
- New content
- Popular content
- Personalized content
- Editorial/priority content if needed
- Similar-to-last-viewed content
-
Rank items automatically Use a scoring model to order candidates for each user. Common ranking inputs:
- Relevance to user interests
- Recency/freshness
- Engagement likelihood
- Diversity/novelty
- Business priorities
-
Mix in exploration Don’t only show what the user already likes. Include some exploratory items so you can learn new preferences and avoid a filter bubble.
-
Apply constraints
- Avoid duplicates
- Limit too many items from one topic or creator
- Ensure freshness
- Filter out already consumed content
- Respect policy/safety rules
Common recommendation strategies
1. Popularity-based
Good for cold start. Show:
- trending now
- most viewed this week
- most saved Simple, but not very personalized.
2. Content-based
Recommend content similar to what the user has engaged with before.
- Use tags, embeddings, or topic vectors
- Strong when you have good metadata
3. Collaborative filtering
Recommend items liked by users with similar behavior.
- Great for personalization
- Needs enough interaction data
4. Hybrid system
Combine popularity, content-based, and collaborative filtering. This is usually the best practical option.
5. Learning-to-rank / ML ranking
Train a model to predict click or engagement probability from user/item/context features.
- Very flexible
- Best at scale
- Requires data and evaluation setup
A practical homepage architecture
A common setup:
- Candidate generation: retrieve a few hundred likely relevant items
- Re-ranking: score and sort with a model
- Business rules layer: enforce freshness/diversity/eligibility
- Fallbacks: if personalization is weak, use popular/trending items
Minimal formula example
You can start with a weighted score like:
score =
- 0.4 × predicted_interest
- 0.2 × recency
- 0.2 × popularity
- 0.1 × similarity_to_past_behavior
- 0.1 × diversity_bonus
Then sort by score and display the top N items.
Important design considerations
- Cold start: new users and new content need fallback strategies
- Feedback loop: log impressions, clicks, and skips
- A/B testing: compare ranking strategies with real users
- Metrics:
- CTR
- dwell time
- conversion
- retention
- content diversity
- long-term satisfaction
If you want the simplest workable version
Start with:
- Popular content by category
- Recency filter
- Basic personalization from past clicks
- Diversity rules
- A/B testing to improve it
If you want, I can also give you:
- a simple algorithm
- a machine learning architecture
- or a Python example for homepage recommendations.