How do I rank search results and recommendations together?

To rank search results and recommendations together, treat them as candidate items from different sources and score them in a shared ranking layer.

Common approach

1) Generate candidates from both sources

• Search: items matching the user query
• Recommendations: items from collaborative filtering, content-based models, trending, etc.

2) Normalize into one candidate set

Give each item a common representation:

• item_id
• source = search or recommendation
• features:
  • query relevance
  • user-item affinity
  • popularity
  • freshness
  • personalization signals
  • diversity / business rules

3) Use one scoring model

Train or define a model that predicts a unified objective such as:

• click-through rate
• conversion probability
• engagement value
• utility / expected value

Example: [ score = w_1 \cdot relevance + w_2 \cdot personalization + w_3 \cdot freshness + w_4 \cdot popularity + w_5 \cdot business_rules ]

4) Calibrate scores across sources

Search and recommendation scores often live on different scales, so normalize them before merging:

• min-max scaling
• z-score normalization
• Platt scaling / isotonic calibration
• per-source score calibration

Without this, one source can dominate unfairly.

5) Add source-aware features

Let the ranker know where the item came from:

• is_search_candidate
• is_recommended_candidate
• query intent strength
• explicit vs implicit intent

This helps the model learn when search should dominate versus when recommendations should.

6) Apply re-ranking for mix quality

After initial scoring, adjust the final list for:

• diversity
• novelty
• deduplication
• source balance
• business constraints

For example:

• if query intent is strong, show mostly search results first
• if query is vague or empty, show more recommendations

---

Practical ranking strategies

A. Single blended ranker

Best when you want one unified list.

• Combine all candidates
• Train one model
• Rank by predicted utility

Pros: simplest user experience
Cons: needs good calibration and training data

---

B. Two-stage with source blending

1. Rank search results separately
2. Rank recommendations separately
3. Merge with a blending rule

Example:

• top 5 search results
• top 3 recommendations
• interleave if scores are close

Pros: easy to control
Cons: less optimal than a true unified ranker

---

C. Intent-aware mixing

Use query context to decide the mix:

• strong navigational query → mostly search
• exploratory query → mix search and recommendations
• no query / home feed → mostly recommendations

This often works very well in production.

---

Important data requirement

To train a unified ranker, you need logs of:

• what candidates were shown
• their source
• their positions
• clicks, conversions, dwell time, etc.

Be careful about position bias and source bias in training data.

---

A simple production recipe

If you want something practical:

1. Retrieve top N from search
2. Retrieve top M from recommendation systems
3. Normalize scores per source
4. Train a learning-to-rank model on combined features
5. Re-rank with diversity and business rules
6. Monitor CTR, conversion, and source mix

---

If you want a very simple formula

You can combine them like this:

[ final_score = \alpha \cdot normalized_search_score + (1-\alpha) \cdot normalized_rec_score ]

Where (\alpha) depends on intent:

• high query intent → larger (\alpha)
• low query intent → smaller (\alpha)

---

If you want, I can also give you:

1. a concrete ranking architecture,
2. a machine learning feature set, or
3. pseudocode for merging search and recommendation results.