What ranking decides
Retrieval finds candidates; ranking orders them. The order comes from combining many signals into a single score.
Signal families
- Textual relevance measures how well query terms match a document, often via BM25.
- Quality and authority capture how trustworthy or popular a document is, like link counts or domain reputation.
- Freshness rewards recent content for time sensitive queries.
- User behavior uses clicks and dwell time as feedback on what people prefer.
- Context includes location, language, and device.
Combining signals
Early systems used hand tuned weights. Modern systems learn the combination from labeled data, a technique called learning to rank. The model takes signal values as features and predicts a relevance score.
Avoiding over weighting
No single signal should dominate. A page stuffed with query terms but low quality should not beat a trusted, relevant page. Balancing signals is the core ranking craft.
Diagram
Key idea
Ranking blends textual, quality, freshness, and behavior signals into one score, increasingly learned rather than hand tuned.