What is Ranking?
Ranking is the process of assigning a score to each candidate result and ordering them from best to worst. Beyond lexical relevance (e.g., BM25/phrases), modern stacks combine freshness, availability, rating, margin, popularity, personalization, and diversity.
How It Works (quick)
- Pipeline: Recall (keyword + vectors) → score with textual + business features → optional re-ranking/LTR → diversity/brand caps → render.
- Features: Text scores, phrase/proximity, attribute matches, price, rating, returns, stock, recency, click/conversion signals, size-in-stock.
- Controls: Hard filters first (ACL/region/OOS), then scoring; log “why this result.”
Why It Matters in E-commerce
- Trust & revenue: The first page shapes conversion and margin.
- Merchandising: Promote in-stock, high-quality items without breaking relevance.
- Speed to product: Better top-k reduces pogo-sticking and facet thrashing.
Best Practices
- Start with BM25F + phrase/bigrams; add LTR as data grows.
- Cap boosts (freshness/popularity) so noise can’t outrank relevance.
- Enforce OOS filtering and brand caps before scoring.
- Use category/locale models; footwear ≠ electronics.
- Measure: NDCG/P@k, CTR@k, conversion, margin/session, returns, and zero-results.
Challenges
- Position bias in labels, seasonality, data leakage, cold start for new SKUs, and latency budgets.
Examples
- Query “merino base layer” → rank in-stock, high-rating, low-return items higher while respecting phrase matches.
- Campaign: apply small, time-boxed boosts to new arrivals with decay.
Summary
Ranking decides which good results go first. Blend relevance with quality and business signals, keep hard rules early, cap boosts, and measure outcomes—not just clicks.
