Objective: use normalized product data to power search that understands users and merch rules.
1) Facets that guide, not confuse
- Curated facets: only expose attributes with good coverage & quality per category.
- Facet order: place high-gain filters first (brand, price, availability); collapse niche ones.
- Value grouping: bucket long tails (e.g., colors) and keep synonyms mapped (Navy ⇄ Midnight Blue).
2) Ranking signals
- Text relevance: title > brand > attributes > description; exact phrase > term match.
- Business signals: availability, price competitiveness, margin, promo flags.
- User behavior: CTR and conversion lift as soft boosts with decay.
3) Synonyms & spelling
- One-way & two-way synonyms: map colloquial terms to canonical names; avoid over-expansion.
- Spell tolerance: small edit distance with brand safelist (don't “fix” brand names).
4) Boosting & demotion
- Campaign boosts: seasonal collections, new arrivals, or overstock.
- Guardrails: never boost out-of-stock; demote poor content quality.
5) Recommendations
- Related items: accessories via attribute rules (compatibility, size, brand).
- Co-views / co-buys: behavior signals with freshness windows.
- Personalization: lightweight (recent categories/brands) with privacy-first storage.
6) Measurement
- Query buckets: head/mid/long tail; track CTR, add-to-cart rate, zero-results share.
- Test & learn: A/B small boosts and synonym changes; rollback on regression.
Outcome
Clean attributes + curated facets + measured boosts = faster findability and higher revenue per visit.
What's next
Next up: Catalog Health Metrics — measuring data quality, coverage and staleness to prioritize fixes.