Search & Discovery Relevance Specification Template

Below 5% is a good target for most e-commerce search implementations. Below 3% is excellent. If you are above 10%, there are significant gaps in your synonym dictionary, spelling correction, or product coverage. Track zero-result queries by volume. A single high-volume zero-result query (e.g., a popular brand you do not carry) can skew the overall rate.

Start with well-tuned keyword search (Elasticsearch, Algolia, Typesense) with synonyms, spell correction, and behavioral boosting. This covers 80% of use cases. Add vector/semantic search as a complement for long-tail and natural language queries, not as a replacement. Semantic search excels at intent matching ("furniture for small spaces") but can hallucinate relevance for specific attribute queries ("48 inch walnut desk"). Most teams get better ROI from tuning keyword search first.

Log these queries and review weekly. High-volume searches for uncarried products signal a merchandise gap. In the short term, show related alternatives ("We don't carry Brand X, but here are similar options"). In the long term, use the data to inform buying decisions or seller recruitment in a marketplace. Never show zero results without an alternative path.

Show 4-6 filters on the initial results page, with the most-used filters visible and others collapsed. Category-specific filters should be contextual: show "size" for apparel, "dimensions" for furniture, "wattage" for electronics. Track filter usage rates to identify which filters users actually need vs. which add visual noise without engagement.

Track mean reciprocal rank (MRR) for known-item queries: how far down the results page is the correct item? Track abandonment rate (searches followed by leaving the site). Track refinement rate (searches followed by adding filters or modifying the query). High refinement rates suggest the initial results are not relevant enough. Use session replay tools to watch real search sessions quarterly. ---