Awesome Search KG

E-commerce Search

5 min read

E-commerce Search

Practice guide for product search — the specific challenges, patterns, and tradeoffs that distinguish e-commerce search from web or enterprise search.

What Makes E-commerce Search Different

The Catalog Is the Corpus

Unlike web search, your corpus is fully controlled. This is a superpower (you can enrich it) and a constraint (vocabulary mismatch is your problem to solve).

Users Don’t Know the Right Words

Sellers describe products their way; buyers search their way. A listing titled “Bohemian Macramé Wall Hanging” needs to surface for “boho wall art” and “woven wall decoration.” Bridging this gap is the core e-commerce search problem.

Relevance Is Multi-Dimensional

A result is “relevant” if the user can buy it, it fits their need, it’s available, it’s in their price range, and the seller is trustworthy. Each dimension may require separate signals.

A common two-way cut is Semantic Relevance (does the product match the query?) against Outcome-Based Relevance (do clicks and purchases confirm it?); see Outcome-Based vs Semantic Relevance. Criteo uses the first as a guardrail threshold and the second to order what passes it. In its own offline benchmark over 400 candidates, re-ranking semantically retrieved products by recent sales share raised the purchased product’s normalized rank score about 2.5× (Leveraging Commerce Data for Outcome-Based Relevancy in Agentic Recommendation Systems, Introducing CLEPR, our model for semantic understanding).

Business Rules Are First-Class

Margin, inventory, promoted listings, new seller support — these business concerns legitimately affect ranking. The challenge is integrating them without destroying relevance.

The 12 Query Types (Baymard)

Understanding which type a query is determines how to handle it:

TypeExampleChallenge
Exact product”iPhone 15 Pro Max”Navigational — precision matters most
Category”headphones”Diversity — show the space
Feature/spec”wireless noise cancelling”Attribute matching + ranking
Thematic”minimalist desk accessories”Semantic, no keyword overlap
Symptom”my back hurts pillow”Requires intent inference
Compatible product”case for MacBook Air M2”Structured attribute matching
Use-case”shoes for marathon running”Activity → product category
Inspirational”gift ideas for dad”Pure exploration

Most search systems optimize for exact + category queries. Symptom, thematic, and inspirational queries are where most systems fail.

The Three Pillars of Quality

Findability

Can the user find what they’re looking for?

  • Key metric: zero-results rate
  • Main causes of failure: vocabulary mismatch, spelling errors, missing synonyms
  • Fixes: synonyms, spelling correction, query relaxation, dense retrieval
  • Caveat: relaxation that empties the no-results bucket can still lose the sale; measure conversion of relaxed queries (Query relaxation - the forgotten hero of eCommerce search, Rohlik)

Relevance / Ranking

Are the right items at the top?

  • Key metric: NDCG, conversion rate
  • Main causes of failure: wrong ranking signal, position bias in training data, business rules overriding relevance
  • Fixes: LTR, implicit signals, debiasing

Engagement signals (clicks, add-to-carts, purchases) are the default ranking proxy, but they bias toward already-popular listings regardless of query fit. Semantic Relevance is a complementary, intent-based signal — Etsy scales it with an LLM judge anchored to human labels, distilled down to a model cheap enough to run in real-time production ranking.

Andreas Wagner also questions NDCG itself as the ranking metric: product grids expose several items per row, with how many depending on the device, so attention drops at row and viewport boundaries rather than along a log curve, and image, price, brand and promotions weaken the metric’s tie to revenue (Why nDCG Misleads eCommerce Search).

Discovery

Can users find things they didn’t know they were looking for?

  • Key metric: diversity, engagement depth, session conversion
  • Main causes of failure: winner-takes-all ranking, filter over-application
  • Fixes: diversification, exploratory UI patterns, personalization

Vocabulary Management

The most common root cause of e-commerce search failure:

  • Sellers write product titles in one vocabulary; buyers search in another
  • Solution stack: synonyms (known pairs) → spelling correction → query expansion → semantic/dense retrieval
  • Don’t jump straight to dense retrieval: synonyms and spelling correction are cheaper and more controllable

See Synonyms and Vocabulary Management and Spelling Correction.

Facets in E-commerce

Facets are the primary UX mechanism for navigating large product catalogs. Key decision: are facet selections hard filters (Boolean, zero results if nothing matches) or soft preferences (boost matching, don’t exclude)?

  • For navigational queries: hard filters are correct
  • For exploratory queries: soft preferences improve discovery
  • Hybrid: hard filter + “include similar items?” suggestion

See Faceted Search.

Broad and Head Queries

High-traffic head queries (“shoes”, “necklace”, “bag”) span enormous intent spaces. Optimizing purely for relevance produces “winner takes all” results — the most popular subcategory dominates.

Approach: detect broad queries → inject diversity → ensure result set represents multiple interpretations.

See Search Result Diversity.

Incremental AI Adoption

Four levels from zero to conversational:

  1. Level 0: keyword + filters only
  2. Level 1: async AI suggestion (“did you mean: Property Type: Condo?“) — zero latency risk
  3. Level 2: AI auto-executes rewritten query — 2-3s added latency
  4. Level 3: conversational chat interface — stateful, multi-turn

Start at Level 1. Only move up if A/B test shows improvement.

Graph View

Backlinks