E-commerce Product Search: Building Relevance Ranking Beyond Keyword Matching

A shopper types running shoes for flat feet into your search bar and gets zero results, or worse, a page of running socks and treadmill mats. That single moment often decides whether the visit ends in a cart or a bounce. For most e-commerce sites, product search is the primary navigation path, and its ranking logic determines which SKUs actually get seen.

Keyword-only search treats a query as a string to match against a title and description, and that breaks down constantly on misspellings, synonyms, and descriptive phrases never written verbatim in a catalog. Building relevance ranking that reflects actual buyer intent requires blending semantic understanding, merchandising judgment, and continuous measurement.

Table of Contents

• Why Keyword Matching Breaks Down in Product Search
• Semantic and Vector Search: How It Actually Works
• Hybrid Search: Combining Lexical Precision with Semantic Recall
• Merchandising Controls: Boosting, Burying, and Business Rules
• Handling Synonyms, Typos, and Query Variants
• Measuring Relevance: Zero-Result Rate, Click-Through, and Conversion by Query
• Rolling Out Semantic Search Without Breaking What Works

Key Takeaways

• Keyword-only search typically drives zero-result rates in the mid-single digits for large catalogs, and a zero-result query carries a much higher abandonment risk than one that returns even mediocre results.
• Hybrid retrieval, combining lexical (BM25-style) scoring with vector embeddings, commonly outperforms either method alone because it catches exact SKU and model-number matches as well as descriptive, intent-based queries.
• Merchandising boosts and buries should sit in a rules layer separate from the relevance model itself, or every model retrain risks silently undoing months of manual tuning.
• Conversion rate by query, not just click-through rate, is the metric that most reliably exposes relevance problems that look fine in engagement dashboards but quietly suppress revenue.

Why Keyword Matching Breaks Down in Product Search

Traditional search scores products by term frequency, rewarding titles and descriptions that repeat the exact words in the query. That works reasonably well when shoppers search by brand or model number. It fails almost immediately once shoppers describe what they want in their own words: gift for someone who bikes to work, dress for a summer wedding, phone case that survives drops. None of those phrases map cleanly onto catalog text, and a pure keyword engine either returns nothing or returns whatever happens to share a word.

The failure compounds with scale. A retailer with a few hundred SKUs can hand-tune synonym lists and get away with keyword matching for years. A retailer with tens of thousands of SKUs across dozens of categories cannot maintain that manually, and in our experience the maintenance burden, not a single dramatic outage, is usually what forces the shift to a smarter architecture.

Semantic and Vector Search: How It Actually Works

Semantic search represents products and queries as dense vectors, numerical embeddings produced by a language model, positioned so items with similar meaning sit close together in vector space regardless of exact wording. A query for lightweight jacket for hiking in the rain can retrieve a product titled packable waterproof shell because the embeddings capture the underlying concept, not just the literal tokens.

In practice, this means encoding titles, descriptions, and attributes into vectors ahead of time and storing them in a vector index, whether a managed vector store or a vector-capable extension on an existing engine like Elasticsearch or OpenSearch. At query time, the shopper's term is encoded through the same model, and the system retrieves nearest neighbors by cosine similarity. The embedding model matters enormously: a generic off-the-shelf model often underperforms one fine-tuned on your own catalog and query logs, since product language carries its own vocabulary and shorthand.

Hybrid Search: Combining Lexical Precision with Semantic Recall

Pure semantic search introduces its own failure mode: it can be too forgiving. A shopper searching for an exact model number wants that exact item, not a semantically similar substitute, and a vector-only system can bury the precise match under conceptually related but wrong products. The practical answer, and the one most mature search stacks converge on, is hybrid retrieval that runs lexical and semantic scoring in parallel and blends the results.

A common pattern retrieves a candidate set from both a lexical index and a vector index, then combines scores with a weighted formula or a learned reranker trained on historical click and purchase data. The lexical side handles exact matches, model numbers, and brand names reliably; the semantic side handles descriptive and long-tail queries. Getting the blend weight right is an ongoing tuning exercise, not a one-time configuration, and it typically needs revisiting whenever the catalog mix shifts meaningfully.

Merchandising Controls: Boosting, Burying, and Business Rules

Relevance to the shopper is only half the ranking problem. Merchandising teams also need to promote overstocked inventory, feature higher-margin private-label products, and suppress items that are out of stock, discontinued, or under a supplier dispute. Those are business decisions, not relevance decisions, and conflating the two causes real damage: baking a permanent boost for a product directly into a relevance model means that boost silently persists, or silently vanishes, every time the model is retrained, and nobody notices until sales on that item drop.

The more durable pattern keeps merchandising rules in a distinct layer applied after the base relevance score: boost multipliers, pinned positions for specific queries, category-level burial for discontinued lines, and inventory-aware demotion when stock falls below a threshold. This lets merchandising teams adjust promotions weekly without retraining anything, and lets engineers improve the underlying model without erasing a campaign marketing is relying on. It also gives an audit trail: when visibility changes, you can tell whether it came from a relevance shift or a deliberate rule. Rules without an owner or an expiry date are the most common source of relevance decay we see in mature deployments.

Handling Synonyms, Typos, and Query Variants

Even with strong semantic retrieval, many queries fail for mundane reasons: misspellings, plural mismatches, regional naming differences such as trainers versus sneakers, and abbreviations. Typo tolerance is usually handled with fuzzy, edit-distance matching at the lexical layer, combined with auto-correction that substitutes a corrected term when confidence is high. Synonym handling can be maintained as an explicit dictionary for known equivalences, but that ages poorly on its own; pairing a curated list with embeddings that naturally cluster near-synonyms gives better coverage with less manual upkeep.

Query logs are the best source of new synonyms and typo patterns you have. Reviewing the highest-volume zero-result queries on a regular cadence, and feeding confirmed patterns back into the synonym dictionary or fuzzy-matching thresholds, is typically far more effective than trying to anticipate every variant up front.

Measuring Relevance: Zero-Result Rate, Click-Through, and Conversion by Query

Zero-result rate, the share of searches returning no products, is the simplest metric to start with, because every zero-result query is an unambiguous failure with an identifiable cause. Tracking it by query and category surfaces gaps fast, and driving it down is usually the highest-leverage early win in a relevance program.

Click-through rate tells you whether shoppers find something worth clicking, but it can mislead on its own; a shopper might click the only appealing option on a poor results page out of limited choice rather than genuine relevance. Conversion rate by query closes that gap: it tells you whether a query's results actually lead to purchases, and it commonly exposes queries where engagement looks fine but revenue quietly underperforms. Segmenting these metrics by query type, navigational, broad category, and long-tail descriptive, is worthwhile since a single blended number can hide problems in any one segment.

Rolling Out Semantic Search Without Breaking What Works

Replacing a working keyword system wholesale is a common and avoidable mistake. The safer path runs semantic and hybrid retrieval alongside the existing system in an A/B or shadow-traffic configuration, comparing zero-result rate, click-through, and conversion by query segment before shifting traffic. This also catches regressions on navigational queries, where a semantic layer can occasionally underperform a simple exact match, before they reach the majority of shoppers.

Relevance engineering is also a continuous program, not a project with an end date. Catalogs change, query patterns shift with seasonality, and a model tuned for last year's assortment will drift. Teams that treat relevance as a one-time implementation typically see zero-result and conversion gains erode over twelve to eighteen months as the underlying data shifts underneath a static model. A practical sequence for teams starting from keyword-only search: instrument the metrics first, add hybrid retrieval on top of the existing lexical index rather than replacing it, separate merchandising rules into their own configurable layer, and establish a recurring cadence for reviewing zero-result and low-conversion queries.

This is exactly the kind of work where the right architecture choices up front save months of rework later: a vector index that scales with catalog size, a rules layer where merchandising and engineering do not step on each other, and analytics instrumented correctly from day one. If your team is evaluating a search relevance overhaul, our engineers at Halkwinds work directly with e-commerce teams on exactly this kind of build, from infrastructure to the ongoing tuning that keeps it accurate as your catalog grows. You can reach out through our contact page to start that conversation.

Frequently Asked Questions

Is semantic search a replacement for keyword search, or does it work alongside it?

In most production systems, semantic search runs alongside keyword search rather than replacing it. Hybrid retrieval commonly performs better than either approach alone because lexical matching still handles exact SKU, model-number, and brand searches more reliably than embeddings do.

How long does it typically take to implement semantic product search?

Timelines vary with catalog size and existing infrastructure, but a hybrid retrieval rollout, including embedding generation, index setup, and a shadow-traffic testing period, commonly takes a few months from initial architecture to full traffic cutover, not counting the ongoing tuning that follows.

What is a reasonable zero-result rate to target?

There is no universal number, since it depends heavily on catalog breadth and query volume, but large catalogs that have invested in synonym handling, typo tolerance, and semantic retrieval typically bring zero-result rates down into the low single digits, compared to noticeably higher rates on keyword-only systems.

Should merchandising boosts live inside the ranking model?

Generally no. Keeping boosts, buries, and pinned placements in a separate rules layer that applies after the base relevance score avoids the common failure where a model retrain silently overwrites manual merchandising decisions a promotions or category team is actively relying on.

Which metric matters most for catching relevance problems early?

Zero-result rate is the fastest signal to act on because the failure is unambiguous, but conversion rate by query is typically the more important long-term metric, since it catches cases where shoppers click results that simply are not converting, a problem click-through rate alone will not reveal.