RelevanceRail
RelevanceRail answers

Query Understanding: Synonyms, Typos, and Intent in Site Search

Shoppers search in their own words: slang, abbreviations, misspellings, and product nicknames. Query understanding translates that language into catalogue terms before ranking starts. Synonym rings handle vocabulary gaps, typo correction handles errors, and intent detection decides whether a query wants a product, a category, or an answer.

Synonyms: rings, not pairs

A synonym ring treats a set of terms as equivalent for retrieval: trainers, sneakers, runners. Rings beat pairwise lists because they stay consistent as terms are added, and they can be scoped per category, since the same word means different things in different departments.

For more on this, see this related guide and our earlier explainer on the same theme.

Build rings from your own search logs, not from a generic thesaurus. The vocabulary gap that matters is the one between your shoppers and your catalogue. Mine zero result and reformulated queries monthly. Every recurring failure is a ring, a redirect, or a catalogue naming fix waiting to be made.

Typo tolerance with a light touch

Shoppers mistype constantly, especially on mobile. Good systems correct obvious errors transparently, showing what was searched for, and never silently substitute a different product intent. Aggressive correction that changes a correct but unusual query is worse than no correction, because it hides the exact item the shopper named.

Set correction strength by query length and catalogue confidence. Short queries are ambiguous and deserve suggestions over substitution. Long specific queries with one near match in the catalogue can be corrected confidently. Log every correction and review them: your correction log is a map of catalogue naming problems.

Intent changes what a good result is

The same words can carry different jobs: a model number wants one product page, a category word wants a browsable list, a problem phrase wants a guide that then links to products. Intent detection uses query shape, length, and known entities to choose the response type, and getting it wrong wastes the shopper's clearest signal.

Readers also found this companion piece from our archive useful.

Measure by intent class, not in aggregate. Search can look healthy overall while problem phrase queries, often your highest intent shoppers, fail completely. Segment the search analytics, or the averages will hide exactly the failures worth fixing.

Maintain it like a product, not a setting

Query understanding decays as catalogues, slang, and seasons change. Assign an owner, review the top failing queries weekly, and ship small fixes continuously: a ring here, a redirect there, a boost for the term everyone suddenly uses after a trend breaks.

Resist the urge to rebuild search annually instead. The systems with the best real world relevance are boringly well maintained ones, where a named person can explain why any given query behaves as it does, and where every manual rule carries a date and a reason.

Frequently asked questions

What is the difference between a synonym and a redirect?

A synonym expands retrieval so both terms find the same products while browsing continues normally. A redirect sends the shopper to a specific page, used when a query has one clearly correct destination, such as a service page or a single product line.

Why does search ignore my exact product name sometimes?

Usually because stemming, synonym expansion, or a popularity boost outranks the exact match. Exact title and SKU matches should carry a protected boost. If they do not on your platform, that is a configuration bug worth escalating, not normal behaviour.

How do I find which queries need synonyms?

Start with zero result searches and searches followed immediately by a reformulation. Both are in your search analytics. Group them by theme, fix the highest volume clusters first, and re measure the same queries after the fix ships.

Should synonyms be bidirectional?

Within a ring, yes for retrieval: every term should find the others' products. Display language can stay directional, using your catalogue's preferred term in titles and facets while still honouring the shopper's word in the search box.

© 2026 RelevanceRail