A synonym strategy that improves search relevance

What a synonym in online store search means

In a linguistic sense, synonyms are words with the same or almost the same meaning. In online store search, the practical question is more specific: can two terms within this catalogue pick up the same products? The result may vary by product group, market and language.

Overlap ‘sneakers’ and ‘sports shoes’, but are not completely equal in every context. A customer looking for running shoes does not automatically expect all casual sneakers. The system should therefore take into account category, query and assortment instead of linking words blindly worldwide.

Synonyms are one instrument in addition to spelling correction, semantic matching, category recognition and recommendations. Use them for clear language relationships. Do not solve missing product data, wrong categories or irrelevant basic rankings with ever wider expansions.

Find candidates in real client questions

Start with search dates. Recurrent zero-result searches, low click ratio, fast reformulations, and terms that customer service often hears show where customer language and catalog language vary. Combine volume with meaning: a little-used term can be important for business customers.

Look at search pairs within the same session. When visitors first search for ‘TV cabinet’ and then ‘seating furniture’ and then click, that can be a candidate relationship. Pattern is not yet proof; check the result sets and product meaning manually.

Involve category experts and local teams. They have abbreviations, regional words, brand language and technical terminology that do not always emerge from a short data period. Document the source of every major candidate.

• Commonly used queries to no avail while appropriate products exist.
• Search queries that are reformulated shortly afterwards.
• Alternative product names from customer service and sales.
• Abbreviations, deregistrations and regional variants.
• Catalogue terms that customers hardly use.

Distinguish between equal, related and hierarchical

True equivalence is the safest: two terms may yield virtually the same candidate products. Related terms require more restraint. ‘Printer’ and ‘ink’ belong together, but are not synonyms. A query to a printer should not primarily show cartridges because the relationship is commercially interesting.

Hierarchical relationships connect broadly and specifically. ‘Shoes’ includes running shoes, hiking shoes and more; the reverse direction is not the same. Treat such relationships as category or query terms, not automatically as a two-way synonym.

Applications are not self-evident synonyms either. A ‘gift for father’ can refer to many product types. There, curated content, recommendations or merchandising may fit better than a permanent word pair.

Choose consciously between one- and two-way rules

In a two-way rule, term A expands to B and B back to A. That is appropriate when both within the catalogue are truly equivalent. With a one-way rule, a specific customer term may find additional products without a broad catalog term automatically getting the same extension.

For example, an abbreviation may safely refer to the full product name, while the full name should not always extend to all the meanings of the abbreviation. Brand names and product categories require extra caution: a brand is not a synonym of the product type.

Limit scope to category, language, store, or query field where possible. A relationship that is correct in Dutch consumer electronics does not have to be valid in a B2B- parts catalogue. Local scope lowers the likelihood of side effects.

• Two-way only with demonstrable equivalence.
• One-way at specific to broader or official terminology.
• Category or shopping scope when meaning is context-dependent.
• Exact brands, models and identifiers protect against unwanted expansion.
• Give temporary campaign extensions an end date.

Test relevance before and after each important rule

A synonym is only good when the result set gets better. For both terms, check which products appear, which exact matches change, and whether irrelevant categories come in. Test not only the query for which the rule was created, but also other queries that contain the same words.

Use a fixed regression set for terms with high traffic, high value and strong exact matches. Where possible, automate checks on zero-result searches and product coverage, but maintain manual relevance assessment. Figures cannot always determine whether the right model, material or purpose is at the top.

After go-live, measure zero-result searches, click ratio, first click position, and conversion where the assignment is reliable. A decrease in zero-result searches is positive as long as the additional results are also relevant. More products without a usable click is not a real solution.

Manage synonyms as a product, not as a separate list

Customer language and assortment change. New brands appear, categories shift and temporary terms lose meaning. Every important rule therefore needs a reason, owner, scope, creation date and evaluation moment.

Make changes traceable and provide a test or approval step before publication. In a multi-store environment, central management can be useful, but local teams need to be able to signal different terminology. Decide which rules are global and which only apply locally.

Clear rules when the underlying product data has been improved or a campaign ends. A smaller, well-founded synonym file is usually better understood and tested than a large historical archive.

Do not manage translations as an automatic copy of the Dutch list. In another language, a relationship can exist more broadly, narrower or not at all. Use real search data and local review by language, while shared product concepts can remain central. This prevents a technically consistent configuration from yielding linguistically incorrect results.

• Source and problem that solves the rule.
• Relationship type and direction.
• Language, category, shopping and time scope.
• Owner, reviewer and evaluation date.
• Results of query testing and effect measurement.