Search synonyms: the highest-ROI hour you'll spend on your store

Most search projects reach for embeddings and reranking before they have spent a single hour on synonyms, which is backwards. Synonyms are the cheapest, most transparent and highest-ROI lever in ecommerce search: they close the gap between the words a shopper types and the words your catalogue uses, and unlike a model you can read exactly what they do. This is how we run synonyms on a multilingual, multi-vendor catalogue — what to curate by hand, what to generate, what to deliberately keep apart, and where to find the next entry to add.

Why synonyms are the best hour you'll spend

Every store has a vocabulary mismatch. The catalogue was written by vendors and manufacturers; the queries are typed by customers, who use colloquial names, abbreviations, older brand terms and the words from a rival's advertising. A shopper searches "plant protein"; the products say "vegan". A shopper searches "fish oil"; the label says "omega-3". Neither typo tolerance nor ranking can bridge that — the words are not misspelled and not mis-ranked, they are simply absent. A synonym maps one to the other in a line you can read, test and revert. The return is immediate: queries that returned nothing start returning the right product the moment you add the entry.

Curated versus generated

There are two kinds of synonym and you need both. Curated synonyms are a small, high-trust dictionary you maintain by hand: the structural mappings that apply across the whole catalogue, like "plant / plant-based → vegan" or a local-language term to its catalogue equivalent. They are deliberate, auditable and few. Generated synonyms are per-product terms produced during enrichment: the colloquial names, abbreviations and common variants for that specific item.

Multilingual mapping: local words to catalogue words

Synonyms are where a small-language market is won or lost. Our shoppers search in Estonian, English and Russian, and the catalogue vocabulary does not match any of them uniformly. The curated dictionary maps local terms onto the catalogue's canonical vocabulary: the Estonian "taimne" and the Russian "растительный" both need to reach vegan products, even when the product text is in English. Two rules keep this honest. First, generate native-language synonyms during enrichment rather than machine-translating the catalogue — customers search in words a translator would not choose. Second, treat the curated cross-language list as the place where subtle product distinctions go to die if you are careless, and review every entry as a merchandising decision, not a translation task.

The deliberate non-synonym

The most important entries in a synonym system are sometimes the pairs you refuse to link. Synonym lists are seductive: it is tempting to bundle everything that looks related, and near-miss product pairs are where that bites. EAA and BCAA look similar and are frequently confused, but they are different products — making them synonyms means a shopper who wants one is shown the other as if it were the same thing. The same holds for adjacent subtypes and for a brand that resembles another brand you carry. Keep a guard list of deliberate non-synonyms and treat it with the same care as the synonyms themselves. This is the vocabulary-level twin of the typo-tolerance rule that disables fuzzy matching on identifiers: both exist to stop the system quietly substituting one product for another.

Mine zero-result queries for candidates

You do not have to guess which synonyms to add — your shoppers tell you. The zero-results report is a candidate list in disguise: every query that returned nothing but describes a product you do stock is a missing synonym waiting to be written. Read it weekly, and for each failing query decide which of three things it is — a synonym to add, a genuine catalogue gap to source, or a spelling problem for typo tolerance. The search analytics pipeline is what makes this loop possible, because it captures the raw queries with their result counts.

One-way and two-way synonyms

A detail that saves a lot of pain: decide whether each synonym is directional. A two-way (symmetric) synonym says two terms are interchangeable — "trainers" and "sneakers" — and a search for either returns both. A one-way (directional) synonym expands a narrower term into a broader one but not the reverse: "hydrolysed whey" should expand to include "whey", but a search for "whey" should not be forced to return only hydrolysed products. Getting this backwards over-broadens specific queries or over-narrows general ones. As a rule, brand and subtype relationships want directional expansion, while casual vocabulary pairs want symmetric. Most engines support both, so use the distinction deliberately rather than making everything two-way by default.

Keep the list healthy

A synonym list is a living asset, not a one-off. Three habits keep it from rotting. Version it, so a change that hurts a query can be traced and reverted. Test it against a small regression set of real queries and their expected top result, so a well-meaning addition does not quietly break an unrelated search. And prune it: catalogue vocabulary drifts as vendors come and go, and a synonym that pointed at a discontinued term is dead weight. Done this way, synonyms stay the highest-leverage hour on your search backlog for a long time — small, readable changes that compound into search that speaks your customers' language.

One more discipline: measure the synonym, do not just add it. After you add an entry, the next week's query-to-click data should show the previously-failing query now producing clicks. If it does not, the synonym reached the wrong products and needs revisiting — an unmeasured synonym list slowly fills with well-meaning entries that quietly point at nothing.

Frequently asked questions