Search relevance: synonyms, typos and the catalogue you have

"Our search is bad" describes three unrelated problems, and buying one tool to fix all three is how stores end up paying monthly for something that solved a third of the issue.

No results. The customer typed a word that appears nowhere in your product data. This is a vocabulary problem and it is by far the most common. It is also the only one where the customer definitely leaves.

Wrong results. The words match, but the ranking puts the interesting products on the second screen. This is a merchandising problem, and its fix is boosts and better product data rather than a different engine.

Too many results, no way to narrow. The customer searched a category term and got four hundred products with nothing to filter by. This is a filter problem and it is fixed on the collection side, not in the search box.

Before spending anything, take a week of query data and sort every top query into one of those three. The distribution tells you what to buy, and more often than not it says the answer is not a search app at all.

This has moved, and a lot of received wisdom about Shopify search is out of date. Semantic search is documented as using "related words, concepts, categories and other contexts to improve and expand search results", analysing product attributes including descriptions, image text and colours to find matches beyond exact keywords.

Shopify’s own example is worth quoting because it shows the scope: a store selling shoes receives a search for "christmas party shoes" and has no products using those words. The system associates "christmas" with red and green, and "party shoes" with pumps, and surfaces red pumps. That is precisely the class of query a synonym list would otherwise have to anticipate by hand.

It is included rather than configured — Shopify states it "doesn't need to be turned on" when the store meets the requirements. Those requirements are documented: fewer than 200,000 products, a Grow, Advanced or Plus plan, and it is not available for predictive search or for the Japanese locale.

Read the last part twice. Predictive search is excluded, which means the dropdown that appears as a customer types and the results page they reach on pressing enter are running different logic. A query that returns nothing in the dropdown may return good results on the full page. Most of the "our search is broken" reports we are sent are screenshots of the dropdown, and most testing we see is done in the dropdown, because that is what people use. Test both, separately, and know which one you are looking at.

The plan requirement matters too: if you are below the tiers listed, semantic search is not part of your baseline, and the gap a third-party app would fill is genuinely larger.

What is documented about matching is narrower than most people assume. The prefix option controls partial word matching on the last search term — set to last, a search for "winter snow" finds resources containing "winter" and terms beginning with "snow", such as "snowshoe" or "snowboard". Set to none, it does not.

On typo tolerance specifically, we would rather describe the mechanism than quote a figure: Shopify does not publish a tolerance specification we can point you to, so the answer for your store is a test rather than a claim. That test is an hour of work and worth more than any vendor comparison table.

Take your thirty highest-volume search queries. Take the misspellings that actually appear in your own query report — not invented ones, the real ones, which are usually brand names, foreign words and anything with a doubled consonant. Run each one through the dropdown and through the results page, on the live storefront, and write down what came back. You now have a factual list of failures instead of an impression, and each one is either fixable with data you control or it is a genuine argument for buying something.

If a customer’s word is not in your product data, no amount of ranking will save the query. So the durable fix is to get the words in, and the place they go matters.

Product titles carry the most weight in any search system and are also read by customers in the results grid, so they are the most expensive place to stuff keywords. Put the word people say here, once: the common name, not the internal SKU convention and not the supplier’s designation.

Product type, tags and metafields are where the alternatives live — regional names, older model numbers, the material, the use case. These also feed your filters, which means one piece of structured work pays out twice. Free-text tags used as a dumping ground for search terms will make a mess of your filter lists, so decide which fields are for people and which are for machines before you start.

Descriptions are read by semantic search along with image text and colours, which raises the value of writing them as a person would describe the product rather than as a spec table. That is also how they read better to customers and to answer engines, so it is rare for this work to be wasted.

Product boosts are the manual override, and they are deliberately limited: a maximum of 10 search terms per product, applied only when the product is available for sale, and not applied at all when a query contains search syntax. That makes them the right tool for your twenty most valuable queries and the wrong tool for a synonym dictionary. If you find yourself maintaining boosts for hundreds of terms, the vocabulary belongs in the catalogue instead.

One honest caveat on this whole area: the configuration surface for search on Shopify has changed more than once, and features move between the app and the platform. Check what your admin actually offers today rather than trusting an article — including this one — about what the options are called.

Unavailable products in search results are a setting with three documented values: show displays them in discovery order, hide excludes them, and last places them after the available results. last is the default, and it can be changed in the Search & Discovery app’s search settings.

This deserves an actual decision rather than a default, because the right answer depends on the catalogue. A fashion store where most products are available in some sizes and not others should almost never hide: the customer searching a product name wants to see it and find out. A store with a long tail of discontinued lines should hide, because a results page of things nobody can buy reads as a dead shop.

The version that causes real damage is show on a catalogue with heavy stock-outs, where the top of the results page is full of products that cannot be bought. If you have never looked at this setting, look at it — it is one click and it changes what a meaningful share of searchers see.

A search for a broad term is a browse session that started in the wrong box, and the thing that rescues it is filtering. The same Search & Discovery limits apply here as on collections: a store gets a maximum of 25 filters combined across standard and custom, each source usable once, and — the one to know — a search producing more than 100,000 results does not display filters at all.

For most catalogues that ceiling is theoretical. The practical point is that the filters available on a search results page are the store-wide set, applied to whatever the query returned, so a search-specific filtering strategy does not really exist. Get the store-wide set right and search benefits automatically.

We have been asked to install a search app several times and recommended the catalogue work instead more often than not. The pattern is consistent: the failing queries turn out to be the customer’s word for a product, the internal name is in the title, and nobody wrote the common name anywhere. That is a copywriting and taxonomy job measured in days, it improves collection pages and filters at the same time, and it has no monthly cost.

Where a search app genuinely earns its place, it is on a large multilingual catalogue with someone whose job is merchandising the results. If nobody owns that, you are buying a control panel that will look exactly as it did on the day it was installed a year from now — which is the real reason most search implementations disappoint, and no vendor will tell you it is the reason.

• Shopify conversion rate optimisation
• The search box is a merchandising surface
• Zero-result searches are a merchandising report