Ecommerce search: filters, Meilisearch and AI

Good ecommerce search helps people find the right product with as little unnecessary clarification as possible. It depends on useful catalogue data, understandable filters and relevance rules. AI and semantic retrieval can extend that foundation, while exact identifiers, compatibility and current availability remain essential.

Visitors arrive with different intentions. One pastes a spare-part code. Another asks for headphones suitable for a noisy office. A third wants a gift associated with a favourite character. Applying the same matching behaviour to every query can either hide the right product or return too many misleading alternatives.

Classify real queries before choosing a tool

Take a representative sample of searches and remove unnecessary personal data. Separate identifiers, names, characteristics, task descriptions and requests for products you do not stock. Mark spelling errors, transliteration and different languages. This collection becomes a reusable evaluation set for any proposed search solution.

An empty result is not always an algorithm failure. The item may be absent, named differently or missing a relevant attribute. Keep these causes separate. Otherwise the team may tune the search engine when it should improve catalogue data or explain available alternatives.

Product attributes shape discovery

A long description cannot replace structured information. Spare parts need device, model, compatibility and quality details. Collectibles may need character, series, manufacturer and scale. These fields affect search, filters, product pages and useful thematic landing pages.

UParts organises selection around devices and spare-part characteristics. Pulsar also gives visitors paths through franchises, characters and manga series. They illustrate different ways to express customer intent. A copied list of filters would not serve both audiences equally well.

Agree a controlled vocabulary for brands, models, units and codes. Variations in capitalisation or supplier abbreviations should not accidentally create several versions of one attribute. Where useful for troubleshooting, preserve the original supplier value alongside the normalised value rather than losing its provenance.

Where Meilisearch and typo tolerance help

Meilisearch supports configurable typo tolerance, including disabling it for selected attributes. A near match can help with a misspelled name while misleading someone searching for a unique code. Its official explanation of typo tolerance discusses this distinction.

Imagine two part numbers differing by one digit. Presenting one as an exact answer for the other can lead to a wrong purchase. Show the precise match clearly and label alternatives honestly. A spelling error in an ordinary product word is a different situation, where assistance can save an unnecessary reformulation.

Engine selection also depends on index size, update frequency, filtering requirements and maintenance resources. A small catalogue may be well served by a simpler implementation. Test with your own data instead of relying on a demonstration built around specially chosen examples.

When semantic retrieval is worth testing

A semantic approach compares meaning rather than only shared words. It is worth evaluating when customers describe a need in everyday language, such as a compact item for travel or a gift for someone interested in space adventures. The catalogue must contain enough meaningful information to support that selection.

Similarity does not establish technical compatibility. A component with a similar description may not fit the required device. Keep hard requirements such as size, fit, availability and sales region as explicit constraints. Compare a hybrid approach with conventional text search on the same tasks.

If a conversational interface is added, evaluate its statements and product links separately. Our article on RAG and website AI assistants covers that layer. A chat interface inherits gaps in the catalogue unless those gaps are addressed.

Filters should explain the available choices

Use labels the target audience understands. Counts, active constraints and reset controls should behave predictably. On a phone, selections should survive closing the filter panel, and the interface should make the resulting number of products clear. Test the actual sequence of opening, choosing, applying and returning.

Zero results deserve a designed state. Explain which constraint can be removed, offer a relevant clarification or provide a route to help. Silently ignoring a filter may look successful, but it can undermine trust when the returned products do not meet the visitor's requirements.

Measure the outcome of a search change

• Relevance: does the correct item appear early for known queries?

• Empty results: which are caused by assortment, data or matching?

• Next action: do visitors open an item or complete a relevant buying journey?

• Reformulation: how often must people try again?

• Freshness: do changes and deletions reach search results in time?

Compare similar categories and account for seasonality, campaigns and assortment changes. Higher sales after a release do not by themselves prove that search caused the increase. A controlled experiment or careful segment comparison provides a more defensible conclusion.

Performance is part of discovery too. Requests during typing, heavy product cards and unnecessary scripts can make a fast server search feel slow. Check weaker devices and realistic networks. The guide to website performance and Core Web Vitals explains how to separate different kinds of delay.

Prepare for ongoing catalogue changes

Assign responsibility for synonyms, attributes and failed-query review. A search configuration that works today may need adjustment after a new product category arrives. Keep a small set of representative queries as regression checks so improvements for one group do not quietly damage another.

Also define what happens when the search index is temporarily unavailable or behind the source catalogue. Customers should receive clear behaviour, and staff should be able to see the last successful update. Search quality includes dependable operation, not just ranking in a prepared test.

Frequently asked questions

Does every shop need AI search?

No. Improve attributes, exact matches, synonyms and filters first. Semantic retrieval should demonstrate a benefit for specific tasks compared with that baseline.

Can search improve without a complete redesign?

Often yes. Review the catalogue source, indexing and frontend integration. A focused component and data update may be sufficient.

What should we bring to an initial assessment?

Examples of failed searches, a product sample and expected results. These make ecommerce development a discussion about an observable customer problem rather than a fashionable tool.