What do you mean?

Shoppers make typos when they type search terms, and those typos usually lead to zero results or irrelevant results. AI Spell Check steps in only when a query returns no results and a correction is available. It then shows results for an alternative query that is close to the original one in terms of either spelling or meaning, but always drawn from terms that actually exist in the catalog.

For example, if a shopper types bisiuts instead of "biscuits", AI Spell Check detects the typo and returns results for the corrected alternative search term.

AI Spell Check combines two layers that run in sequence:

• A lexical layer that looks for the closest match by spelling.
• A semantic layer, also called AI Semantics, that understands meaning and intent rather than just spelling, and steps in when the lexical layer doesn't retrieve any results.

Together, these layers help shoppers reach relevant products even when a query is heavily misspelled. Both search a standalone embeddings feed built from the product catalog, but limited to the fields that actually help with correction, such as product names, descriptions, categories, and brands, rather than full product records. This keeps the feed focused on the vocabulary shoppers search for, instead of every detail stored in the catalog.

AI Spell Check corrects queries that contain some kind of typo, but that error doesn't have to be a single misspelled keyword. It also supports longer, multi-term queries, including more complex ones that describe what shoppers want rather than naming it exactly, or that use natural language, as long as they still contain a typo. For example, the lexical layer corrects the composite query "bitamin c 1000mg" with no results to "vitamin c 1000mg", even when the typo lands on the very first letter of the word. The semantic layer can go further: it corrects the query "anti gaing cresm" with no results to "anti aging cream" products, even when typos spread across several words make it harder for the lexical layer to find a close match.

warning

AI Spell Check doesn't trigger if the current query returns any results, even if that query is misspelled or mistyped.

When AI Spell Check corrects a query, a message can appear on the SERP to let shoppers know that the original query returned no results and show the closest alternative that replaced the original query. So, you can identify AI Spell Check under messages such as Did you mean…? or There were no results for "bisiuts". Showing results from "biscuits".

Deprecation notice

AI Spell Check replaces the classic Spell Check feature, which relied on Elasticsearch's own suggester capabilities. That classic, spelling-only correction feature is now deprecated. AI Spell Check's lexical layer takes over these spelling-based corrections, now running against a standalone embeddings feed, and adds a semantic layer on top, making it substantially more capable than the previous engine, including resolving typos that land on a word's very first letters.

Spot the difference

AI Spell Check comes into play when it determines that the input query might be misspelled, leading to no results, and replaces the original query with an alternative search term. AI Spell Check doesn't recommend search term suggestions; don't confuse it with the Query Suggestions feature, which displays a list of query suggestions while shoppers type to refine and improve query formulation.

Don't confuse AI Spell Check's semantic layer with Semantics Recommendations, either. Semantics Recommendations finds queries that are semantically close to the original one and surfaces the products for those alternative queries as carousels, a query-to-query search type. AI Spell Check's semantic layer improves on that approach: it goes straight from the query to matching products, a query-to-product search type, without an intermediate step of suggesting alternative queries. Note that Empathy Platform is moving away from Semantics Recommendations as the AI Spell Check semantic layer takes over.

Don't reach for Synonyms either. Synonyms, available from the Empathy Platform Playboard, maps specific terms to a guaranteed, business-defined match regardless of what AI Spell Check's layers find on their own. Since the semantic layer now resolves many misspellings and vague phrasings automatically, reserve Synonyms for mappings you want to control directly rather than leave to automatic correction.

note

Other features, such as AI Carousels, Related Prompts, Recommendations, or Partial Results, also help prevent zero-results pages and minimize shopper frustration.

Try AI Spell Check to…

• Help shoppers find what they're looking for despite spelling mistakes (e.g., "running shoos" instead of "running shoes").
• Catch typos in long, descriptive, natural-language queries that name several attributes at once (e.g., "waterproof yacket for hiking in the rain" instead of "waterproof jacket for hiking in the rain").
• Fix typos that land on a word's very first letters (e.g., "bitamin c" instead of "vitamin c"), a case that spell checkers relying on prefix matching often fail to resolve.
• Skip the frustrating No results found page and surface close, relevant matches that meet your shoppers' intent.
• Leverage zero-results searches, providing a smooth search experience and avoiding abandonment.

The inner workings of AI Spell Check

When a query returns no results, the Search microservice looks for and compares close terms that approximately match the original search term, from either a lexical or a semantic perspective, to invisibly launch a new alternative query and return relevant results.

AI Spell Check runs against a standalone embeddings feed, not against the live product catalog directly. This feed is built from the same source data as the catalog, but extracts only the fields needed for correction, including product names, descriptions, categories, and brands. Extracted terms go through a deduplication process so that each unique term in that feed gets its own vector embedding. This means the index represents recoverable vocabulary rather than complete product records.

Then, the embedding phase converts each unique term into a vector. Because the feed stores one embedding per unique term instead of one per product, it avoids re-processing the brand names, categories, and common words that repeat across many products, and keeps the vocabulary focused on what shoppers actually search for. Each term's vector, together with the list of products it comes from, is stored in a single document per term.

Once the embeddings feed is indexed and stored, when a search returns no results, AI Spell Check tries a lexical match first. It uses approximate string matching algorithms to determine the similarity, or distance, between the alternative terms and the original search term, and finds correct spellings of the misspelled term in the embeddings feed. The lower the distance between the original search term and the alternative one, the more similar they are. If it finds a good alternative, it takes the associated products, retrieves the matching results from the catalog managed by the Index microservice, and displays them on the SERP.

If the lexical layer doesn't retrieve any results, the semantic layer takes over and runs a semantic (kNN) search instead. It looks for the closest semantic matches to the original query inside the same embeddings feed, based on meaning rather than spelling. If it finds matches, it goes back to the catalog to retrieve the corresponding products and displays them on the SERP.

Resilience and fallback behavior

AI Spell Check is designed to fail safely. If the embeddings service used to compute the correction doesn't respond in time or returns an error, AI Spell Check doesn't produce a correction, and the search response isn't affected. It never returns a broken or empty page as a result of this dependency.

If AI Spell Check still doesn't return results, other fallbacks, such as AI Carousels or Related Prompts, can display so shoppers never land on a blank results page.

Deactivating the semantic layer for regulated catalogs

Because the semantic layer proposes an alternative based on meaning rather than exact spelling, it can occasionally match a misspelled query to a real, but different, product when two catalog terms are lexically close. In a general commerce store, for example, a shopper searching for a "Phillips" screwdriver could conceivably be redirected toward "Philips" branded electronics instead, since the two brand names differ by only one letter.

In safety-sensitive verticals such as pharmacy, the semantic layer can be deactivated to rely on the lexical layer alone and prevent it from correcting a query toward a wrong product or medication whose name differs by only a couple of characters from the one the shopper meant.

---

FAQs

What is AI Spell Check?

AI Spell Check is a feature that corrects zero-result queries by proposing an alternative query drawn from terms that exist in the catalog. It combines a lexical layer, which corrects by spelling, and a semantic layer, also called AI Semantics, which corrects by meaning.

When does AI Spell Check trigger?

AI Spell Check only triggers when a query returns no results and a correction is available. It doesn't trigger if the current query already returns results, even if that query is misspelled.

What happened to the classic Spell Check feature?

The classic, spelling-only Spell Check is deprecated. AI Spell Check's lexical layer covers the same types of corrections, now based on an embeddings feed, and adds a semantic layer on top.

How is AI Spell Check's semantic layer different from Semantics Recommendations?

Semantics Recommendations finds queries that are semantically close to the original one and surfaces the products for those alternative queries as carousels, a query-to-query search type. AI Spell Check's semantic layer goes straight from the query to matching products, a query-to-product search type.

Can the semantic layer be disabled?

Yes. In safety-sensitive verticals such as pharmacy, the semantic layer can be deactivated to rely on the lexical layer alone, since it can occasionally match a misspelled query to a real, but different, product when two catalog terms are lexically close.