5 min read

Vectorized Recommendations overview

“Whoops! No “tweed” found, but you might be interested in plaid jackets instead.

Vectorized Recommendations are relevant product suggestions based on query semantic affinities. They usually show up as product carousels to amaze shoppers with product discovery inspirations.

Vectorized Recommendations

Vectorized Recommendations preview product results of queries semantically related to the original searched query. Shoppers only need to navigate the recommendations’ carousel to get product inspirations. When clicking on a product preview, shoppers directly land on the product’s description page.

Style recommendations

Vectorized Recommendations are displayed mostly as product carousels, but they can also be styled in multiple ways, for example, in a grid or slider. For product carousels, you configure the style settings in the [X-controls section]](/play-with-empathy-platform/configure-empathy-platform/configure-x-controls.md) from the Instance Management Console.

Vectorized Recommendations are helpful in many situations, but especially when there’re no results to show, helping shoppers overcome the roadblock to their search journey. For example, if a shopper searches for the term “tweed”, but there are no direct results, the underlying vector search service in the backend comes into action by displaying results for semantic similar searches like plaid, plaid jacket, and blazer.

You can identify Vectorized Recommendations mostly under sections such as Other similar products.

Vectors as a no results fallback mechanism
You might usually think of zero results as a bad shopping experience. If shoppers couldn’t find the products they’re looking for in your commerce store, most of them would jump off your store and move on to the next one. As a search manager or merchandiser, you can see it as a complicated situation, but you can find low findability as an opportunity as well. Here’s when Empathy Platform’s Vectorized Recommendations come into play.

With Vectorized Recommendations, Empathy Platform’s catalogue of search features offers a different fallback mechanism for no-result pages. Spell Check and Partial Results present product result alternatives based on partial or fuzzy keyword matches of the shopper’s search query. However, Vectorized Recommendations return relevant product results by interpreting search queries as vectors and identifying semantic neighbors to search terms.

Zero-results streamline: Synonyms vs Vectorized Recommendations

Empathy Platform offers a complete catalogue of features to overcome zero-result pages. Traditionally, you use the Playboard’s Synoymize tool to define search term synonyms manually. But it requires some time in the re-indexing process. You use Vectorized Recommendations instead to speed up the zero results fallback experience by dynamically getting semantical synonyms to shoppers’ queries, ensuring a seamless, relevant search experience.


Learn more on how vectors come into play in the Inner workings of Vectorized Recommendations section below.

Spot the difference

Vectorized Recommendations are part of the Empathy Platform’s recommendations catalogue. Even though the core purpose for all recommendation types is to ease product discoverability, their scope is totally different. Unlike Top Product Recommendations, which are not related to the search queries or the shopper intent, as they showcase the most frequently visited products in your commerce store, Vectorized Recommendations offer products that align semantically with the shopper's search intent.

Try Recommendations to...

  • Improve product findability. Help your shoppers find the products in your product catalogue they’re looking for.
  • Handle shopper frustration. Skip the frustrating No results found page and suggest alternative close search results that semantically match your shoppers’ intent.
  • Streamline zero results search and avoid abandonment.
  • Enhance shopper experience. Provide a smooth shopping experience based on relevant product discovery journeys on the fly.
  • Meet shoppers’ intent. Bridge the gap between shoppers and your product catalogue when they don’t speak the same language.

The inner workings of Vectorized Recommendations

When a shopper launches a search and there’s no exact keyword matching in your product catalogue, leading to no results at all, the Vectorized Recommendation system comes into action.

First, the Semantics Query component from Interface X detects if the input query returns any results. As no results are returned, the Semantics Query component communicates with the Semantics API to get some search term suggestions, semantically similar to the original query, that may bring meaningful results.

The underlying vector search service performs a vectorization process, based on machine learning (ML) and natural language processing (NLP) techniques, to represent search terms in a numerical format, called vectors. Trained semantic models are used to establish associations between search terms and the original query, where textual information is transformed into numerical representations that capture semantic meaning. So, the service calculates the semantic similarity between the search term vectors to identify and suggest semantical-closed terms as search alternatives to the shopper’s initial query. For example, for the query “joggers” the Semantics service may bring query suggestions such as sweatpants, jogging pants, tights or even shorts.

Finally, the Semantics API returns the semantical alternative queries found for the Interface X Query Preview component to display a preview of the product results related to these close search terms, allowing the shopper to explore relevant options that really match their search intent.

Privacy matters

Empathy Platform’s Vectorized Recommendations system is built on a native stack, respecting shoppers’ data privacy. Empathy Platform uses anonymized tagging events to ensure an ethical and contextualized search experience. So, semantic models are based on learned patterns and relationships from your store dataset only, enabling to understand the semantic similarity between different search terms based on the catalogue domain.


When there are no results available on query time, the Search microservice first check whether a typo during query formulation exists. In that case, the Spell Check feature looks for close terms that lexically match the original query to offer fallback product results, instead. Note that the Vectorized Recommendations system doesn’t come into play in this situation.