As ecommerce professionals and online shoppers, we’ve all felt the limitations of keyword search. You optimize your product titles, clean up metadata, and shoppers still type something like “hydrating face cream for winter” and get irrelevant results. Why? Because legacy search engines aren’t built to understand language the way humans use it.
Shoppers don’t just search. They express needs, preferences, and even moods. And site search needs to pick up on these signals and serve both relevant and attractive results.
That’s where transformers come in.
Best known as the “T” in ChatGPT, transformer models have quietly become one of the most impactful breakthroughs in ecommerce search and discovery. They’re what enable AI to move beyond literal keyword matching and understand the intent behind a query — whether it’s vague, misspelled, or could apply to many different product types.
In this post, we’ll walk through how transformers work, how we’ve integrated them into our search and discovery engine, and how they improve user experience and conversion rates.
Transformers are a type of deep learning model first introduced by Google researchers in 2017. Unlike traditional models that process language word-by-word, transformers process all words in parallel, enabling them to understand the context and meaning behind language.
This architecture underpins modern large language models (LLMs) like GPT and BERT. In plain terms, transformers don’t just read what shoppers type. They interpret what they mean.
We’ve embedded transformer models directly into the foundation of our platform, not as an add-on, but as a core building block. Paired with real-time clickstream data, our transformer-driven engine continuously learns from every shopper interaction.
This makes search results feel both relevant and personal by:
When shoppers search with specific long-tail queries like “popular books for 8 year olds” or “mother of the bride dress,” transformers help an AI-powered search engine interpret the nuance and intent behind these queries and return results that actually make sense.
The search “popular books for 8 year olds” returns a range of age-appropriate children’s books on Target AU’s website.
The same query can mean different things to different people.
By combining transformer models with a behavioral data-first approach — what users click, scroll past, and purchase collectively and individually — Constructor can calculate an attractiveness score and rank products based on what an individual user is most likely to click, rather than by boost and bury rules alone or simply product “popularity.”
Let’s break down what this shift looks like in practice:
Traditional search might match “black hoodie” only if those exact words appear in the product title. Transformer-based search understands that “comfy black pullover” or even “what to wear on a chilly night” could lead to the same item.
People type like humans, not robots. Transformers help normalize messy input like misspellings, shorthand, and brand nicknames so shoppers don’t hit dead ends.
Mixing English and Spanish in a search query like “dry champu” returns attractive results for Sephora shoppers.
Semantic search (often powered by transformer models) maps the meaning of a query to product attributes. So, even abstract searches like “ethical wedding gift” can surface relevant, inspiring results.
It’s important to note that transformers are not a popular approach. They’re more modern. Some vendors still rely on outdated methods, such as traditional vector models, to achieve similar outcomes.
Transformers help connect not just queries to products, but also products to each other, thanks to attribute affinities, purchase-together trends, or complementary use. That means autocomplete, “You Might Also Like,” and dynamic filters feel more curated, personally relevant, and helpful.
As everyday web users are becoming more familiar with LLMs like ChatGPT in their everyday lives (and ChatGPT incorporating native shopping search capabilities), they’re also becoming more comfortable with shopping through chatbots like Constructor’s AI Shopping Assistant as an alternative to the traditional search box.
The Shopping Assistant is where transformers really shine. It enables natural, flowing dialogue between shopper and storefront when shoppers want to combine product questions with their discovery queries. For example:
“I’m looking for a water-based foundation without SPF that’s long-wearing but not full coverage.”
“I need new snowboard bindings in size Small, my board has a 4x4 mounting pattern.”
“What is the safest product to clean bird droppings off my car without damaging paint?”
Kmart Australia’s shopping assistant doesn’t skip a beat when queried “I need travel essentials for my trip to Germany in the fall,” producing attractive search results like packing cubes, travel pillows, and converters.
The AI Shopping Assistant understands follow-up questions, context, and intent and matches results using the same attractiveness scoring we use in search and discovery. And because it’s trained on real clickstream data and works on a constant feedback loop, the more it’s used, the smarter it gets.
The landscape of LLMs is evolving quickly, with new models emerging all the time — each with different strengths. Rather than locking into a single one, we use a number of different providers as well as train our own LLMs and select the best option for each use case.
This means the AI powering Constructor stays current and optimized based on real data. As the underlying models improve, so does the quality of your site’s recommendations, personalization, and product discovery.
Transformer-powered discovery doesn’t just improve search, it also transforms outcomes. By better extracting shopper intent from keywords and ranking products not just by keyword relevance but (more importantly) attractive options, each individual shopper has a better chance at converting. And customers who have a quick and painless search (or chat) experience are more likely to return — even coming to your site before Google.
Want to explore what transformer-powered search and discovery could look like on your storefront? Request a demo of Constructor.