How Transformers Can Enhance Search and Product Discovery

Providing an exceptional, highly personalized search and product discovery experience can make the difference between a sale and an abandoned cart. As customer expectations continue to rise, traditional keyword-based search solutions often fall short. Enter transformer models, a technological advancement that's revolutionizing how customers find products online.

At Constructor, we've integrated transformer technology into our product discovery platform to help merchandisers and product owners deliver more personalized, intuitive shopping experiences that drive revenue. In this article, we start by briefly exploring what transformer models are and then deep dive into how they work within our platform as well as their tangible business benefits. 

What Are Transformers? 

Transformers are a type of neural network architecture capable of understanding and processing natural language queries simultaneously. They have a leg up on their predecessors, which processed words sequentially (i.e., one after another). 

This retailer’s search engine ranks ‘high rise sweatpants’ despite the query specifically saying ‘not high rise.’ That’s because it processes the search phrase word by word and not as a whole. 

Because transformers process all words simultaneously, they’re able to better understand context and meaning in language. By analyzing the relationships between words in a query, transformers can grasp the intent behind what someone is searching for — not just match keywords.

A different retailer who does use a search engine powered by transformers is able to return much more contextually relevant product results for the same query ("womens sweatpants not high rise"). This not only fulfills the searcher’s request, but increases their chances of converting.  

In ecommerce, this means the difference between showing a customer irrelevant results that happen to contain their search terms and understanding what they're actually looking for, even if they use imprecise language, make typos, or phrase their query in an unexpected way.

Benefits of using transformers in product search

Transformer models offer several key advantages over traditional keyword-based search:

• Better understanding of intent: Transformers capture the contextual meaning of words, enabling the system to understand what shoppers are really looking for, not just what they typed
• Improved product discovery: By understanding semantic relationships, the system can surface products that are genuinely relevant to a query, even if they don't contain the exact search terms
• Handling language complexity: Transformers excel at understanding natural language, including multi-lingual colloquialisms, industry-specific terminology, and even typos or grammatical errors

For a more detailed recap on the role of transformers in AI search, click here. 

How Constructor Uses Transformers 

At Constructor, we've integrated transformer models into the recall phase of our search architecture through the Cognitive Embeddings Search (CES) pipeline. While that might sound technical, the concept is straightforward: we use transformers to better understand what your customers are looking for.

Transformers start by understanding semantic meaning through embeddings 

When a shopper searches on your site, we use transformer models to convert both the search query and your product catalog items into what we call “embeddings,” or mathematical representations that capture meaning.

Think of embeddings as placing products and search queries in a multidimensional space where similar concepts are positioned close together. This allows our system to find products that are conceptually related to a search, even when they don't share exact keywords.

For example, if someone searches for "summer dress," our system understands that "floral sundress," "lightweight cotton dress," and "beach outfit" might all be relevant results, even if those exact phrases don't appear in the product title or description.

This UK retailer’s search engine returns a variety of dresses for the query “summer dress,” based on the shopper’s brand affinities, behaviors, etc. 

Transformers finesse the search experience from there 

We then create a mapping between common search queries and your catalog items based on these embeddings. Here's how it works:

• Building the knowledge base: We enrich your catalog dataset (with behavioral interactions and user queries) and process your entire product catalog offline, transforming each product into its embedding representation and mapping common queries to sets of relevant products 
• Real-time search: When a customer searches on your site, our system:
  • Checks if we've seen this or a similar query before
  • Retrieves the most relevant products based on both semantic understanding and real user behavior
  • If it's a new query, our system quickly creates an embedding and finds the closest matching products
• Continuous learning: Our system continuously improves as it learns from how shoppers interact with search results, creating a virtuous cycle of better understanding and more relevant results.

This approach is more efficient than traditional search methods, as we don't need to perform complex calculations from scratch for every search query.

Further exploration of improved architectures 

We're constantly refining our transformer models to provide even better search experiences. 

For example, we implemented real-time transformer processing for all queries. This means there is improved accuracy and relevance of search results, especially for unique or complex queries. 

Now, we’re working toward generating multiple embeddings per product. Rather than representing each product with a single mathematical representation, our goal is to create multiple embeddings per item to capture more nuanced product attributes and use cases.

How Clickstream and Transformers Enhance Search Performance 

What makes Constructor's approach truly powerful is the combination of transformer technology with real user behavior data, or clickstream data.

Clickstream: the power of real user behavior 

While transformers provide advanced language understanding, clickstream data shows us how real shoppers interact with your site. This includes:

• Which products they click on after searching
• What they add to cart
• What they ultimately purchase
• How they navigate through your site

By analyzing patterns in this behavior, we can predict which products are most likely to satisfy your customers' search intent, creating a more personalized shopping experience.

Why this combined approach works better 

The integration of clickstream data with transformer technology creates a search experience that's:

• Intent-focused: The system understands what customers are looking for, even if they express it imperfectly
• Personalized: Results are tailored based on what similar customers have found valuable in the past
• Resilient: The approach handles typos and different phrasings, ensuring customers find what they need even with misspellings
• Results-driven: By providing more relevant results, this method improves key business metrics like conversions and revenue

Traditional keyword matching and basic vector search simply can't match this level of sophistication. They typically rely on exact word matches or simplistic similarity measures, missing the context, intent, and actual preferences that drive purchasing decisions.

Business Impact of Transformers in Ecommerce 

Implementing transformer-based search technology isn't just about having the latest tech — it delivers measurable business results. Our customers have seen significant improvements across key metrics by using our AI-native tools enhanced with advanced technologies like transformers:

• Higher conversion rates: When customers find exactly what they're looking for quickly, they're more likely to purchase. Constructor clients have seen over 20% increases in search-driven conversions
• Increased revenue per search: More relevant results mean more purchases, with many clients seeing double-digit lifts in revenue per search session
• Reduced zero results: Transformer models better understand intent, dramatically reducing the frustrating "no results found" experience for shoppers
• Improved customer satisfaction: When search works intuitively, customers have a better shopping experience, leading to increased loyalty and repeat purchases
• Merchandiser efficiency: With AI handling the heavy lifting of search relevance, your merchandising team can focus on strategic initiatives rather than manually managing search rules

The Future of Transformers in Ecommerce Is Here 

As transformers continue to evolve, their applications in ecommerce expand as well. Here are some of the exciting ways transformers are already improving the overall ecommerce industry:

Conversational commerce 

As transformer models become more sophisticated, they're enabling truly conversational shopping experiences. Shoppers who ask complex questions like "What would look good with the blue dress I bought last month?" are now able to receive helpful recommendations tailored specifically for them.

Constructor is enabling this reality with our AI Shopping Agent (ASA). This Generative AI-powered technology mimics the experience of interacting with a knowledgeable store associate. It uses sentiment analysis to better gauge customer emotions and provide contextually appropriate, personalized responses. 

ASA can even be integrated into your search bar, suggesting and showing items aligned to a shopper’s needs, tastes, and intent within the native Search experience. 

By ensuring that every customer feels supported and understood, businesses that implement it improve conversions, AOV, LTV, and brand trust.

Learn more about conversational commerce and what it means for product search and discovery here. 

Visual search enhancement

Transformers are bridging the gap between text and images better. Customers are no longer limited to textual criteria to find products they want. Thanks to Google Lens-like Image Search capabilities, now shoppers can more easily bring what they see in real life to reality, creating a more intuitive discovery experience.

Voice-driven shopping

Likewise, as Voice Search grows in popularity, transformer models will be crucial in understanding spoken queries, which tend to be more conversational and less structured than typed searches.

Ready to Drive Revenue With Your Search Experience? 

Transformer technology, combined with real user behavior data, offers a powerful solution for ecommerce businesses looking to get and stay ahead.

Constructor's AI-native product discovery platform leverages the power of transformers along with clickstream data to deliver search experiences that truly understand customer intent and drive business results.

Ready to see how Constructor can transform your search and product discovery? Request a complimentary Search Experience Audit to uncover opportunities for improvement and revenue growth.