Introduction

In e-commerce, true personalization begins with deep customer understanding, knowing not just what customers click, but why. Traditional customer understanding models predict affinities such as style or brand preferences from behavioral data like clicks, searches and purchases. However, they rely on fixed taxonomies, require extensive training data and often miss implicit patterns or latent interests that are not directly expressed in behavior.

For example, a customer living in a NYC studio apartment who searches for a foldable dining table and a sofa bed is likely optimizing for space, even though “space saving” never appears in her searches.

At Wayfair, our vision for personalization is both simple and ambitious: “Know The Customer Better Than They Know Themselves.” This means building systems that

• Anticipate customer needs and stay one step ahead
• Deliver intuitive, personalized experiences
• Foster trust by aligning with customers’ core values

To move beyond traditional models, we turned to generative AI (GenAI) to help interpret behavior in richer, more human ways. Powered today by Google’s Gemini large language model (LLM), our system generates free-form and personalized customer understanding (known as interests), such as “Space-optimizing furniture”, “Boho chic bedroom accents” or “Modern earthy decor aesthetics”.

These LLM-generated interests are scalable, free form and dynamic, allowing the system to evolve with changing customer needs, behaviors, lifestyles and trends. Their inherent explainability provides deeper insight into how these interests form, enabling more transparent and adaptable personalization systems.

Our new interest-based product carousels on the homepage are already driving measurable gains in engagement and revenue and this is just the beginning.

What Are Customer Interests?

Customer interests are nuanced, contextual insights about a customer’s preferences that go beyond surface-level understanding. An interest can include (but is not limited to):

• Functional Needs: Usage or purpose-driven requirements (e.g., “ergonomic workspace setup”,  “child-friendly furniture”)
• Preferences: Tangible attributes such as material, color or price range (e.g., “oak finishes”,  “budget-conscious décor”)
• Values: Lifestyle or aesthetic principles (e.g., “eco-conscious living”,  “minimalist modern design”)
• and much more

By leveraging LLMs, we can interpret customer interactions such as searches, product views and add to carts to reveal nuanced and evolving interests in a natural and free form way. This goes beyond rigid taxonomies and enables personalization that feels intuitive, relevant and uniquely tailored to each customer.

How Are Customer Interests Generated?

We feed customers’ search queries, viewed product details, Add to Cart (ATC), wishlist and purchase signals from the historical data into Gemini using curated prompts to generate customer interests. An evolving challenge in the GenAI space is the trade-off between context, cost and performance (quality of results). We developed an in-house model that compresses the historical data by up to 70% yielding significantly lower costs without compromising with the available information. We then cluster the generated interests into semantically similar themes to remove duplicates and ensure high-quality results.

Following metadata about each interest is generated and stored.

• Confidence Level: LLM is prompted to provide low, medium and high confidence levels indicating the strength of the generated interest, typically correlating with frequency and consistency in recent activity. 
• Reasoning: Provides a human interpretable reasoning for the generated interest. This helps to quality check the model outputs and revisit prompts on a regular basis.
• Query: Provides a search query to fetch product list using Wayfair’s proprietary semantic search model. 
• Title: Provides a quirky short title that can be used over a product carousel that populates the product results of the generated query. Prompt is designed to generate titles that resonate with customers and are compliant with brand guidelines.

A key design consideration in the interest generation pipeline is determining how often customer interests should be refreshed – it directly impacts the operational cost and the customer experience. While many interests are durable and reflect long-term preferences, customer lifestyles, behaviors and needs naturally evolve over time. For example, a customer living in a compact city apartment may be drawn to space-efficient decor, but after moving to the suburbs, their preferences could shift toward larger, more substantial furniture. To balance personalization accuracy, system efficiency and computational cost, we experimented thoroughly and found a biweekly cadence for regenerating customer interests based on the interests’ durability analysis works the best.

Design

The Interest Generation System is built to operate efficiently at scale, generating personalized interests for millions of Wayfair customers. When a customer lands on Wayfair’s homepage, the Dynamic Page Constructor retrieves their interests and associated metadata. It processes and ranks them before passing to the UI Composer that creates personalized, interest-driven product carousels. The products in the carousels are retrieved by Wayfair’s search model.

These interests are generated offline through large-scale batch jobs that process rich customer engagement data and are then stored for real-time use.

Interest Validations

We’ve conducted multiple offline experiments to validate the performance and reliability of the Gemini-powered interest generation system. These evaluations focus on areas like assessing Gemini’s ability to comprehend and reason over customer context. Measuring how well the generated interests align with customer intent and whether they remain valid over time. We validate generated interests across multiple dimensions — consistency, alignment and predictive power through the following experiments.

1. Alignment with Predictive Models: Evaluating interests against existing predictive models such as class prediction, price and style affinity models. We verified user journeys and matched their preference with generated internet overlaps to ensure the interests don’t diverge significantly.

2. LLM as Judge: Running bi-weekly/monthly validation using specific tasks (Interest-Activity Alignment, Temporal Relevance, Concept Mapping) on a small sample to minimize cost.

3. Out-of-Time (OOT)/Delayed Validation: Assessing predictive power by observing customer purchases one to three months later.

Together, these evaluations ensure that generated interests are accurate, durable and predictive of future behavior.

Use Cases

• Interest-Based Product Carousels: The generated customer interests power personalized product carousels on the homepage, enabling proactive discovery of items aligned with each customer’s overall shopping journey, including their searches, browsing patterns and purchase behavior. We leverage Wayfair’s in-built semantic search to retrieve products that best align with each customer’s interests and surface them as personalized carousels. These experiences feel intuitive and emotionally resonant, helping customers feel seen and understood. The descriptive carousel titles also improve explainability, increasing the likelihood that customers will connect with and engage with the recommended products. 

• Anchor-Based Retrieval for Product Detail Page Recommendations: Using a bank of interests generated from millions of active customers, we identify interests for each product and surface contextually relevant recommendations directly on the Product Detail Page. Shoppers at this page have a clearer idea of their general need and characteristics, by recommending relevant yet diverse products, we help them to compare similar products, look for add-ons or get inspired to explore more products.

Future Steps

We’re continuously advancing how Wayfair uses AI to understand and inspire our customers. In the next phase, we plan to expand the system in several exciting directions:

• Enhanced Prompt Intelligence: Incorporating third-party demographic data and in-house model outputs (e.g., brand affinity) into the LLM prompt layer to refine the quality and precision of generated interests.
• Customer Segmentation: We aim to extend customer interests data beyond on-site personalization to enhance Wayfair’s marketing and audience intelligence systems. We can dynamically group customers into segments based on shared emerging themes, such as “space-optimizing furniture” or “modern rustic decor.” Marketers can then target these segments with campaigns and emails tailored to specific interests or product categories. Interest-based campaigns not only increase relevance and engagement but also inspire customers with ideas that align with their tastes and needs, encouraging proactive product discovery.

The authors would like to acknowledge and thank Praveen Hegde, Iain Greer and the product recs MLS and engineering teams for their suggestions to Customer Interest Generation.