Ecommerce ops · Production

Wayfair uses Google Gemini LLM to generate free-form personalized customer interests powering homepage product carousels

The problem

Traditional customer understanding models relied on fixed taxonomies and required extensive training data, missing implicit or latent interests not directly expressed in customer behavior.

Workflow diagram · grounded in source
1
Customer behavioral signals ingested
trigger
“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.”
2
In-house data compression
ai_action
“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.”
3
Gemini generates free-form interests
ai_action
“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…”
4
Interests clustered and deduplicated
ai_action
“We then cluster the generated interests into semantically similar themes to remove duplicates and ensure high-quality results.”
5
Interest metadata generated
ai_action
“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…”
6
Bi-weekly LLM-as-Judge validation
validation
“Running bi-weekly/monthly validation using specific tasks (Interest-Activity Alignment, Temporal Relevance, Concept Mapping) on a small sample to minimize cost.”
7
Personalized carousels displayed
output
“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.”
Reported outcome

Wayfair's new interest-based product carousels on the homepage are already driving measurable gains in engagement and revenue, while the in-house data compression model yields significantly lower operational costs.

Reported metrics
Historical data compressionup to 70%
Operational costsignificantly lower costs
Homepage engagement and revenuemeasurable gains in engagement and revenue
Reported stack
GeminiDynamic Page ConstructorUI Composersemantic search model
Source
https://www.aboutwayfair.com/careers/tech-blog/smarter-shopping-starts-here-how-ai-understands-what-youre-looking-for
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Wayfair's new interest-based product carousels on the homepage are already driving measurable gains in engagement and revenue, while the in-house data compression model yields significantly lower operational costs.

What tools did this team use?

Gemini, Dynamic Page Constructor, UI Composer, semantic search model.

What results were reported?

Historical data compression: up to 70%; Operational cost: significantly lower costs; Homepage engagement and revenue: measurable gains in engagement and revenue (source-reported, not independently verified).

How is this ecommerce ops AI workflow structured?

Customer behavioral signals ingested → In-house data compression → Gemini generates free-form interests → Interests clustered and deduplicated → Interest metadata generated → Bi-weekly LLM-as-Judge validation → Personalized carousels displayed.

Related ecommerce ops cases