Every day, Etsy buyers visit our platform with a diverse selection of unique, handmade, and vintage items. However, traditional search and recommendation systems struggle to capture the nuanced interests that make each Etsy buyer unique, from specific styles to aesthetic preferences. At Etsy, our goal is to deliver engaging, personalized experiences that cater to each individual buyer’s unique interests.
To address this challenge, we recently explored utilizing large language models (LLMs) to create detailed buyer profiles based on user browsing and purchasing behaviors. These profiles are built in strict alignment with privacy laws and agreements such as GDPR and CCPA, ensuring that buyers’ personal information remains protected. In addition, we will provide buyers with a clear opt-out option, allowing them to opt-out of having a profile generated.
Understanding Buyer Profiles:

Buyer profiles anonymously capture nuanced interests for each user, including preferred styles, product categories, shopping missions, and other insights that help us better understand our users. For example, do they prefer minimalist styles? Do they mainly shop for home decor, jewelry, or a combination of both? They also offer insight into specific shopping missions, such as finding the perfect fall sweater. These profiles help us personalize the Etsy experience and make it feel tailored to each individual buyer.

Technical Implementation:

Our process begins by

Every day, shoppers from Etsy’s community of nearly 90M buyers visit our marketplace to search for unique, handmade, and vintage items. But with over 100 million listings, how do we help each buyer find exactly what they’re looking for? Traditional search and recommendation systems often fall short of capturing the nuanced interests that make each Etsy buyer unique, from specific styles to aesthetic preferences. At Etsy, understanding our buyers’ interests is central to delivering engaging, personalized experiences.

Recently, we explored enhancing our personalization by leveraging large language models (LLMs) to create detailed buyer profiles based on buyers’ browsing and purchasing behaviors. We build these profiles in strict alignment with privacy laws and agreements such as GDPR and CCPA. In addition, we will be providing buyers with a clear opt-out, so that they can choose not to have a profile generated at all.

Understanding Buyer Profiles

Buyer profiles anonymously capture nuanced interests for each user, like preferred styles and product categories. For example, do they tend to prefer minimalist styles? Do they mostly shop for home decor, or jewelry? They also offer insight into specific shopping missions a buyer may be on, like finding the perfect fall sweater. These profiles help us better understand our users, making the Etsy experience feel tailored to each individual buyer.

Technical Implementation

Our process begins by retrieving user activity data and listing information from internal sources like our feature store and BigQuery. This includes users’ recent searches, item views, purchases, and favorites.

Once we have the users’ session data with interactions, we then prompt the LLM to interpret this data and generate structured buyer profiles.

In the case that the LLM does not have enough supporting data to infer a user’s categorical interests (i.e. these interests would also not be obvious to a human looking at this user’s interactions), these fields and the confidence hashmap can be empty. The LLM will always include an explanation and any observed interests in the explanation section of the buyer profile.

Bringing Personalization to Scale

Initially, using this method to generate profiles for our roughly 90 million buyers would have been exceedingly costly and taken weeks to update.
We optimized this by:

• Shifting our listings data source from API endpoints to BigQuery tables that are clustered and partitioned for efficient querying.
• Decreasing the volume of input tokens. Initially, we were including about 2 years of session data. Now, we’ve reduced this to just the last 9 months. The 9 month timeframe also allowed us to lighten the prompt corrections (and cached input tokens) by reducing the weight of holiday shopping.
• Increasing LLM and BigQuery batch sizes for data processing.
• Introducing parallel processing with managed concurrency to avoid request rate limits.
• Scaling up computational resources for session data retrieval and LLM processing tasks.

These improvements reduced buyer profile generation time dramatically — from 21 days down to 3 days for 10 million users.
Cost management was also crucial. By adjusting the prompt to get high quality results with a smaller model, we significantly lowered the cost and made large-scale personalization economically feasible. Through these various cost management techniques, we were able to reduce the estimated cost by 94% per million users.

Additionally, to scale and schedule buyer profile refreshes, we used Airflow as an orchestration tool. By batching and staggering tasks by user_id, we parallelize some tasks while avoiding too many concurrent requests on BigQuery & OpenAI APIs (as visualized in this Directed Acyclic Graph).

Experiments and Applications

With these powerful insights in hand, we’ve started to explore several key applications of buyer profiles to make the Etsy search experience feel unique to each shopper.

Query Rewriting

Query rewriting is the process of transforming a user’s submitted search query on the backend to better represent their intent to the underlying search systems. We already use query rewriting for different use cases at Etsy. It can correct spelling, add specifics to ambiguous queries, adjust niche terminology, or complete incomplete queries (where the user accidentally pressed enter before they finished writing their query).

To further personalize Etsy’s search experience, we enrich search queries with predicted interests from buyer profiles. For instance, a simple query like “cool posters” becomes “cool posters + hippie|boho|vintage|nature,” significantly boosting relevance.

Search results without personalization

Search results with personalization using example buyer profile

Refinement Pills:

On the Etsy website, “refinement pills” work as an interactive query reformulation. When a user clicks on an option, the path will be prefixed to the user’s query and a new search will be executed. A mix of different types of refinement pills are generated: single pills (just one word) and grouped pills (a dropdown of options). We present high-confidence buyer interests as clickable filters or refinement pills in the search interface, allowing users to effortlessly refine their searches based on predicted preferences.

Refinement pills without personalization

Refinement pills with personalization using example buyer profile

For example, if a user clicks on a “daisy” refinement pill as in the example below, the search query “car accessories” will be reformulated to “daisy car accessories” and the new, more specific search will be executed.

Measuring Success

Validating Profile Accuracy

As we roll out buyer profile-based personalization, understanding the accuracy of our predictions is crucial. Through systematic experimentation, we’re working to establish how well our LLM-generated profiles reflect actual user interests and shopping behaviors.

Key Accuracy Metrics:

• Click-through Rate (CTR) Lift: Comparing CTR on personalized search results versus baseline helps us quantify immediate engagement improvements
• Conversion Rate Impact: Tracking purchase rates from personalized searches reveals whether we’re surfacing items users actually want to buy
• Refinement Pill Engagement: Monitoring clicks on suggested refinement pills helps measure how well we’ve predicted user interests
• Search Query Reformulation Success: Analyzing subsequent user actions after query rewriting indicates whether our enrichments align with user intent
• Profile Refresh Strategy: We work to maintain profile accuracy through several methods:
• Dynamically refreshing timing based on user activity levels. These include the number and frequency of recent interactions. For example, does the user click on the refinement pills proposed? Are the personalized items viewed?
• Detecting interest drift when search patterns change significantly
• Taking into account seasonal considerations, such as the winter holiday shopping season
• Monitoring performance degradation signals when personalization effectiveness drops below thresholds of engagement listed above

Future Work

The Cold Start User Problem

To extend this personalized experience to new users who lack extensive session data, we experimented with “inheritance profiles” using collaborative filtering. By matching early interaction signals (like brief session data or initial search terms) to existing profiles of similar users, we could predict interests for new users earlier in their relationship with Etsy.

This user-based collaborative filtering approach would allow us to have profiles for as many users as possible, even those with minimal behavioral data. Below is a diagram for a sample new user who might inherit aspects of a similar profile as the one above, based on just a few searches.

Conclusion

As we continue to refine these profiles and expand their applications across the Etsy marketplace, we’re excited about the possibilities, from more intuitive search experiences to discovery features that surface hidden gems aligned with each buyer’s unique tastes.

With nearly 90 million active buyers and over 100 million listings, the challenge of personalization at Etsy’s scale is immense, but that’s exactly what makes it worth solving. Every improvement we make impacts millions of shopping journeys, helping buyers discover that perfect vintage find, custom wedding gift, or miniature dragon they didn’t even know they were looking for.

Acknowledgements

Huge thanks to Orson Adams, David Blincoe, Jugal Gala, Davis Kim, Haoming Chen, Yinlin Fu, Julia Zhou, and the entire Search Mission Understanding team for their invaluable support and guidance.