Kurly Boosts 'No-Result Search' Revenue by Over 300% with Vertex AI Search Adoption

Kurly's Data Service Development Team adopted Google Vertex AI Search to solve No Result search cases. Previously, about 6-7% of all searches ended without results due to typos, spacing errors, and synonyms, leading to lower customer satisfaction and lost revenue.

Vertex AI Search Implementation and Data Configuration

Kurly chose Vertex AI Search for its ease of integration with the GCP environment and its hybrid search capabilities (text matching and embedding similarity analysis). To overcome the limitations of website crawling and improve accuracy, Kurly used BigQuery as its own data source, designating product names, detailed descriptions, and related keywords as search fields.

• Field Optimization: Assigned Key Property to product names and detailed descriptions to raise search priority, and set sale status as an Indexable field for filtering.
• Data Updates: Used the Python SDK to periodically synchronize BigQuery data to the Vertex AI Search data store.

Model Tuning and A/B Test Results

Beyond simple text matching, Kurly performed embedding vector tuning to teach the model the underlying meaning of thematic product collections, such as 'Gather, PSY Generation'. This enabled the model to capture semantic associations not present in surface-level features.

Over approximately two weeks of A/B testing, the experimental group outperformed on all metrics, including click-through rate, cart conversion rate, and purchase rate. In particular, as of one month after deployment, cart additions, purchase quantity, and revenue via AI search in No Result (NR) search cases increased by over 300%.

Future Tasks

Going forward, Kurly plans to establish a prohibited word restriction system, reduce costs through search result caching, and advance custom embeddings that reflect Kurly's domain characteristics.