Target rebuilds product discovery on Google Cloud Spanner

Target has rebuilt its retail discovery platform on Google Cloud Spanner Graph, replacing a fragmented mix of search and transactional systems with a single data platform.

The change affects the systems behind product discovery features run by Target's Guest Product Confidence platform team, including ratings, reviews and its Gift Finder chat agent on the retailer's website and app.

Target had previously relied on Elasticsearch clusters for search and separate NoSQL datastores for transactional data. According to the company, that setup created persistent synchronisation problems between search, vector and transactional databases and increased the engineering work needed to maintain data consistency.

It also made expansion harder, as new retail data domains required additional database collections, more complex joins and looser transactional links across systems. Under the earlier design, the retailer could not query graph relationships, vector similarity and keyword search indexes within a single transaction.

Unified platform

After assessing standalone vector and graph databases, Target chose Spanner Graph as the basis for what it describes as an enterprise ontology. The system combines semantic data, graph relationships, vector embeddings and transactional records in one database environment.

Under the new architecture, one layer captures catalogue data and metadata from multiple internal sources and applies generative AI-based enrichment to product information. A second stores entity nodes, relationship edges and vector embeddings in Spanner Graph, while a third supplies structured context to the large language models used in conversational shopping tools.

The arrangement is designed to support product relationships and shopper intent in ways that go beyond keyword search. Gift Finder is one example, using conversational prompts to help shoppers identify items based on context rather than basic category browsing.

Migration process

Target moved the platform in four stages to avoid downtime for shoppers. It first mapped retail entities such as products, categories, brands and guest preferences into a Spanner Graph schema.

It then built a parallel data integration pipeline to replay live transactional updates, apply schema changes, generate embeddings and write the results into Spanner Graph in real time. Target gradually shifted read traffic through a canary deployment before completing the cutover and retiring its legacy Elasticsearch stack.

The final stage removed the maintenance burden associated with those clusters and consolidated search and transactional workloads that had previously been spread across separate technologies.

Operational gains

Target reported a 50% reduction in infrastructure maintenance after combining its transactional NoSQL and search index databases into Google Cloud's managed service. Developers now spend less time on database administration and upkeep, according to the company.

The new setup also lets teams query relational catalogue data and graph relationships together using SQL and Graph Query Language. That removes some of the duplication and extract, transform and load work previously required to bridge different data models.

Another goal was to improve how AI systems retrieve product context. Target said traditional retrieval-augmented generation methods that rely mainly on vector similarity can miss structured links between items, such as compatibility or age suitability. Combining graph traversal with vector search, it said, gives its language models more precise grounding.

The retailer linked that approach to improved recommendation relevance, stronger customer satisfaction and a higher Net Promoter Score, though it did not disclose figures. It also said Spanner's autoscaling handled sharp traffic swings during major retail peaks, including Black Friday and Cyber Monday.

The project reflects a broader shift among large retailers toward combining transactional databases, search and AI retrieval in fewer systems rather than stitching together separate platforms. For Target, the result is a single source of truth for both operational retail data and semantic product intelligence.