Target has deployed a generative AI-based accessory recommendation system to address the growing complexity of pairing complementary products across its large retail catalog. Developed by the Product Recommendations team, GRAM, a GenAI-based Related Accessory Model for the Home category, uses large language models to identify which product attributes matter most when recommending related accessories, helping shoppers find items that go well together.
As emphasized by Target engineers, the sheer size of the company’s catalog and the diversity of item attributes make providing high-quality accessory recommendations a complex challenge. Depending on the product, shoppers may prioritize color, material, brand, intended audience, or other factors. Traditional rule-based approaches can be brittle and difficult to scale, particularly when hundreds of attributes influence what constitutes a suitable accessory. GRAM addresses this challenge by automating the assessment of attribute importance across categories, reducing the need for manual curation while maintaining recommendation relevance.
At the core of GRAM’s design is the use of large language models to analyze structured product data. The model evaluates which attributes are most significant for each core-accessory pairing, assigns weights accordingly, and generates relevance scores that determine which items are presented to shoppers. For example, when recommending a throw pillow to go with a sofa, the system might emphasize color and material, while suggesting a battery for a toy prioritizes compatibility and kid-safety features.
To ensure business relevance, Target combined the generative AI model with a human-in-the-loop process, allowing merchants to provide curated lists of co-purchased or seasonally relevant items. This approach ensured that while the LLM handles scalability and discovery, human input guides cross-category pairings and maintains alignment with merchandising goals.
Adnan Awow, Senior Director of Data Science at Target, mentioned this in a LinkedIn post:
With GRAM, our GenAI-powered Related Accessory Model, we automatically prioritize hundreds of product attributes such as color, material, and brand to surface the most relevant add-ons first. The system also captures aesthetic cohesion, suggesting visually harmonious pairings like pillowcases that complement sheets. Designed to operate at massive scale, GRAM scores products at the type level, handling hundreds of thousands of Home SKUs, and is ready to expand across categories, said Adnan Awow, Senior Director of Data Science at Target.
As reported in the blog post, early A/B testing, GRAM produced measurable improvements in engagement and conversion. Add-to-cart interactions for Home category accessories increased by approximately 11 percent, display-to-conversion rates rose by 12 percent, and attributable demand grew by 9 percent. Following these results, the system was deployed fully into production, reflecting Target’s strategy of integrating AI into digital shopping experiences to enhance personalization and recommendation quality.
Product recommendation before and after GRAM model (Source: Target Tech Blog)
Target engineers say GRAM has created a scalable system for recommending accessories across diverse categories. The approach improved user engagement and enabled more seamless shopping experiences. The model was fully deployed to production in April 2025, demonstrating Target’s ongoing efforts to enhance personalization and recommendation quality across its digital channels.
