Grocery recommendations move from items to categories — and cart-adds rise 7.5%
Online grocery has cyclical repurchase patterns that item-level models struggle to scale on; category-level recommendations with constrained decoding lifted production cart-adds by 7.5%.
Key numbers
- Problem: online grocery recommendations must capture cyclical buying behaviour and diverse intents, while traditional item-level methods face scalability and accuracy limits — category-level recommendation is more structured and practical.
- Method: GROCLM uses a two-stage LoRA training strategy to encode cyclical purchasing patterns directly into model parameters (better use of repurchase signals than prompt conditioning), plus trie-based constrained decoding over a predefined category space for valid, controllable output.
- Validation: consistent gains over strong baselines on proprietary production data and a public benchmark, generating all categories jointly to keep inference efficient.
- Production result: in a live restocking task, a 7.5% relative improvement in cart-adds per impression.
Method and sample
arXiv paper “GrocLM: Grocery Category Recommendation in E-Commerce with Large Language Models” (2026-06-04): a two-stage LoRA fine-tuned category recommender with trie-based constrained decoding, deployed in a real production environment and evaluated on proprietary data, a public benchmark and a live restocking task.
What it means for brands
Retail competition in AI is moving from page impressions to being placed into the cart: entering the category-level candidate set matters more than rank position. Models rely on structured product attributes and repurchase signals — the reason product data is worth cleaning up.
How SourceSight handles it
For retail and e-commerce projects SourceSight uses purchase-decision prompts (which to choose, compare, pricing, is it right for me), prioritises review and directory sources off-site, and structures category, specification and certification data so models have usable fields.
Limits
The 7.5% cart-add lift comes from one platform's restocking task and depends on category mix and traffic; the result reflects that production setting rather than retail generally.
References
Run it on your own project
What the same prompts return across 10 engines is only knowable by running them.
Related research
Cross-category recommendations: +36% add-to-cart on item pages, +15% on cart pages
Recommending general merchandise to grocery shoppers — a milk frother with the milk — has long been hard; LLM retrieval plus real-time cart context moved production A/B metrics.
Industries · 2026-10-07Home services: 22% of homeowners already use AI to find a provider, and 80% of businesses don't know how to be found
The local services funnel is moving upstream: before a homeowner opens a website or calls, AI has already synthesised reviews, reputation and service descriptions into one answer.