Food delivery recommendation is now two-stage: guess the craving, then rank the vendors
For repeat orders, ranking is no longer pure CTR prediction: cuisine prediction comes first, vendor ranking second, and the LLM works in the re-ranking step.
Key numbers
- Framework: MARS splits repeat-order recommendation into two stages — cuisine prediction, then vendor ranking — combining LightGCN global preference signals, Swing-based local peer evidence, geospatial filtering and prompt-driven LLM reasoning over behavioural, temporal and geographic context.
- Evaluation uses two real-world Delivery Hero benchmarks (DHRD-SE, DHRD-SG) against heuristic, sequential, graph-based and food-delivery-specific baselines.
- One conclusion: strong pre-trained backbones paired with lightweight collaborative retrieval are already competitive for repeat orders, and the LLM's value sits in re-ranking rather than replacing the pipeline.
Method and sample
arXiv paper “MARS: Multi-Agent Re-ranking for Repeat-Order Food Delivery Recommendation” (2026-07-28): a modular multi-agent re-ranking framework performing two-stage cuisine prediction and vendor ranking with collaborative, peer, geospatial and LLM signals, evaluated on two real-world delivery benchmarks with public implementation and evaluation protocols.
What it means for brands
The online front door for restaurants is layered: making it into the “what do I feel like eating” layer decides whether you reach the vendor-ranking layer at all. What matters is not a single impression but whether you are thought of at all for a category or occasion.
How SourceSight handles it
For food and local projects SourceSight uses real ordering questions (nearby, directions, opening hours, price level, delivery), reports mentions and position per engine, and fixes the question types where you are entirely absent.
Limits
The evaluation uses two enterprise delivery benchmarks focused on repeat orders; the findings concern recommendation ranking internals rather than how consumers phrase questions to AI assistants.
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.