Advancing Menu Content with AI: How DoorDash uses AI to generate menu descriptions

Yilin Ren, Zheng Hu, Ying Yang, Joel Moniz30 Jun 2025

Problem

Restaurants need a description for every menu item, because descriptions drive online visibility and orders, but owners rarely have time to write them. An LLM can write at scale, yet the text must be accurate and sound like the restaurant. An item with only a name gives the model nothing to ground on, so it writes generic text or invents ingredients. Quality must also be checked continuously.

Example from the article: DoorDash generates menu item descriptions for merchants, who can pick a tone (descriptive, concise, playful or enthusiastic) and regenerate instantly.

Solution in brief

Build three connected parts. First, retrieve context for each item: its options, photos and descriptions of similar items in the same cuisine. Second, assemble the prompt from whatever context exists, add few-shot examples from similar merchants, and generate in the chosen tone. Third, filter outputs automatically, review by hand, and A/B test each model before scaling.

Questions this case answers
  • How to generate accurate, individually styled text for catalog items at scale when many items have almost no source data
  • How to control the quality of generated content at scale when automatic checks cannot catch every error
  • How to generate accurate item descriptions with an LLM when many items have little or no source data?
  • How to make generated text match each seller's own voice and style?

Tools and methods

  • Structured data extraction — item options (protein choices, sizes), categories, metadata
  • Vision model — pulls ingredients and cooking methods out of item photos
  • Embedding similarity search — finds descriptions of similar same-cuisine items for cold-start items
  • Retrieval-augmented generation (RAG) — adds retrieved descriptions to the prompt when item data is sparse
  • Conditional prompting and few-shot examples — prompt built from available data plus examples from similar merchants
  • Automated filters, human review, A/B tests — hybrid quality control
Continue reading — it's free

Google or a one-time email code. No password.