Retrieval-augmented generation.In plain English.

The term comes from a 2020 research paper by Lewis and colleagues, presented at NeurIPS 2020, which combined a language model with an index of documents it could retrieve passages from before writing an answer.

Retrieval pulls passages, not whole pages, so a page whose sections each answer one question clearly is easier to use.

AI search products describe the same pattern in their own words. Google says AI Overviews and AI Mode may issue multiple related searches, which it calls query fan-out, and show links to supporting pages. Microsoft says Copilot fetches information from Bing when the web helps, and Perplexity says each answer includes numbered citations to its sources.

Being retrievable comes first. Google says a page only needs to be indexed and eligible to show a snippet to appear as a supporting link, with no extra technical requirements, and OpenAI says sites that block its OAI-SearchBot crawler won't be shown in ChatGPT search answers.

Example: someone asks which plumbers near them fix tankless water heaters. The assistant searches, pulls the paragraph from a plumber's tankless service page that names the brands it repairs, and cites that page in its answer.