Key takeaways
- Patrick Lewis and colleagues described RAG in a paper first submitted to arXiv on May 22, 2020. [1]
- The paper combined a pretrained generation model with a searchable Wikipedia index. [1]
- As of Oct. 29, 2026, Google calls its retrieval-based Search technique RAG or grounding. [2]
Definition: Retrieval-augmented generation combines a generative model with an external information source, allowing retrieved material to inform the generated response rather than relying only on knowledge stored in model parameters. [1]
Where does the term come from?
Lewis and colleagues’ paper, Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, was accepted at NeurIPS 2020. [1] The arXiv version was last revised on April 12, 2021. [1]
The paper’s architecture paired a pretrained sequence-to-sequence model with a dense vector index of Wikipedia accessed through a neural retriever. [1]
How does it work in Google Search?
As of Oct. 29, 2026, Google says its core Search ranking systems retrieve relevant, current pages from its index. [2] Its systems then use information from those pages to generate responses with links to supporting pages. [2]
Google’s crawler documentation describes grounding as supplying indexed content to the model when a prompt is processed. [3]
What does the evidence support?
| Question | Supported answer | Limit |
|---|---|---|
| Did the original research compare RAG with a model-only baseline? | The authors report more specific, diverse, and factual language on their generation tasks. [1] | The abstract does not establish performance for today’s commercial search products. [1] |
| Does Google document RAG in Search? | Its optimization guide explicitly describes the technique. [2] | This establishes Google’s stated mechanism, not every engine’s implementation. [2] |
What matters for publishers?
As of Oct. 29, 2026, Google’s eligibility guidance requires pages to be indexed, eligible for snippets, and included in Search generative AI features. [2] Google cautions that satisfying these requirements does not guarantee indexing or display. [2]
AVJ’s methodology also cautions that a citation is not proof that a source shaped the answer, while an uncited source may have done so. [4]
Related terms
Grounding · Query fan-out · Citation
How this was made: AVJ used AI tools to help draft this definition and compare it with the cited sources. Nhi Phan reviews the published version.
Sources
[1] Lewis, Patrick, et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arXiv:2005.11401; first submitted May 22, 2020; revised April 12, 2021; accepted at NeurIPS 2020; accessed Oct. 29, 2026. https://arxiv.org/abs/2005.11401
[2] Google Search Central. Optimizing your website for generative AI features on Google Search. Updated July 10, 2026; accessed Oct. 29, 2026. https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
[3] Google. List of Google’s common crawlers. Updated Oct. 15, 2026; accessed Oct. 29, 2026. https://developers.google.com/crawling/docs/crawlers-fetchers/google-common-crawlers
[4] AI Visibility Journal. AVJ Citation Index — Methodology. Methodology v0, published Oct. 23, 2026; accessed Oct. 29, 2026. /index/methodology/