Definition
Grounding
Grounding is the practice of tying a language model's output to specific, verifiable sources — retrieved web pages, documents or a knowledge graph — supplied at generation time, so that each claim in the answer can be traced to evidence rather than to the model's training data. In answer engines, grounding is what turns a retrieved passage into a cited one.
The term comes from the model vendors' own APIs. Google's Gemini API offers Grounding with Google Search, which it describes as connecting the model to real-time web content so it can give more accurate answers and cite verifiable sources beyond its knowledge cutoff; the response carries grounding metadata that maps spans of the answer to source URLs. Anthropic's web search tool and OpenAI's web search tool return the same kind of span-to-source annotations. Consumer answer engines expose them as footnotes.
For a content owner, grounding is the stage where being retrieved becomes being cited. The model is handed a handful of passages and asked to answer using them. A passage that states the answer plainly, with the entity named and a figure or date attached, is easy to ground on: the model can quote or closely paraphrase it and point to it. A passage that hedges, buries the claim, or depends on context elsewhere on the page gets read for background and dropped from the attributions.
That is why answer-first writing and retrieval-augmented generation belong together: retrieval decides whether a model sees your passage, grounding decides whether it credits it. Grounding is also the main defence against hallucination, which is why engines prefer sources whose claims they can check against one another.
Frequently asked questions
Is grounding the same as retrieval-augmented generation?
They overlap but are not identical. RAG is the architecture — retrieve passages, then generate. Grounding is the goal that architecture serves: an answer whose claims are anchored to, and attributable to, the supplied sources. A RAG system can still produce ungrounded statements if the model ignores or contradicts what was retrieved.