---
term: "Grounding"
canonical: https://citable.wiki/glossary/grounding
license: CC BY 4.0
---

# Grounding

**Definition:** 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](https://ai.google.dev/gemini-api/docs/grounding), 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](/guides/how-to-write-answer-first-content) and [retrieval-augmented generation](/glossary/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](/glossary/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.

## Related guides

- [What is Answer Engine Optimization (AEO)?](https://citable.wiki/guides/what-is-answer-engine-optimization)
- [How to write answer-first content that LLMs can quote](https://citable.wiki/guides/how-to-write-answer-first-content)

---

Source: https://citable.wiki/glossary/grounding