Models & technology

Grounding

Grounding ties an AI's answer to a verifiable source — search results, documents you supplied, or a database — in order to reduce hallucination.

Answers tied to something checkable

Grounding means connecting what an AI produces to a source you can verify. Rather than answering from general knowledge alone, the model uses search results, documents you provided, internal data, or a database as its basis.

The purpose is to reduce hallucination — plausible content that is not factual — and to make answers more accurate, more trustworthy, and more usable.

What counts as a source

  • Current information found through web search
  • PDFs, meeting notes, and manuals the user supplied
  • Internal documents and product databases
  • Figures pulled from an external service such as an analytics platform

A notebook-style tool that answers from documents you uploaded and shows where each claim came from is grounding in practice.

How it relates to RAG

RAG retrieves information relevant to the question and adds it to the model’s input. RAG is one of the main ways grounding is implemented.

Grounding is the broader goal of tying answers to evidence, and it can be achieved through web search, APIs, database queries, or tool calls as well.

Citations are not proof

Grounding does not guarantee a correct answer. The model can consult an outdated document, pick a passage only loosely related to the question, or misread the source.

For anything important, open the link or the quoted passage and check that the answer actually follows from it. Grounding does not remove the need to verify — it makes verification possible.

Related terms

Sources and review information

Last reviewed July 17, 2026

Back to the AI Glossary

Search this site