Grounding
In short
Grounding is tying an AI model's answer to specific, checkable sources, such as search results or a company's documents, rather than letting it answer from memory alone. Grounded answers usually come with links or citations you can check.
- Also called
- AI grounding, Grounded generation, Grounding with search
- Topic
- Using AI tools
- Level
- Some background helps
A language model's built-in knowledge is frozen at training time and can be wrong. Grounding feeds it relevant material at the moment of asking and has it answer from that. Google's cloud documentation defines grounding as "the ability to connect model output to verifiable sources of information", and says it reduces the chances of the model inventing content1.
In practice grounding comes in a few forms. Google offers grounding in Google Search, Google Maps data, or a customer's own documents through retrieval-augmented generation (RAG)1. Its Gemini API returns grounded answers with inline citations that link each piece of text to a source URL, and can answer about events after the model's knowledge cutoff2.
Grounding is a goal and RAG is one way to reach it. RAG is the mechanism of fetching passages and adding them to the prompt; grounding describes the result, an answer anchored to those sources, ideally with links that show where each claim came from.
It is not a guarantee. A grounded model can still misread a source, pick a poor one or add details the source never said. Citations make that easier to spot, not impossible to happen.
Grounding has a cost. Google bills developers for each search query Gemini 3 and newer models run, with 5,000 free a month and then $14 per 1,000 searches, as of October 20262.
Example
Ask a grounded chatbot who won Euro 2024 and it searches the web first, then answers with citations linking to the pages it used. Google's own Gemini API documentation uses this very example2. An ungrounded model can only answer from its training data, which may be out of date.
Why it matters
Grounding is what lets you check an AI answer rather than take it on trust. When a tool shows its sources, you can see whether the claim really says what the AI says it does.
Questions people ask
What is the difference between grounding and RAG?
RAG is a technique: retrieve relevant passages and add them to the prompt. Grounding is the aim: an answer tied to verifiable sources. Google lists RAG over your own documents as one way to ground a model, alongside Google Search and Maps1.
Does grounding stop AI hallucinations?
It reduces them. Google says grounding reduces model hallucinations and anchors answers to your data sources1. But the model can still misquote or misread a source, so checking the linked citation is still worthwhile.
What is grounding with Google Search?
A Gemini API feature that lets the model run Google searches before answering. It returns the answer with inline citations linking parts of the text to source web pages, and can cover events after the model's training cutoff2.