GLOSSARY
Grounding
Grounding / RAG
One sentence of definition, and how it differs from the concepts it gets confused with. The definition is written to be quoted as it stands.
DEFINITION
Grounding is the technique of generating an answer based on actual documents fetched by search, also known as retrieval-augmented generation (RAG). It improves accuracy and currency, and makes source citation possible.
Easily confused with
It is not training data
Training happens once when the model is built; grounding fetches documents anew with each question. So the route by which a page you fixed yesterday can appear in an answer today is grounding, not training.
Grounded does not mean accurate
Fetching a source and reproducing it faithfully are different things. Statements not present in the body are observed being attributed to a source even when a citation link is shown.
In practice
- To keep the grounding path open, allow the search crawler, not the training one. They are controlled separately in robots.txt.
- State an update date on documents where currency matters — grounding tends to prefer recent documents.
Related terms
Where this leads
- Should you allow AI crawlers in robots.txt? Whether to open or block GPTBot, ClaudeBot and PerplexityBot — how to decide by business type, with working configuration examples.
- Which websites does ChatGPT find? AEO read from OpenAI's official documentation Taking OpenAI's ChatGPT Search documentation as the primary source: how ChatGPT searches the web (query rewriting, follow-up searches, source citation), the official conditions for a site to be found, and what those facts mean in practice.
What is AI saying about your brand right now?
Our free audit finds out. We ask real questions in the major answer engines and report what was cited and what was said.
We reply within one business day.