What is grounding?
Also called: search grounding, grounded generation
Definition
Grounding is the process of tying an AI model's answer to external sources retrieved at answer time, such as live web search results, rather than relying only on what the model memorized in training. A grounded answer can include current information and cite the pages it used.
Explained
Google uses the term for Gemini's "grounding with Google Search" feature, and the idea applies to ChatGPT search and Perplexity too. When grounding is on, what ranks and what is written on retrievable pages today directly shapes the answer.
For brands, grounding is good news: it means you can influence answers within days or weeks by changing what is on the web, instead of waiting for a new model to be trained. It also means the same prompt can give different answers with and without search.
Example
Asked "best note-taking app" without search, a model lists apps it learned about in training. With grounding on, it reads this month's comparison articles and may include a newer app that launched after its knowledge cutoff.