Grounding

A language model learns from a huge pile of text, and then it stops learning. Without help, it can only answer from that memory. Grounding hands it fresh material to work from.

Google describes it as retrieving relevant, up-to-date pages from its Search index to make AI answers more accurate and fresh. The pages it used can then be shown as links next to the answer.

Say someone asks an AI assistant when a local bakery opens. A model working from memory might guess. A grounded model first searches, finds the bakery’s own page and answers from that, with a link to it.

Grounding is the moment your page can be picked. Google says its AI features rely on its core Search ranking systems to retrieve the pages they use.

Bing even reports it: its AI Performance report shows "grounding queries", the phrases the AI used when it retrieved pages that were then cited.

Google also uses the word for Gemini. The Google-Extended setting controls whether Google may use your content for grounding in Gemini Apps.

• Make sure your page can be crawled and indexed. Google says a page must be indexed and eligible to show a snippet to appear in AI Overviews or AI Mode.
• Put the answer to a likely question in plain text near the top of a section.
• Keep facts like prices, opening hours and addresses up to date on your own site.
• In Bing Webmaster Tools, check the AI Performance report to see which of your pages were cited.

RAG

RAG is the technique; grounding is what it achieves. Google’s guide even calls RAG "also known as grounding".

Training

Training happens once, before a model is released. Grounding happens again for every question.