Grounding
Definition: Grounding is the process of connecting an AI model's answer to verifiable source material, such as search results or a document set, so the response reflects current facts instead of only the model's training data.
Also called: AI grounding, grounded generation, grounding with search
Updated
Grounding versus training
A model learns from its training data once, and that knowledge is frozen at a cutoff date. Grounding happens at the moment a question is asked: the system fetches source material, often from a search index, and hands it to the model so the answer reflects what those sources say now.
Google says grounding means providing content from the Google Search index to the model at prompt time to improve factuality and relevance. Its Google-Extended robots.txt token controls whether content Google crawls may be used for grounding in Gemini Apps and in Grounding with Google Search on Vertex AI (Google’s crawler documentation). It does not apply to AI Overviews or AI Mode, which are part of Search.
Why it matters to site owners
Grounded answers are the ones that cite sources, which makes grounding the mechanism behind AI referral traffic and attribution. An answer produced from training data alone usually names no source at all, so even a brand the model knows well gets no link.
Each assistant grounds on its own index. Google’s products ground on Google’s index; ChatGPT search relies on OpenAI’s search crawler, OAI-SearchBot, among other sources; Perplexity uses PerplexityBot. Blocking one of those crawlers removes you from that assistant’s grounding pool, even if you remain in its training data.
How to check it
- List which search crawlers your robots.txt and firewall allow, assistant by assistant.
- Ask each assistant a question your page answers and see whether it cites a source at all, and which one.
- Repeat over several weeks; grounded results change as indexes update.
The AI crawlers guide maps each crawler to the product it grounds.