Glossary
Search grounding
A hosted web search the model runs inside one vendor request, then answers from, with citations.
Search grounding is a web search the model runs inside the vendor’s request. The harness enables the tool. The vendor searches, and the model answers with citations. Google’s grounding guide, read November 11, 2026, names this google_search and bills Gemini 3 and newer per query. OpenAI’s web_search and Claude’s web_search are the same shape on those APIs: the search happens before your code sees a snippet.
That is not RAG. RAG is a lookup your application runs and then places in the prompt. It is not a snippet tool either. A snippet tool returns text the harness can cut, store, or refuse. Grounding hides the page inside the vendor’s loop. The cap, when the vendor prints one, is a tool-call budget. Instructions hiding in a fetched page are indirect prompt injection.
The wiring choice, with the November 11 prices, is What is the best web search tool for an AI agent harness?. The chat products are Best AI agent for web search.