Definition
Grounding is Grounding is connecting an AI model to outside sources, such as web search results or your own documents, so its answer is based on that material and can point to where it came from.
Grounding means giving an AI model something real to stand on before it answers. Instead of relying only on what it absorbed during training, the model is handed current or trusted material and told to base its reply on it.
A grounded answer usually takes three steps. First, the system decides what information the question needs. Second, it fetches that information from a source: a web search, a company's documents, a database or a product catalog. Third, the model writes its answer using the fetched text, and often attaches links or citations to the passages it used.
The fetching step is often retrieval-augmented generation, and many people use the two terms almost interchangeably. A useful distinction: retrieval describes the lookup mechanism, while grounding describes the goal, an answer tied to a source.
A model on its own has a knowledge cutoff, so it knows nothing about later events. It can also produce a hallucination, a fluent answer that is simply wrong. Grounding attacks both problems. Fresh sources fix the recency gap, and a cited source gives you something to check.
It is not a guarantee. A model can still misread a source, quote it badly or ground itself in a page that is wrong. Treat the citation as a trail to follow, not proof.
Ask a plain chatbot what a software library changed last week, and it may guess. Ask a grounded assistant, and it searches the library's release notes, reads them and summarizes, with a link to the page it read. Same model, very different reliability.
You see grounding in AI search products, in assistants that read your files, and in developer features that let an app switch on a search tool for a model. Some providers now give website owners a separate setting for grounding, distinct from the setting for training data, because letting a model read your page to answer someone is a different act from teaching a model with it.
Grounding sits alongside tool use, since a model usually fetches sources by calling a search or lookup tool. It also depends on the context window, because whatever was fetched has to fit in it. And it is a core part of AI search, where grounded answers replace a list of links.
When you rely on a grounded tool, ask three things. What sources did it actually read? Are they current and trustworthy? Does the answer say what the source says? If you can click through and confirm in under a minute, the grounding is doing its job.
They overlap. RAG is a technique that retrieves documents for a model to use. Grounding is the broader goal of tying an answer to a source, and RAG is one common way to get there.
It reduces them but does not remove them. A model can still misread or misquote its sources, so check the citations.
Google released Nano Banana 2.1 on October 6, an update to its Flash-tier image model that, by Google's own evaluations, holds multi-character scenes together far better. If you make images with people in them or infographics with real text, test it this week.
A new robots.txt token, __usage_ai-training, lets a site state one preference against AI training, and OpenAI, Google, Microsoft and Anthropic say they will support it. It is a request, not a lock, and it does not cover AI answers.
Google says its Project Suncatcher prototype satellite ran Gemma 4 on a data center class TPU in orbit and sent back a haiku, a month after its Oct. 1 launch. The poem is a demo. The radiation and heat tests that come next are the real story.