Search Grounding

An AI language model only knows what it saw during training. Training means: the model once read a huge amount of text and learned patterns from it. After that, this knowledge is frozen — the model simply doesn’t know about new events. Search Grounding solves this problem: before the model answers, it retrieves current search results from the internet and incorporates this information into its response. The model invents less and stays closer to verifiable facts.

The expiration date of stored knowledge

Every language model has a so-called training cutoff — a cutoff date after which it has learned nothing new. This can be months or even more than a year in the past. Questions about current stock prices, an ongoing court case, or yesterday’s football result simply cannot be answered correctly by the model.

On top of that, there’s the phenomenon known as hallucinating: when a model doesn’t know something, it tends to produce a plausible-sounding answer anyway — which can be wrong. Search Grounding counteracts this because the model has to cite an external source instead of speculating on its own. This is a crucial difference, especially for applications in finance or news.

Process: searching, reading, answering

When a question comes in, the system first decides whether a search is necessary. Simple factual questions — such as asking for a mathematical formula — are answered directly by the model. For time-dependent or specific questions, it initiates a search query, usually via a connected search engine.

The search results — text excerpts from found web pages — are then passed to the model together with the original question. The model reads both and formulates an answer based on the retrieved content. Many systems also output source links so that users can verify the information themselves.

Technically, this approach is also called Retrieval-Augmented Generation, or RAG for short — roughly meaning “text generation supplemented through retrieval.” Search Grounding is the best-known variant of this, in which the source is the open web. However, RAG can also access private databases or company documents.

Search Grounding in products and headlines

Google's AI assistant Gemini uses Search Grounding directly via Google Search. Microsoft has built the same principle into Copilot, which relies on Bing. Both companies promote it as an argument against their models' hallucination problem. Perplexity AI is a startup that has made Search Grounding the central idea of its product: it works almost like a search engine, but formulates results as coherent text with source citations.

In the tech press, the term mainly comes up when new models or AI services are introduced. Search Grounding is often highlighted as a quality feature — a model that can search is considered more reliable than one that relies solely on its stored knowledge. Whether this is true depends heavily on the quality of the search results: bad sources produce bad answers, even if the model summarizes them flawlessly.