Grounding is the anchoring of a language model's answer in concrete, verifiable sources instead of free generation from the model's parameters. In practice, this means that before answering, the system retrieves documents, inserts them into the context, and the model formulates the answer solely from them, maintaining a link to the source for every claim. Grounding is precisely the reason why links to websites appear in AI answers at all, and why it makes sense to optimise for generative search engines. For content, the key point is that the model will favour passages that confirm a claim unambiguously and without having to fill in gaps: specific figures, one-sentence definitions, validity dates, parameter tables. Vague marketing phrasing cannot be grounded, and so it will not appear in the answer, even if the page otherwise ranks well in classic search.
See also: Content chunking, Reranking, Information gain.
