Grounding

Grounding is the difference between "the model said it" and "the model said it because this page said it." Mechanically, a grounded answer pipeline retrieves a set of source passages (via web search, an embeddings index, or a fixed corpus), passes them to the model as context, and asks the model to answer using only those sources, with citations.

For publishers, grounding is the hand-hold. An engine that does not ground will rarely cite you. An engine that grounds heavily (Perplexity, ChatGPT Search, Google AI Overviews) produces a visible "Sources" list, and the contents of that list are the new ranking surface.

What gets retrieved into the context is the variable you can influence. The retriever favors pages with clear semantic structure, current dates, entity anchoring (schema.org), and content that directly answers the parsed query. The model favors passages with self-contained factual claims it can quote cleanly. Optimization is for both: be retrievable, and be quotable.