AI search

Grounding

answer grounding · retrieval grounding

Definition

Grounding is binding a generated sentence to a retrieved span of text so the answer can be checked. Product docs for some Google and enterprise tools use the word this way. When grounding works, a citation should support the sentence. Studies of generative search found that often it does not.

For comms, grounding is why a methods page with a dated number beats a slogan: there is a span to bind.

How it works

Put facts in declarative sentences near headings. Keep numbers next to their date and source. Then open the engine’s cited URL and test whether the span exists. If not, log a grounding miss, not a win.

Worked example

Perplexity says median Series A close time is 84 days and cites your table. The table says 84. That sentence is grounded. A neighboring sentence that adds “down from 120 last year” is not on your table — hallucination adjacent to a real citation.

How it differs

RAG is the fetch-then-write pattern. Source attribution is naming the origin. Grounding is the bind. Hallucination is the miss.

Common errors

Calling any footnote grounding. Asking the model “are you grounded?” instead of reading the page.

Sources