Grounding
Grounding is the practice of tying an AI system's answer to specific retrieved sources so its claims can be traced and verified.
A grounded answer is built from documents the system fetched; an ungrounded one is produced from model memory. Grounded answers are more current, more verifiable, and usually carry citations.
For a business, grounding is the mechanism by which its own website can influence what an engine says about it.
Why it matters
Ungrounded answers about your business reflect whatever the model absorbed, which may be years old or simply wrong. Being retrievable is how you get grounded answers instead.
An example
Two assistants are asked for a restaurant's opening hours. One answers from memory and gives last year's hours. The other searches, retrieves the restaurant's own page, and gives this year's hours with a link to where it found them. The second answer is grounded; the first is not, and the difference is whether the restaurant's page could be retrieved and read.
Common mistakes
- Assuming a confident answer is a grounded one. Fluency says nothing about sources.
- Hiding key facts in images or scripts, so there is nothing for a grounded answer to draw on.
Questions
- How can I tell whether an AI answer is grounded?
- Look for citations or source links attached to specific claims. An answer that shows where each fact came from was grounded in retrieved documents; one with no sources may have come from model memory.
- How do I make answers about my business grounded?
- Make the facts that matter — services, prices, locations, hours — available in plain server-rendered HTML on pages that AI search crawlers are allowed to fetch.