Glossary
What is Grounding?
Grounding is when an AI model bases its answer on retrieved, real sources (web pages, documents) rather than only its training memory. Grounded answers cite those sources — so being a page the model retrieves and grounds on is how you earn citations in tools like Perplexity and AI Overviews.
Why grounding matters
A model answering from memory can be out of date or simply wrong, and it has nothing to cite. Grounding ties the answer to documents fetched at question time, which keeps it closer to current facts and lets the engine show its sources. For a brand, this is the route to citations: you need to be a page the engine retrieves and trusts enough to build on.
How grounding works
A retrieval step, usually a web search or an index lookup, finds candidate pages. Passages from those pages are placed in the model's prompt, the model is told to answer from them, and the citations link statements back to the passages used. Grounded does not mean correct: a model can misread a passage or generalize too far, and a cited page does not always support the claim beside it.
Because grounding depends on which passages get retrieved, clear, self-contained passages with the key fact stated plainly have the best chance of being used as they were written.
A hypothetical example
For example, someone asks about an online shop's return policy. A model answering from memory recalls an older version of it. A grounded engine fetches the shop's current policy page and cites it. If that page is blocked to crawlers or only exists as an image, the engine grounds on a third-party article instead, which may describe last year's rules.
Common mistakes
- Assuming that being in training data is enough. Many answers are grounded on live retrieval instead.
- Hiding key facts in images, scripts or gated content that retrievers cannot read.
- Publishing contradictory versions of the same fact across pages, so the engine may pick the outdated one.
- Treating a citation as proof of accuracy. Check what the answer actually claims about you.
Common questions
Does grounding prevent hallucinations?
It reduces them but does not eliminate them. A grounded answer can still misinterpret a source, blend sources together or cite a page that does not support the sentence.
How can I tell whether an answer is grounded?
Look for visible citations or a source list. Many assistants only search the web for some questions, so the same assistant can give a grounded answer to one question and a memory-based one to another.
See if AI assistants recommend your brand
Run a free scan to see which AI engines mention your brand and who they cite instead of you.
Scan my site free