Glossary
Grounding
Grounding is tying an answer to a specific source you can go and check, such as a retrieved document or a database record, rather than letting the model answer from what it absorbed during training.
In plain terms
It is the difference between a colleague saying they think the returns window is thirty days and one saying it is thirty days and here is the policy page. Both may be right. Only the second can be verified in ten seconds, and only the second is safe to forward to a customer without reading the whole thing again yourself.
Why it matters
It is the practical dividing line between an AI feature you can put in front of customers and one you cannot. Without it, checking an answer means knowing the subject well enough to have written the answer yourself, which removes most of the saving. With it, checking means following a link. It also changes what a wrong answer costs: a grounded mistake can be traced to a source and fixed at the source, where an ungrounded one leaves you knowing only that the tool was wrong about something, once.
How it works
Something fetches material before the model answers. A search runs, a document store is queried, a database record is looked up, and the results are placed alongside your question. The model then answers from what is in front of it rather than from what it happens to have absorbed, which is the whole of the technique.
The instruction matters as much as the material. A model handed relevant documents and no constraint will happily blend them with everything else it knows. The line that does the work is a plain one: answer only from the supplied material, and say so where it does not cover the question. Without the second half, gaps get filled rather than reported.
Citations are the visible part and are not the same thing. A citation is a claim about where a statement came from, and it can perfectly well be attached to a sentence its source does not support. Grounding is the arrangement that fetched the material and constrained the answer; the citation is the label on the result. Products can produce the label without the arrangement.
It is a matter of degree rather than a state. Fully grounded means every claim in the answer traces to something supplied. Partly grounded, which is what most tools actually do, means the model uses your material where it applies and falls back on its own recall where it does not. Both arrive in one continuous piece of prose that gives no sign of where the boundary was.
It buys traceability rather than correctness. Where the retrieval step misses the right passage, or the source itself is out of date, the result is a confidently wrong answer with a citation attached. That is in some ways worse than a wrong answer without one, because it survives a casual check. The remedy is spot-checking a few citations rather than trusting their presence.
Four statements, three markers, and one that slipped through
Seen in the wild
Load your own documents into a notebook tool and ask something they do not cover, then see whether the answer says so or reaches past what you supplied.
NotebookLMFollow the numbered links on a research answer back to the pages they came from, and check that each page actually says the sentence it is attached to.
PerplexitySearch a company workspace and follow an answer back to the document it came from, which is the same arrangement applied to material only your colleagues can see.
Glean
Common misconceptions
People assume
A citation means the answer is grounded.
In fact
It means a source has been named. Whether that source supports the sentence it is attached to is a separate question, and one that only checking answers. Treat citations as an invitation to verify rather than as evidence that somebody already did, particularly on the claims that matter.
People assume
Grounding stops the model making things up.
In fact
It reduces it substantially and does not remove it. When the retrieval step fails to surface the right passage, the model still has to produce something, and what it produces arrives dressed in the same format as the grounded sentences around it. Retrieval failure is the case worth designing for.
People assume
We uploaded our documents, so we are grounded.
In fact
Uploading gives the system something to retrieve, which is one of three parts. The other two are instructing the model to stay within that material and to say when it does not cover the question, and having something check that it did. Without those, an assistant with your documents is an assistant with better raw material and the same habits.
Telling them apart
Grounding vs RAG
Grounding
The goal: answers tied to checkable sources. Achievable by retrieval, by a database lookup, or by simply pasting the document in yourself.
One well-established technique for reaching it, in which material is searched for and supplied automatically before the model answers.
All retrieval of this kind is aimed at grounding; not all grounding needs retrieval. A person pasting the contract into the conversation has grounded the answer without building anything.
Grounding vs Citation
Grounding
The arrangement: what was fetched, what the model was told to do with it, and what happens when the material falls short.
The label on the output: a marker pointing at where a statement is said to have come from.
One is plumbing and the other is signage. A product can install the signage without the plumbing, which is why a numbered link is a starting point for checking rather than the end of it.
Questions
- How can I tell whether an answer is actually grounded?
- Pick two claims, follow their sources, and read what the source says. Then find a sentence with no source attached and ask where it came from. Answers that blend supplied material with recall usually give themselves away in that third step, because the unsourced sentences are the ones doing the most confident work.
- Does grounding eliminate made-up answers?
- No, and expecting it to is the mistake that lets grounded systems through without review. It removes most of the fabrication that comes from having nothing to go on, and leaves the case where retrieval fetched the wrong passage or nothing useful. The answer that results looks exactly like a good one.
- What should I ask a vendor about it?
- What the assistant is allowed to answer from, what it does when the material does not cover a question, whether every claim in an answer is traceable or only some, and whether you can see which documents were retrieved for a given answer. The last one is the most revealing and the least often volunteered.
- Is grounding worth it for internal use as well?
- Usually more so, because internal questions are exactly the ones where a model's own recall is thinnest. It also gives you something a general assistant cannot: a way of finding out that the underlying document is wrong, since a traceable answer sends somebody to the source and occasionally they fix it.
Key takeaways
- Grounding ties an answer to material you can go and check, instead of to the model's own recall.
- It needs three things: relevant material fetched, an instruction to stay within it, and a stated fallback.
- A citation is a label; grounding is the arrangement behind it, and a product can produce one without the other.
- It is a matter of degree, and most tools blend supplied material with recall in one continuous piece of prose.
- It buys traceability rather than correctness, so spot-checking a few sources is still the job.
Tools that use this
- NotebookLM
Answers drawn from the sources you supply, which is the strict case.
- Perplexity
Numbered links you can follow back to the page, which is what makes checking cheap.
- Glean
The same arrangement over material only your colleagues can see.
Last checked July 2026