Answer groundedness: meaning, examples and practical use
Answer groundedness describes how well the supplied source material supports an answer’s claims. Plausible wording or an attached citation is insufficient: the source must support the particular statement and its conditions. Groundedness alone does not establish that a source is current or true.
Also known as: groundedness · source support
At a glance
- A citation and evidence supporting a claim are different things.
- A faithful answer can still be outdated when its source is old.
- Check individual claims, their evidence and the conditions attached.
On this page

What does groundedness assess?
The question is whether the answer follows from the material made available to it. Google’s evaluation documentation likewise defines groundedness through the relationship between a response and its supplied context. Particular scoring schemes vary by system. Evaluation criteria
Separate faithfulness to a source, factual correctness and completeness. Every statement in a short answer may be supported while the response leaves an important part of the question unanswered.
Example: what counts as free delivery?
Imagine an online shop whose terms say “Free domestic delivery on orders over €80.” The answer “All orders ship free” concerns the same subject but is unsupported: it drops the spending threshold and geographical restriction.
An accurate account retains both conditions. If a customer asks about an overseas address, the domestic rule alone cannot answer their question. The system needs further evidence or an explicit acknowledgement that the information is missing.
How to check an answer
Separate the checkable claims: prices, deadlines, eligibility and comparisons. Find the specific passage supporting each. Look for omitted exceptions, and check whether a possibility has become a promise.
An AI citation is a useful starting point, but linking a related page is not itself proof. Poor chunking can also separate a rule from the exception needed to interpret it.
For a numerical statement, check the units and period as well as the number. “Twenty bookings in September” does not support a claim of twenty bookings every month. These small changes can alter the commercial meaning substantially.
What does a strong score not establish?
The source may be mistaken, or the rule may have changed. Faithfully summarising an old price list does not necessarily answer a question about today’s price. Check source freshness separately.
Groundedness is one perspective on AI hallucinations, not an all-purpose quality measure. A useful business answer must also be understandable, current and responsive to the actual question.