Groundedness (LLM answers)

Groundedness is whether the factual claims in a generated answer trace back to an identified source a reader can open and check, and as a metric it is the share of those claims that do. It scores an answer against whatever sources the system named or was handed, which makes it wider than faithfulness, a number computed only against the passages one retrieval step returned for that question.

The judgement underneath is binary at the level of a single claim, since a source either supports it or does not. Rashkin and co-authors formalized that judgement as Attributable to Identified Sources, in Computational Linguistics in 2023, and their annotation runs in two stages: decide whether a sentence is interpretable as a standalone claim about the world, then ask whether “According to the source, x” holds. Aggregating binary verdicts produces a proportion, so groundedness inherits what a proportion owes, starting with a declared denominator and an interval. The screening stage is where that denominator moves. Sentences carrying no external claim and sentences too vague to judge both drop out of the count, so a system that hedges harder posts a better rate on the fewer claims it leaves behind.

On the generation side of a pipeline, groundedness sits beside whether the answer addressed the question that was asked and downstream of what share of the relevant material the retriever actually surfaced. Where it sits relative to the other five scores, and which question each of them answers alone, is worked out one score at a time. A claim can only rest on evidence the run actually obtained. Quote it with the source set it was allowed to count and the claims it managed to judge, or the figure describes an answer set nobody can reconstruct.

How to calculate groundedness

Split the answer into claims, one assertion each, then drop the ones that assert nothing about the external world: a hedge, a restatement of the question, an offer to help further. For each claim that survives, name the source it is supposed to rest on: a passage in a retrieval pipeline, a tool return or a database row in an agent run, the page behind the citation marker in a search-backed answer. Judge whether that source supports the claim, then divide supported claims by judged claims.

Deciding which sources count has no counterpart in a faithfulness score. Someone chooses the admissible source set by hand. A faithfulness check takes the retrieved context as its reference and asks nothing about where that context came from, while a groundedness check has to declare that set before it can score anything. The indexed corpus counts, a tool return usually counts, and the user’s own earlier message rarely should, though it sits in the same window and entails plenty. Write the set down before the first run, because quietly widening it raises the score without changing the system.

Report the judged-claim denominator and the screened-out count next to the rate, and put the interval on answers rather than on claims. Claims inside one answer share a question, a context window and a single generation, so scoring each as an independent trial gives an interval narrower than the evidence supports. Collapse each answer to a single verdict and the pass-rate interval calculator returns the Wilson or Clopper-Pearson bound on that count. Where a model does the judging, measure the judge first, with agreement against human labels and then a correction for its sensitivity and specificity. Publish the share of requests that got an answer at all beside the rate too, because refusals and empty generations leave the denominator, and grading only the questions a system agreed to answer flatters the one that declines the hard ones. All of it assumes the reference is settled, and the term most often swapped in has that reference fixed for it in advance.

Groundedness vs faithfulness

Faithfulness in RAG evaluation is the proportion of an answer’s claims that can be inferred from the retrieved context, computed as supported statements over extracted statements against the passages the pipeline actually put in front of the model. Groundedness asks the same shape of question with the reference left open, which is why it still means something outside a retrieval pipeline.

Where the two disagree, one case leaves faithfulness with nothing to score at all. An agent answering from tool returns, or a model that names a checkable public source while retrieving nothing, has a groundedness result and no faithfulness result. A harness pointed at “the retrieved context” finds an empty reference there and reports silence rather than a failure. That case is ordinary in an agentic architecture where evidence arrives across several hops instead of one retrieval call.

Running the other way, an answer can be perfectly faithful to a context that should never have counted as evidence. Every claim follows from a chunk, the chunk turns out to be a pasted customer email or a policy page superseded two quarters ago, and faithfulness scores 1.0 because entailment held. Groundedness fails that same answer once the admissible source set excludes what it leaned on. The two also move on different schedules: faithfulness shifts when the generator or the prompt changes, groundedness when the corpus, the tools or the citation layer change.

Groundedness vs citation coverage

Citation coverage counts markers, meaning the share of sentences in an answer that carry a citation of any kind. Groundedness counts support, a different question about the same sentences, because a marker asserts that a source exists rather than that the source says what the sentence says.

The gap between the two has been measured directly.

Liu, Zhang and Liang put four generative search engines in front of human annotators for Findings of EMNLP 2023. On average, only 51.5% of generated sentences were fully supported by their citations, with 74.5% of citations fully supporting the sentence they were attached to. Those are point estimates over four systems as they behaved in 2023, published without intervals, and each product has shipped many versions since. Read them as evidence that the gap between a marker and its support is real, never as a current rate for anything.

A check that only confirms a marker is present rewards appending markers, which is how a fabricated claim with a tidy reference survives review in the taxonomy of RAG pipeline failures. Score each claim against the cited span rather than against the document that span was taken from.

Grounding as a feature and groundedness as a measurement

The shorter form of the word names a product capability. Google’s Gemini API documentation calls its search connection Grounding with Google Search and describes it as connecting the model to real-time web content. That is a claim about what sources a model can reach rather than about how much of an answer rested on them. Switching the feature on changes the admissible source set, and the share of any given answer that actually traces to it is still yours to measure.

Cognitive science and multimodal machine learning reached the word earlier. In that older sense grounding is the link between a symbol and the thing in the world it refers to, which reads as an ancestor of this sense rather than a rival to it.

Teams still want one groundedness number for a whole product, and there isn’t a defensible way to hand them one. The rate depends on which sources were admissible and on the mix of questions asked, so a support assistant reading tight policy documents and a research assistant reading the open web cannot share a scale. Report it per source class and per question type, each segment carrying its judged-claim denominator and its interval, and the number starts describing something checkable.

Groundedness is the property, and faithfulness is the instrument most teams reach for to measure it inside a retrieval pipeline, so a report quoting either one should say which sources it was allowed to count. The fuller account of scoring retrieval and generation separately instead of reading a single blended number lives in evaluating a RAG pipeline stage by stage.