What is an atomic fact?

An atomic fact is the smallest unit of an answer that still makes a claim. It names one thing and says one thing about it. "The Pro plan costs $99 a month" is atomic. "The company, founded in 2015, sells a $99 Pro plan" is not, because it bundles two claims that could be true or false independently.

The term comes from factuality research. Min et al. defined it in FActScore (EMNLP 2023), which scores a response as the share of its atomic facts that a knowledge source supports. Related work uses the names atomic claims or statements for the same idea.

Decomposition runs in three steps.

1. Split. A model or a reviewer breaks the answer into standalone claims, each rewritten so it makes sense without the surrounding text.
2. Check. Each claim is compared with a trusted source and labelled supported or not supported.
3. Count. The labels become a rate.

Say an AI answer about a hypothetical software company reads: "Example Co, founded in 2015 in Berlin, offers a Pro plan at $99 a month." Decomposition yields four atomic facts: founded in 2015; based in Berlin; has a Pro plan; the Pro plan costs $99 a month. If the company's own records show a 2017 founding, three of four facts are supported.

The main metrics all divide by a count of claims. FActScore divides supported facts by all atomic facts. RAGAS faithfulness divides supported claims by total claims. HaluEval 2.0 reports two rates: hallucinated statements over all statements, and responses with at least one hallucination over all responses. For the full mechanics, see our explainer on how FActScore measures factual precision.

Being named in an AI answer is half the picture. The other half is whether the answer gets your facts right. AI platforms do get specific facts wrong about specific brands, and a wrong price or a wrong acquisition story costs more than silence.

Atomic facts make that measurable. Instead of "ChatGPT seems off about us," you get a list: this answer stated the wrong founding year, that one quoted a retired plan. Each line points to a fix. Often the cause sits on your own site, where the pricing page and the footer disagree. Our guide to consistent brand facts covers finding and fixing those contradictions.

What stays outside your control: the model's training data, and whether a given answer used live retrieval at all. What is in your control: stating each key fact once, clearly, on a page crawlers can fetch. That gives an engine a sentence it can quote. It does not guarantee the engine will use it.

"Not supported" means false

It doesn't. FActScore's not-supported label covers claims the source contradicts or lacks evidence for, and does not separate the two. An AI stating your headcount when you never published one is an unverifiable claim, not a proven error. Read more on whether every flagged claim is actually false.

Every atomic fact matters equally

In the benchmarks, yes. FActScore states every atomic fact has equal weight. Weighting a wrong price above a wrong headcount is a product decision, not research practice.

A high precision score means a good answer

Not always. The SAFE paper (Wei et al., NeurIPS 2024) added a recall term because precision alone does not penalize systems that abstain often or state fewer facts.

A citation supports every fact near it

A cited source can sit beside a claim it never makes. See why AI citations sometimes don't support the claim.

• FActScore: the metric that introduced atomic-fact scoring.
• Unverifiable claim: a fact no trusted source confirms or contradicts.
• Hallucination: a stated fact that is fabricated. Our guide on AI hallucinations about brands covers causes and responses.
• Faithfulness: the share of an answer's claims supported by the documents it was given.

What counts as a trusted source for checking brand facts? For brand claims, the trusted source is usually the brand's own verified record: the current pricing page, the about page, official filings. Research metrics like FActScore used a fixed knowledge source such as Wikipedia. Whatever you choose, keep it consistent, because an atomic fact can only be judged against something. If your own site states a fact two ways, no checker can say which version is right.

Can I decompose answers into atomic facts by hand? Yes, and for a handful of answers it is often the clearest method. Copy each AI answer, split it into one-claim sentences, then mark each against your records. It gets slow quickly, which is why research uses a model to split and check. Automated splitting can merge or drop claims, so spot-check a sample by hand either way.

Is a missing fact the same as a wrong fact? No. An answer that leaves out your founding year has no atomic fact about it to score. An answer that states the wrong year has a not-supported fact. Precision metrics only judge what was said, which is why SAFE added recall against a target number of facts. Omissions and errors call for different fixes.

How do I see which facts AI gets wrong about my brand? Ask the engines the questions your buyers ask, then check each stated price, date and name against your records. The free checker on this site runs up to three buyer questions and scores the answers, as a one-time snapshot rather than ongoing tracking. Treat any single run as a sample, since answers vary by query and by whether live search was used.