What "cited answers" actually means when accuracy matters

Most AI products that touch documents will tell you they offer citations. The implementation varies enormously. In any setting where accuracy matters, the variance matters, because the standard for "verifiable" is harder than the standard most products are built to meet.

The three levels of citation

It’s worth being precise about what people mean when they say a model "cites." There are three common implementations, and they are not interchangeable.

Level 1: Document reference

The model tells you which document an answer came from. This is the floor. It is also the ceiling for many tools, and it isn’t enough. If the document is 400 pages, "the answer is in this document" is barely better than no citation at all — the verification step still costs the user most of the time the model was supposed to save.

Level 2: Page or section reference

The model tells you the page or named section. This is the middle. It is what most "decent" implementations achieve. It compresses verification significantly, but it relies on the user being able to read fast enough through a section to confirm the specific claim. For complex domains where claims are layered and conditional, that gap is still costly.

Level 3: Claim-level deep linking

The model points to the specific paragraph, table cell or figure that supports the specific claim it made. This is the standard Dash is built to. The user is one click from the exact piece of evidence, and verification takes seconds rather than minutes. At scale, the difference between Level 2 and Level 3 is the difference between an AI tool that gets used and one that gets quietly abandoned after the second time it’s wrong.

Why the level matters more than the answer

A useful way to think about a document-intelligence system: the answer is only as defensible as the citation behind it. A correct answer with a vague citation is worth less to a team than a correct answer with a paragraph-level citation, because the verification overhead is what determines whether the team will rely on the system in time-pressured work.

A correct answer with a vague citation is worth less than the same answer with a precise one. The work the team has to do to trust the system is the cost that matters.

The hidden second feature: legible gaps

The other half of citation is knowing when there is nothing to cite. A system that always produces an answer — even when the evidence isn’t there — trains the team to be suspicious of all of its answers. A system that says "the knowledge base doesn’t support a confident answer to this question" earns trust quickly, because the team learns it can rely on the ones the system did answer.

Dash refuses to answer without source support. That refusal is a feature, not a limitation. In serious knowledge work, knowing where the corpus is silent is as valuable as knowing where it speaks.