What Is Citation Integrity?

What does the definition mean in practice?

Three parts of that definition carry the weight.

Relied upon means an AI system can lean on the information without importing an unsupported claim into its answer. Correctly attributed means the information can be traced back to the source that actually supports it. Published information means the material you have put into the world — pages, documents, disclosures, data — not your brand, your reputation, or your marketing claims about yourself.

How is a citation different from citation integrity?

A citation is a pointer. Citation integrity is whether the pointer is trustworthy in both directions: the source is dependable, and it genuinely supports the statement attached to it.

Research on attributed generation splits the question in two.

A system can cite constantly and still be wrong, or be right and cite nothing. A 2023 evaluation of generative search engines found that a large share of generated statements were not fully supported by the sources cited beside them.

Citation Integrity is the publisher-side counterpart to that finding. If AI systems attach citations imperfectly, the material they attach them to needs to withstand the check.

Why does it matter to an organisation?

When an AI assistant answers a question about your market, your product or your category, it makes a selection. Sources it can retrieve, parse, attribute and defend are usable. Sources that are unreachable, unattributable, unsupported or of uncertain origin are risky for the system to lean on.

NIST's Generative AI Profile (AI 600-1), published in July 2024, describes confabulation and information-integrity risks as characteristics that AI actors are expected to manage. Managing them means being selective. Selectivity is why the dependability of your published evidence has become a commercial question, not just an editorial one.

What does Citation Integrity assess?

CiteAbility assesses Citation Integrity across eight public dimensions: Source Authority, Entity Authority, Organizational Probity, Evidence & Citations, First-Hand Experience, Content Quality, Technical Accessibility, and Integrity Analysis. The complete guide explains how they relate; the framework overview sets out the dimensions in product terms.

Two points about scope belong in the definition itself.

First, identity is part of it. Evidence is assessed alongside the identifiable entity that published it, because "who is saying this and why would they know" is inseparable from "can this be relied upon".

Second, standing is part of it. Organizational Probity considers material verified and dated information about conduct, transparency or standing — an official finding, an adjudicated sanction, a company's own disclosure. Where no such information exists, the result is neutral. Absence is never treated as an adverse result, and an unresolved allegation is never treated as an established fact.

What is Citation Integrity not?

• Not a ranking metric. It says nothing about position in a results page.
• Not a mention count. Counting appearances measures what happened; Citation Integrity assesses what your evidence merits.
• Not fact-checking. Fact-checking adjudicates a claim. Citation Integrity assesses whether a body of published information, and its publisher, can be relied upon.
• Not a guarantee. No assessment can commit an AI system to citing you. Systems differ, change without notice, and do not publish their selection criteria.
• Not a technical audit alone. Machine-readability is necessary and insufficient. Perfectly structured pages with no verifiable evidence behind them have low Citation Integrity.

How do tested evaluation and observed citation differ?

These are separate, and conflating them is a frequent error in this field.

A tested AI evaluation asks AI systems questions under controlled conditions and records how they assess the available evidence. It indicates citation readiness at the time of the test. An observed citation is a recorded instance of an AI system citing a source in live use.

A model stating that it would cite an organisation is a test result. It is not proof that it has cited it, and not a forecast that it will.

How is it assessed in practice?

A Citation Integrity Audit begins with retrieval: what an AI system can actually reach on a site, and what happens when it cannot. From there it examines whether claims are substantiated, whether sources support what they are attached to, whether the publishing entity is identifiable and appropriate, and whether the material can be parsed and attributed. The Citation Integrity Score expresses assessed position across the dimensions over time, and Citation Integrity Monitor re-checks pages so regressions surface. Research and testing transparency sets out how we test.

What is deliberately absent is a published formula. The dimensions and their questions are public. The underlying signals, weightings and thresholds are not — a published formula becomes an optimisation target and stops measuring the property it was built to measure.

What do people most often get wrong?

"If my content is good, my Citation Integrity is good." Quality is one dimension of eight. Good content that cannot be retrieved, or whose claims cannot be substantiated, does not clear the bar.

"This is just structured data." Structured data helps machines parse and attribute. It does not make an unsupported claim supportable.

"We were mentioned by an AI assistant, so we're fine." A single mention is an observation, not an assessment — and not a durable one.

"A neutral probity result means we were investigated and cleared." No. Where there is no material verified information, the result is simply neutral.

In one line

Citation Integrity is not about persuading AI systems to use you. It is about publishing evidence that holds up when they check — and being honest that the checking is theirs, not yours.

Definition current as at September 2026.