Why You Need to Audit the Citation Checks Behind AI Quality Scores

An AI-generated response can read smoothly, address the prompt directly, and still lean on a citation that a domain expert would immediately disqualify. Evaluating whether an answer sounds helpful is fundamentally different from verifying the integrity of its sources.

We can see this distinction clearly in the prelaunch evaluation of Evrópuvefur, a public-information platform run by the University of Iceland. Researchers fed the same 287 questions into two configurations: one restricted to a curated document collection, and another connected to open-web search. Subject-matter experts then audited a subset of the generated responses and their accompanying citations.

Out of 187 audited web-search responses, 65 included at least one citation flagged as either irrelevant or untrustworthy—roughly 35%. While that is a notable takeaway for those specific web responses, it should not be extrapolated as a universal failure rate for web-grounded AI.

Meanwhile, the underlying study notes that only 16 of the 262 audited curated-retrieval responses received source flags, or about 6%. However, reviewers only checked curated citations for whether they were outdated, whereas they scrutinized open-web citations for relevance and trustworthiness. Because the auditors were looking for different flaws, you cannot use these figures to claim that curated retrieval isnearly six times safer or more reliable.

Compare the criteria before comparing the percentages

Two percentages can only be weighed against each other if they measure the exact same phenomenon. Flagging a citation for being out of date is not the same event as flagging it for being irrelevant or untrustworthy.

Think of it like auditing one group of web pages for broken hyperlinks and a second group for pricing errors. Even if both groups contain the same number of pages, a lower failure rate in the first group does not prove those pages are of higher overall quality. The audits simply targeted different failure modes.

That exact dynamic is at play here. The divergence in flagged issue types was partly built into the instructions given to the evaluators. It would be a mistake to read the results as if experts ran identical audits and naturally uncovered different signature weaknesses for each retrieval method.

To run a true head-to-head trust comparison, you have to apply identical, relevant evaluation criteria and review protocols across both datasets. Even so, the study offers a valuable signal: domain experts found questionable citations in more than a third of the reviewed web-search answers. That observation holds plenty of weight on its own without misinterpreting the curated figures as a direct safety benchmark.

Track answer quality and citation quality separately

To conduct the audit, five subject-matter experts graded responses across a seven-criterion rubric while logging separate, citation-level flags. In total, they produced 551 individual evaluations covering 449 generated answers.

Keeping these two tracks separate is critical because a strong composite answer score can easily mask a broken citation. A model might synthesize a coherent, well-structured explanation while citing a dubious blog, or it might link to a prestigious institution whose actual webpage says nothing about the claim being made.

When designing your own evaluation pipeline, always log the specific claim, the linked URL, and the exact failure mode together. Slapping a generic "bad citation" label on an error gives engineers nothing actionable to fix. Replacing a stale policy document requires a different fix than re-grounding a claim that lacks evidentiary support, which in turn differs from filtering out low-authority publishers.

It is also worth remembering that the expert flags in this paper represent subjective professional judgments rather than absolute verdicts on every URL. Even expert human review requires strict rubric definitions and a mechanism for resolving inter-annotator disagreement. Simply labeling a workflow "expert review" does not automatically make its output categories uniform.

Prompting for preferred domains is not a hard allowlist

The research team also experimented with prompt engineering to nudge the open-web search toward preferred domains. Across their tested prompts, the share of citations pointing to those target domains rose from 12% to 21%.

While that confirms the prompt shifted the citation distribution, it tells us nothing about the quality of the remaining 79% of citations, nor does it prove that the preferred-domain links actually substantiated the model's claims. Most importantly, it does not show that the prompt instruction enforced a strict retrieval boundary.

If your compliance standards demand that an assistant pull exclusively from vetted sources, look closely at where that constraint is enforced. Asking a model to prefer certain websites via a system prompt is architecturally worlds apart from restricting a retrieval engine to an indexed, approved corpus. And even when you lock retrieval down to a strict allowlist, you still have to keep the corpus updated and verify that the retrieved chunks actually answer the user's query.

Lock in your audit rubric before declaring a winner

When building a user-facing assistant, begin by mapping out the exact queries it needs to answer and the specific citation failures that would break user trust. When benchmarking different retrieval architectures, subject every pipeline to the exact same citation rubric. Track query coverage, factual grounding, citation relevance, and content freshness as distinct metrics rather than blending them together.

Make sure you also log abstentions. A closed-corpus system that refuses to answer out-of-scope questions and an open-web system that attempts an answer anyway are making fundamentally different product trade-offs. Relying on a single blended quality score will obscure those operational realities.

The easiest sanity check you can run today requires no complex tooling: pull a random sample of high-scoring responses and actually click through their citations. Verify whether the text on the destination page genuinely backs up the specific claim the model attached to it. That simple exercise will quickly reveal whether your top-line quality metric actually captures evidentiary rigor.

Context note: The referenced paper evaluated a prelaunch, government-backed Icelandic platform operating within a single language and policy domain. Because auditors applied different flagging rubrics to open-web versus curated sources, the headline 35% and 6% failure rates cannot be read as an apples-to-apples comparison of system trustworthiness.