35% of web-search LLM answers cited untrustworthy sources

The University of Iceland's Evrópuvefur service ran a controlled expert evaluation before its public launch, and the numbers are unambiguous. Reported on arXiv, the study found that 35% of AI-generated answers produced via open web search cited at least one source flagged as untrustworthy or irrelevant. Curated retrieval against a governed local corpus produced flagged citations far less often. Five domain experts scored 551 evaluations across 449 answers on a seven-criterion quality rubric. The verdict is not subtle: retrieval architecture determines source trust, and source trust determines whether an institution's AI service is fit to publish.

This matters because the study is not a theoretical exercise. Evrópuvefur was built to answer Icelandic citizens' questions about the EU ahead of a referendum on 29 August 2026 on whether to resume accession talks. The stakes of a wrong, or badly sourced, answer are measurable in civic terms. That context makes this one of the cleaner real-world tests of a question most enterprise AI deployments have not yet seriously asked: when an LLM cites something, does it cite something worth citing?

The mechanism behind the failure rate

Open web search hands the retrieval decision to the model. The model optimises for apparent relevance, not for editorial quality, institutional credibility, or accuracy. A government press release, a partisan commentary site, and a peer-reviewed legal summary can all look similar to an embedding-based retrieval step. The curated corpus, by contrast, encodes a prior human judgment: these sources have been assessed and included deliberately. The expert flagging rate collapses accordingly.

The 35% figure is not a marginal edge case. It means that more than one in three answers produced via web search contained a citation a domain expert would not trust. In a consumer product, that is a reputational problem. In a public information service advising citizens on a constitutional decision, it is a governance failure waiting to be published.

The trade-off named in the study title is real, though. Curated corpora have coverage limits. A local knowledge base built around EU accession law will not surface a breaking development from last week. Web search adds recency and breadth; it also adds noise, misinformation, and sources with no editorial accountability. The study treats this as a genuine trade-off, not a solved problem. That honesty is the most useful thing about it.

What this means for brands publishing in AI-mediated channels

For senior communicators at multilateral institutions, financial services groups, and major industrial bodies, the implication cuts in two directions.

First, if your organisation operates or is building an AI information service, retrieval architecture is a content governance decision, not an infrastructure one. Leaving source selection to live web search is the equivalent of publishing without editorial review. The University of Iceland study gives you a defensible benchmark: 35% flagged-source rate for open web retrieval versus a materially lower rate for curated RAG. That comparison belongs in a procurement conversation, a risk register, or a board-level briefing on AI deployment.