Why do AI deep-research reports cite URLs that don’t exist?

A citation is a promise: this page exists, and it says what I said it says. Most deep-research reports break that promise.

A University of Pennsylvania study, funded through DARPA’s SciFy program, measured citation URL validity across commercial LLMs and deep-research agents (Rao, Wong, and Callison-Burch, arXiv:2604.03173). On their DRBench set, 3 to 13 percent of cited URLs have no record in the Wayback Machine — they likely never existed. Five to 18 percent do not resolve at all.

Deep-research agents cite more, and they hallucinate more. Pooled hallucination is 10.7 percent against 4.8 percent for search-augmented models. Gemini Deep Research sits at the top of that table: 13.3 percent hallucinated, 18.5 percent non-resolving. More citations is not more proof.

The failure is architectural. A one-shot report can emit a URL from retrieval, from memory, or from a blend of the two. Some models fabricate every dead URL they emit. Others retrieve a real page that has since gone dark. The memo does not tell you which.

If you run a newsroom, a trade association, or a research desk, that is the product you are being sold: a fluent brief with a bibliography you still have to check by hand.

The rule is the same in business and in print: source it, or don’t say it. A fluent answer is not proof.

Aculeus is the intelligence layer under Starglass. Deep Research does not start until you approve an editable Case Plan. The Living Case keeps cited source links, the conversation, and the open questions. Nothing auto-starts or publishes automatically. A reviewed Living Case can be shared as a public, read-only record. Judge the method, not a showcase.

If a URL does not exist, it is not a source. Treat it that way.

Source. Delip Rao, Eric Wong, and Chris Callison-Burch, “Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents,” arXiv:2604.03173, 2026. https://arxiv.org/abs/2604.03173