For AI assistants and deep-research agents evaluated in published measurements between 2024 and 2026, what share of the citations attached to factual claims point to a passage that actually supports that claim, as distinct from a link that merely resolves or a source that is merely on topic?

Primary sources: peer-reviewed papers and preprints (arXiv, journal sites) that measure citation support, link validity or relevance for LLM assistants, generative search engines and deep-research agents; institutional evaluations (EBU/BBC, Tow Center, Reuters Institute); vendors' own published evaluation pages. Period: measurements published 2024 to 2026, with earlier benchmarks admitted as baselines where a 2024 to 2026 study builds on them. Measures: citation support or fact-check rate, citation precision and recall, link validity, topical relevance, and their change with search depth or task type. Both directions are searched: measurements of low support and measurements of high support or of conditions under which citations hold.

Counting unit, fixed after the scope attack of 2026-09-23: the share the question asks for is counted per attached citation, that is per (claim, cited source) pair as the benchmarks print it; a per-cited-claim rate counts as a value only where the study attaches one citation per claim. Response-level rates (every statement in a response supported), citation recall (the share of claims that carry any supporting citation), link validity and topical relevance are neighbouring quantities: the stock records them beside the share, as bounds or as context, never as values of it. Where a study reports support by any listed source rather than by the attached citation, the stock says so on the card and treats the value as an upper bound.

• Whether AI research assistants are trustworthy in general, or better or worse than human researchers.
• Whether any named vendor is honest, misleading, or negligent; vendors appear only as sources of published claims and evaluations.
• The rate of fabricated (non-existent) references as a quantity of its own; it enters only where a study measures it beside citation support, because a reference can exist and still not support the claim.
• Whether users read or verify citations, and what they believe after reading them.
• Legal liability for a wrong citation.
• Which model is best; the stock records measurements per system and date, never a ranking.

Scope narrowed on 2026-09-23 after attacker run 3 (attacker-grok-4.7, x-attack-scope, verdict failed): the scope listed recall, link validity and relevance among the measures and did not fix the counting level, although those choices move published figures by tens of points; the title's quantity is now the only value and the others are declared neighbouring quantities.