For assistants and deep-research agents measured in 2025 and 2026 the published share of citations that support their claim runs from 24.4 to 94.04 percent. Within one benchmark the printed spread between systems is 52.4 points (24.4 to 76.8) and 54.7 points (39.36 to 94.04); one deep-research product moves 31 to 32 points between benchmarks, against spreads of 29 and 12 points between deep-research products within one benchmark. Counted per response instead of per statement, the same answers of one system in a dataset of 2024 give 38.4 against 75.7 percent, which is a different quantity and not a correction. Where automatic support judges were validated against humans they disagree on 2 to 22 percent of decisions; those validations concern other judges than the ones behind the 24 to 94 span. An answer to the question is a range with its conditions, not a number.

The range derivation contributes the span of printed values and the spreads within a benchmark (76.8 minus 24.4, 94.04 minus 39.36). The same-product derivation contributes that for deep-research products the benchmark moves one product as far as products differ within a benchmark; the step does not extend that to all systems, where the within-benchmark spread is larger. The unit derivation contributes that statement-level and response-level rates cannot be set side by side. The judge derivation contributes the decision-level disagreement of the judges that were validated in its parents; the step does not transfer that band to the judges behind the span, three of which print a validation on their anchors (DeepTRACE: Pearson 0.62 on 100 tasks; DeepResearch Bench: 96 and 92 percent agreement on 100 pairs; the 14-agent benchmark: rubric judges validated at F1 0.75 in a separate study) while ResearcherBench prints none. The step is a conjunction and adds no premise beyond the four. It does not rank systems and does not say which benchmark is closest to ordinary use.

A human-rated measurement on a representative sample of ordinary user queries across current products; it would give the question a value with an interval and make this card a statement about benchmarks only. Query: any 2026 human audit of per-citation support on sampled real-world queries.

About X percent of AI citations are wrong. Any single X quoted for this is one cell of a table whose rows are benchmarks, units and raters.

Attacker run 3 (attacker-grok-4.7, x-attack-step): applied, the step over-generalised; three of the four benchmarks behind the span print a judge validation, ResearcherBench does not. Conclusion unchanged.