Six citations, one measurement
A citation dashboard reports where an assertion appears. It does not report where the assertion came from. A survey result repeated by five news organizations looks, on a dashboard, like five points of corroboration. It may in fact be one measurement, quoted five times.
AATL ran a 100-case experiment to test how often that gap actually occurs. Each case was an assertion that appeared across at least three publisher domains. Each was traced upstream to its underlying evidence.
What collapsed does not mean
Collapsed does not mean false, unreliable, or poorly sourced. It means the apparent multiplicity of publishers did not survive as multiplicity of evidence. A dataset, a survey, a company disclosure, or a single measurement sat underneath publisher-domain counts that looked, from the outside, like independent confirmation.
Forty-five of the 100 cases retained meaningful independent verification: court orders, regulatory filings, executed agreements, and events independently observable by multiple parties. One case was contested, where genuinely independent evidence supported conflicting conclusions.
Same source ecosystem, different topology
The clearest finding was that evidence independence belongs to the assertion, not the publisher. Two claims from the same article, about the same company, citing the same set of domains, produced different classifications. A regulatory approval retained, because the record itself is direct evidence. A trial result reported alongside it collapsed, because every publication traced back to the same single experiment.
The same pattern held across a national unemployment figure repeated by a dozen outlets, an incident-scale estimate attributed back to one company, and a survey result quoted by five separate newsrooms. In each case, publication independence was real. Measurement independence was not. A cybersecurity case ran the opposite direction: reporting drew on multiple distinct sets of investigators and technical evidence, and that assertion retained. The protocol does not treat everything as derivative by default — genuine multi-path evidence survives it.
What the 100 cases do not establish
The cohort was constructed to include a range of assertion types, not sampled at random from the web. Fifty-four percent describes what happened inside these 100 cases, not a rate across the internet or across AI citations generally. The study did not measure whether AI systems distinguish independent evidence from repeated evidence, whether expensive AEO tools outperform cheaper ones, or whether any AEO product improves citation outcomes. None of those were tested.
Adjudication was applied uniformly after all 100 cases were collected: one classification changed on review, moving from collapsed to retained after stronger securities documentation surfaced. Ninety-nine of the original 100 classifications held. The one correction ran against the study’s own headline finding, not toward it.
What commercial tools currently document
Public materials from AEO and AI-visibility products reviewed during this research describe mentions, citations, cited URLs, source domains, sentiment, and competitive visibility. We did not identify a documented capability equivalent to the assertion-level evidence-independence adjudication used in this experiment. That is a statement about documented capability, not about what any vendor’s internal systems may or may not do.
The blind spot
A rising citation count or a growing list of source domains answers where an assertion is being repeated. It does not answer how many independent reasons exist to believe it. For anyone deciding what to trust in an AI-generated answer, or what to build a business claim on top of, those are two different measurements — and only one of them shows up on the dashboards currently being sold.