A forensic review of KPMG's October 2025 "Total Experience" report by AI detection firm GPTZero found 5 of 45 citations resolved to the cited source, in a pattern GPTZero labels "vibe citing."
KPMG told enterprises how to deploy agentic AI. Its own report, audited citation by citation, did not clear a much lower bar: only 5 of 45 references pointed to the source they claimed to cite. The rest were mangled, invented, or vague enough to be unverifiable, according to a forensic review by AI-detection firm GPTZero, as reported by The Register's Carly Page on 12 June 2026.
The document under review is KPMG's "Total Experience: Redefining Excellence in the Age of Agentic AI," published in October 2025. It is the sort of artifact a corporate board pays six figures for: a roadmap for deploying autonomous AI agents inside large organizations, packaged with case studies from established brands. On the cover, the firm is selling institutional credibility. Inside, the footnoting does not survive a click-through.
GPTZero reviewed all 45 citations in the report. Five correctly resolved to the cited source. The remaining 40 ran a spectrum from minor mangling to outright fabrication: invented titles, stitched-together fragments of real papers, and references that looked plausible until a reader tried to retrieve them. GPTZero called the pattern "vibe citing," the citation analog of vibe coding, where a generative model produces an output that reads as competent until someone checks the seams.
The cleanest demonstration is on page 42. The report attributes a customer-experience vignette about Emirates' virtual assistant "Sara" to a specific industry case study. The passage has the texture of a real case write-up. The cited source, when located, does not contain the quoted material. The example is a composite stitched from fragments of other Emirates coverage, with a reference dressed up to look authoritative. KPMG's report, in other words, exhibits the exact failure mode it warns clients about.
The same pattern shows up across the report's other named case studies, including the UBS, SBB, and TfL examples cited alongside Emirates. Each is plausible on the page. None, per GPTZero's audit, holds up under retrieval. That is the technical story: a consulting firm advising the world's largest organizations on AI adoption produced a flagship document that cannot be trusted at the level of "does this footnote go where it says it goes."
The Register frames the KPMG report as the next entry in a pattern that includes a Deloitte precedent: a refund to the Australian government after AI-generated content slipped into a taxpayer-funded report. The Deloitte episode is not independently re-verified here, but the broader pattern it gestures at, of firms advising on AI governance while their own quality control has not caught up, is the throughline.
GPTZero is not a neutral arbiter. It sells AI-detection tooling and has a commercial interest in keeping the question of machine-generated content salient. Its audit is reproducible, however: the citations are listed, the targets are checkable, and the 5/45 finding is the kind of claim a reader can falsify in an afternoon. Treat the audit as a labeled, named-source result, not a verdict from a disinterested party. The number is the story. The firm behind the number is the firm that did the counting.
The fix is unglamorous. Citation provenance, retrieval-augmented verification, and a human reviewer who actually clicks each link before sign-off are the table stakes. None of those steps require new research. They require that consulting firms treat their own publications with the same skepticism they recommend to clients. KPMG has not, on the public record available, commented on GPTZero's findings.
The watch item is whether the report gets quietly revised, formally retracted, or left in place as a downloadable artifact on a client portal. Each of those responses is a different signal about how seriously the firm treats its own evidentiary standards. For now, the document is still circulating as a reference for agentic-AI adoption, footnoted in the format of authority, accurate at the level of 5 out of 45.