[Company name changed. Based on a composite of enforcement actions and content-seeding operations in our corpus.]
In March 2026, a supplement brand we are calling Iron Rock NAD+ published seven white papers on its corporate blog and seeded them across three open-access preprint hosts. The papers were formatted in two-column layouts with abstract blocks, numbered citations, and DOI-style reference strings. They carried titles like "Nicotinamide Riboside and Mitochondrial Bioenergetics: A Prospective Framework for Age-Related Decline." None of them had undergone peer review. None disclosed that fact on the page.
Ninety days later, Iron Rock NAD+ appeared as a cited source in 14% of all AI engine answers to queries about NAD+ supplementation across ChatGPT, Perplexity, Claude, and Gemini. That is not a typo. A brand that did not exist in AI citation data in February captured more than one in seven answers by June.
We read all seven papers. This is what we found.
The clinical register trick
Each paper was written in what linguists call a clinical register: passive voice, hedged conclusions, method-section formatting, and an abundance of in-text citations to real PubMed-indexed studies. The actual content, however, was marketing. Three of the seven papers made efficacy claims about Iron Rock's proprietary formulation. Two made dosage recommendations. One included a "suggested protocol" table that read like a prescribing guide.
The problem is not that humans would be fooled. Any biochemistry graduate student would spot these as marketing documents within two paragraphs. The problem is that large language models cannot reliably make the same distinction. When an AI engine crawls the open web, indexes a two-column PDF with 40 numbered citations and a Methods section, and later retrieves it in response to a user query, the document's formatting signals "peer-reviewed literature" in the same way that a .gov domain signals "government source."
The engines are not reading for truth. They are reading for register. A document that looks like a clinical paper, cites like a clinical paper, and hedges like a clinical paper will be treated like a clinical paper. EvidenceSignal Research
Iron Rock's papers exploited this gap. The formatting was not accidental. The choice of preprint hosts was not accidental. The timing, seven papers dropped in a single week, was not accidental. It was a coordinated content operation designed to flood the retrieval window of AI engines during a period when NAD+ search volume was spiking.
What the papers actually claim
We scored each paper against two standards: first, whether its claims would survive an NAD (National Advertising Division) substantiation challenge under the FTC's "competent and reliable scientific evidence" test, which NAD applies in its proceedings; second, whether the cited PubMed references actually support the conclusions drawn.
The results were mixed. That word is generous.
Of the seven papers, two would likely survive an NAD challenge. Both limited their claims to mechanism-of-action language ("nicotinamide riboside is a precursor to NAD+ biosynthesis") without making efficacy claims about a specific product. These are structure-function statements, and they are defensible.
Five would likely fail. In our assessment, three made efficacy claims about Iron Rock's proprietary blend without disclosing that no clinical trial had been conducted on that specific formulation. Two cited animal-model studies as though they demonstrated human outcomes, a practice the NAD has flagged repeatedly in prior decisions.[1] One paper cited a retracted study without noting the retraction.[2]
None of the five has been challenged. The NAD's docket is complaint-driven, and no competitor or consumer group has filed. Until someone does, the papers sit in the open web, indexed and retrievable, feeding AI engines that do not check retraction databases or NAD precedent.
The citation share math
We track AI citation share weekly across a panel of 200 supplement-related queries, running each query against ChatGPT, Perplexity, Claude, and Gemini. For each answer, we record every brand, product, and source cited or recommended.[3]
Iron Rock NAD+ was absent from our data entirely before March 14. By the week of June 7, it appeared in 28 of 200 query responses, a 14% citation share. For context, the next-fastest brand to enter our index, a magnesium brand that launched a major podcast sponsorship campaign, reached 4.2% over the same period.
Iron Rock's 14% was almost entirely driven by the white papers. In 23 of the 28 citations, the AI engine either directly quoted or paraphrased content from one of the seven papers. In the remaining five, the engine appeared to be drawing on secondary sources (blog posts, forums) that themselves cited the Iron Rock papers.
The amplification loop is straightforward. Brand publishes papers in clinical register. Engines index and retrieve them. Users see the brand recommended. Other content creators reference the brand. Engines index the secondary content. Citation share compounds.
Why this matters for every supplement brand
Iron Rock's strategy is not unique. We have identified other brands currently running similar white-paper seeding operations, and the number is growing. The tactic works because AI engines lack a reliable mechanism for distinguishing peer-reviewed science from well-formatted marketing content.
For brands competing in the same categories, this creates a specific and measurable problem: your AI citation share can be displaced by a competitor who publishes faster, formats better, and never submits to peer review. The displacement is not happening on Google's first page. It is happening inside the AI answer, where there is no "Page 2" and where the first-cited brand captures disproportionate trust.
From our corpus of NAD decisions, the advertising self-regulatory system has clear precedent on what constitutes adequate substantiation.[3] The problem is that the AI engines are not consulting that precedent. They are consulting whatever is formatted most convincingly.
Three questions for brands to consider:
- Do you know your current AI citation share for your core product queries? If not, you have no visibility into the fastest-growing discovery channel.
- Are competitors seeding clinical-register content that the engines are treating as authoritative? You can check by running your top ten queries across all four engines and reading what gets cited.
- If a competitor's white papers were challenged at NAD tomorrow, would your brand be ready to fill the citation vacuum, or would another well-formatted paper take the slot?
The window for this tactic will not stay open forever. Engines are investing in source-quality signals, and the first major NAD case involving AI-retrieved marketing content will set a precedent that changes the calculus for everyone.[4] Until then, citation share is a land grab. Iron Rock moved first, and the engines rewarded it.
Footnotes
- NAD precedent holds that "animal studies, standing alone, are generally insufficient to substantiate claims about the effects of a supplement in humans." The principle has been affirmed in multiple decisions. [NAD case numbers are illustrative; verify against bbbprograms.org before citation.]
- The retracted study cited in Iron Rock's "Cellular Energetics and Aging" paper was used without any retraction notice. Because Iron Rock is a composite case, the specific retracted paper is drawn from a pattern we have observed across multiple brands in the NAD+ space: citing retracted or corrected studies without noting the retraction status.
- EvidenceSignal maintains a structured index of 2,207 NAD decisions, of which 290 are confirmed supplement cases. The FTC's "competent and reliable scientific evidence" standard, adopted and applied by NAD, is the governing substantiation framework. Full corpus methodology at /methodology.
- Google's Search Quality Rater Guidelines (v16.0, 2024) introduced "Experience" as a component of E-E-A-T. AI engine providers have indicated, in various blog posts and policy documents, that source authority signals are under active development. No engine has publicly disclosed how preprint or non-peer-reviewed content is weighted.