Citation laundering: tracing AI-draft stats back to the real source

Tracing a statistic back past the blogs to the page it first appeared on.

A cold feeling, low in the stomach. That is what a statistic in an AI draft should give you when nobody can say where it came from. In their 2026 paper on GenAI citations, David Resnik and Mohammad Hosseini list cases that were caught: PLOS ONE retracted a study in which 18 of 76 citations were hallucinated, and a perspective piece in Digestive Diseases and Sciences carried 12 hallucinated references out of 14. Caught, yes. But published first.

Marketing content has no retraction desk. A figure enters a blog, another blog quotes that blog, an AI draft quotes them both, and somewhere along the way the number, its year or its population stops matching the document that produced it. I call this citation laundering. I do not believe subject-matter review stops it. A reviewer asks whether a number sounds right. Laundered numbers always sound right.

The same authors point to something quieter, almost worse. A 2025 study they cite checked 176 citations generated by GPT-4o in mental health: 56% contained errors, and one in five was hallucinated outright. So the fix cannot be a newer model alone. It has to be a habit, and a stubborn one. One plain question per figure: where is the first page?

What is citation laundering, and where does it begin?

Citation laundering is the process by which a statistic loses its primary source as it passes through AI drafts and secondary pages, yet still looks credible on the final page.

Before you trust any figure in a draft, ask four plain questions:

1. Who first published this number, and in which document?
2. How many pages sit between that document and my draft?
3. Do the number, its year and its population still match the original?
4. Can I open the full text, or only someone's summary of it?

In 2025, the UNMC McGoogan Library described a study in which 98.1% of citations generated by GPT-3.5 were fabricated and 20.6% of those from GPT-4 were fake. I read that twice. Then a third time. One in five, from the better model! The common assumption is that a fake citation looks fake. It does not. The same post explains that a chatbot may pair a real journal with an invented article, or put a real expert's name on a paper that person never wrote.

UNC Charlotte's Atkins Library describes the machinery coldly. A model trained on real citations builds "a statistically probable citation that fits with the topic requested in the prompt." Its sharpest line is shorter: "The tool doesn't care if the answer is accurate, just plausible." Plausible. I keep returning to that word. Plausible is exactly what a tired reviewer approves at the end of the day.

And then the second stage begins, quietly. The library warns that some hallucinated citations are already appearing in Google Scholar, so a listing there proves nothing. One academic editing firm describes these references as metadata borrowed from real titles, authors and venues, stitched into something that never existed. Once such a figure is indexed, the next writer copies it. Then another. Each hop washes it a little cleaner.

Put 5 of these sources side by side and one pattern repeats: the fabrication happens once, but the laundering happens at every copy. Researchers who study misconduct point out that so many people now draft with GenAI that many publications may already carry generated citations, hallucinated ones among them. The implication is uncomfortable. So look first at the page your writer copied from, not the chatbot: does that page name its own source, and when was it published?

At AEO Content we have run 26,577 real AI-visibility audits, and I would rather lose a strong number than publish one I cannot trace to its first page. Ask one question of every figure before it goes into research-driven content built on evidence: who counted this, and where did they first print it? Which leaves the stranger problem, the one that kept me awake: how a true number changes while nobody along the chain is lying.

Why don't citation checkers, AI detectors or tracing tools catch laundered statistics?

Most tools ask whether a source exists. None of them asks the harder question: does the cited passage state the same number, for the same people, in the same year?

I wanted to believe the tools. Truly. Then I read a PREreview of Shi et al. and their MedResearchBench, a benchmark for medical manuscripts written by language models. Citation hallucination rates across the 7 systems tested ran from 2.9% to 36.8%. Adding a citation verification and repair pipeline to the best system lifted its citation integrity score from 40.0 to 90.9. A triumph, it seemed to me, for about a minute.

The reviewer saw what the pipeline actually did. It confirmed that replacement references exist in CrossRef or PubMed. It never checked whether they support the claims sitting beside them. The reviewer's judgment was that the benchmark is "easily gamed in ways that may not improve integrity." Worse, the reviewer recalculated one of the paper's own rates as 4.08% where the paper printed 2.9%. A number drifted inside a single document, before any blog had touched it!

Tracing tools fail in a different, quieter way. Cobus Greyling describes OLMoTrace as a system that finds verbatim matches between a model's output and its multi-trillion-token training data, "in seconds." Impressive. And yet he concedes the obvious point: tracing an output back to training data does not guarantee truth, because the corpus itself can hold errors. A laundered blog post traced perfectly is still a laundered blog post.

Even the checkers stumble. A reference-manager vendor that sells its own citation checker reported that the tool flagged one ChatGPT reference as unmatched when the reference was in fact real. A provenance vendor makes a similar admission about its product: a timestamped record shows what you were working with and when you checked it, not whether the citation was correct. I respect that honesty. It frightens me a little, too.

A commenter in one Reddit thread on spotting AI writing argued that detectors can be fooled but workflow evidence cannot; half right, perhaps, since a trail shows what you opened, not whether it was true. The number itself is another matter. The research section above walks through these findings. If you want the wider picture of how answer engines weigh the pages they cite, the AEO knowledge base guides are a calmer place to begin. But the comparison itself still waits for a person, a person with the original open on the desk. Who does it, and in what order?

What verification workflow should a marketing team use before publishing a statistic?

Write from the primary document, never the chain. Our pilot cut web sources per article from 30 to 38 down to zero, and every product claim checked out.

Six steps, in this order. Skip none, least of all when you are tired, which is exactly when you will want to.

1. Pull every number, date, percentage and count out of the draft into a separate list.
2. Find the page the draft took each figure from.
3. Follow that page's citation, then the next one, until you reach the document that produced the figure. A line crediting one outlet's coverage of somebody else's report means you have not arrived yet.
4. Open the full text and find the exact passage.
5. Compare the number, the population, the model, the date and the method with your own sentence.
6. Record one verdict: keep it, qualify it with its year and context, or cut it.

Why such stubbornness? Because the secondary source is often wrong in the most reasonable way. Our knowledge-base pilot read a product end to end, from the admin setting to the screen the user actually sees. The admin page described two-factor sign-in as using an authenticator app. The real sign-in screen sends a one-time code by text and email. Strange. Nobody had lied. The documentation had simply drifted from the thing it described, and an article written from that documentation would have repeated the drift with perfect confidence.

The first round of that pilot still had web research. About a third of one article became regulation, attack statistics and industry framing, and each article carried 15 to 18 first-person opinions. With the code-only rule enforced, the second round had zero web sources, zero opinions and zero outside mentions. I felt something close to relief reading that. Less to verify. Less to launder.

None of this asks for genius. It asks for patience, and a builder's temperament. Alex and I bring a combined 40+ years of SEO and content infrastructure experience to AEO, and our team includes a full-stack engineer and content infrastructure architect with 20 years of building enterprise systems. The scoring behind our work spans 11,000+ domains across 15 sectors and 28 categories. Ask the writer one thing for every figure: which page of which document, and may I see it?

For plainer answers on how AI engines treat sourced pages, the AEO FAQ on AI search optimization makes a good companion to this list. And when step five fails? When the original says one thing and the draft says another? Then the draft loses. Every time, without a word of argument.

Where does citation laundering go from here?

Verification is moving from a polite courtesy to a requirement, and a figure without a reachable primary document will soon cost more to publish than it ever earns.

Cobus Greyling imagines a day when every chatbot answer carries a "source" button. I want that day. And yet a button only points somewhere. It cannot tell you whether the page it points to copied its number from another page, which copied it from a third. Quillcademia, a company that sells academic writing tools, points to a study reported in Nature: suspected hallucinated citations rose from 0.3% of papers to 2.6% in 2025.

I sometimes think editors would sleep better with one rule pinned above the desk. No first page, no number. Perhaps that is naive. Perhaps not.

So start with the next draft on your desk. List its numbers. Then go looking for the first page of each one, and notice, quietly, how many you cannot find.