When a Claim Traces to Nothing

A figure with two reviews and no measurement under it

The claim is that GHK is present in human plasma at about 200 ng/mL at age 20, declining to about 80 ng/mL by age 60. It appears in the third sentence of the introduction to Pickart, Vasquez-Soltero and Margolina's 2015 review in BioMed Research International, with no citation attached to that sentence. Two sentences later, the statement that GHK-Cu was isolated in 1973 from human albumin does carry one; the sentence between them, on that decline coinciding with a decrease in regenerative capacity, carries nothing either. The review's first reference marker sits two sentences past the number that needed it.

The same pair of figures opens the abstract of a 2020 review in Aging Pathobiology and Therapeutics, there attributed to serum rather than plasma, again with nothing attached. PubMed searches for a primary study measuring GHK concentration in human plasma stratified by age returned no such study. Two reviews five years apart, one pair of numbers, two different matrices named, and no located measurement underneath either statement. The figure is not obviously wrong. It is simply not traceable to anything that measured it.

That pattern is the most common defect in this subject matter: a specific number, repeated in sources that look authoritative, resting on nothing a reader can open. It is common enough that every record on this site carries a field for it. What follows is how the pattern forms, and why a claim that traces to nothing is published here rather than removed.

The circuit

The chain observed most often runs in four steps. A figure appears on a vendor product page with no source. An aggregator repeats it and attaches a citation — a real journal, a plausible year, the right subject area — to a paper that does not contain the figure. Forum posts quote the aggregator, usually keeping the number and dropping the citation. A later product page cites the forum thread, or repeats the aggregator's citation without opening the paper. Four independent-looking appearances, one origin, no measurement anywhere in the sequence.

Nothing in that sequence requires dishonesty. Each step is the ordinary behaviour of someone assuming a number in several places has been checked by at least one of them. Bikhchandani, Hirshleifer and Welch formalised the structure in 1992: when actors observe the choices of those ahead of them but not the private information behind those choices, they discount what they know and copy what they see. Public information stops accumulating after the first few decisions, and the convergence that follows carries no more evidence than the first decision did.

What the network looks like when someone maps it

Greenberg's 2009 analysis in the BMJ is the closest thing to a full map of this process inside peer-reviewed literature. It assembled the complete citation network of every PubMed-indexed English-language paper addressing one belief: that beta amyloid is produced by, and injures, skeletal muscle in inclusion body myositis. The network held 242 papers and 675 citations, with 220,553 citation paths supporting the belief. Citation bias operated against the papers that refuted or weakened it. Amplification expanded the belief system through papers presenting no data addressing it. Invention converted hypothesis into fact through citation alone.

That network was then extended into the text of NIH-funded grants obtained under the Freedom of Information Act, where the same phenomena were present and sometimes used to justify requests for funding. The conclusion named the mechanism: citation is social communication as well as scholarly method, and distortions in its social use can generate information cascades that produce unfounded authority. The setting matters. That was indexed, peer-reviewed literature at the best-governed end of the range. A product page has no review at all.

Repetition, measured

Why a repeated figure feels credible has been measured for decades. Dechêne and colleagues pooled 51 studies of the repetition-induced truth effect in a 2010 meta-analysis and found it on both of the criteria that literature uses, larger when repeated statements were compared against new ones than when the same statement was rated twice. The account usually offered is processing fluency: a statement encountered before is easier to process, and ease is read as a marker of truth.

Two later findings narrow that account. Fazio, Brashier, Payne and Marsh reported in 2015, from an experiment with forty undergraduates, that illusory truth effects occurred even when participants demonstrated knowledge of the correct answer on a post-experimental check, a failure they termed knowledge neglect. Fazio, Rand and Pennycook followed in 2019 with a simulation and a preregistered experiment indicating that repetition raised perceived truth across all levels of plausibility, and that the larger effect seen for ambiguous items reflected the psychometric properties of the task rather than a mechanism blocking repetition for implausible ones.

Hassan and Barber tested how much repetition the effect needs. Participants — 51 in the first experiment and 57 in the second, after attrition and exclusions — saw trivia statements up to nine times and up to twenty-seven times respectively. Perceived truthfulness rose as repetitions increased, but logarithmically, and the largest single increase came from encountering a statement for the second time. Pennycook, Cannon and Rand had shown in 2018, in two online samples of 949 and 940 adults, that a single prior exposure to a fabricated news headline raised its perceived accuracy within the session, that the effect was still present in the 566 participants who returned a week later, and that labelling stories as disputed did not remove it. The second appearance does most of the work.

The citation is rarely opened

A citation attached to a repeated figure is checked less often than its presence implies. Simkin and Roychowdhury modelled the propagation of misprints in citation strings, the same typographical error recurring across many reference lists, found the misprint distribution followed a Zipf law, and estimated that only about 20 percent of citers had read the original. Jergas and Baethge screened 559 studies for their 2015 meta-analysis of quotation errors, included 28, and estimated a total quotation error rate of 25.4 percent, with major errors at 11.9 percent.

Errors also survive formal correction. Frasco and colleagues examined 245 papers retracted over 28 years from nine anaesthesiology journals; of the 211 retracted before January 2020, Scopus recorded 1,307 post-retraction citations. Eighty percent of those citations affirmed the validity of the retracted work, and 5.2 percent acknowledged the retraction or the misconduct behind it. Rekdal's 2014 account of academic urban legends traced the belief that spinach is a strong dietary source of iron to an apparent decimal point error, then followed its reproduction through decades of respectable publication.

Five ways a claim fails to trace

The untraced entries in this ledger fall into recognisable types. First is the uncited review assertion, of which the GHK plasma figure is one instance. That same 2015 review supplies another: it states that GHK is capable of up- and downregulating at least 4,000 human genes. The primary work it goes on to describe, Campbell and colleagues in 2012, identified 127 genes whose expression tracked regional emphysema severity across 64 human lung tissue samples taken from eight lungs of smokers with COPD, and used the Connectivity Map, a gene-expression connectivity resource, to nominate GHK as a compound reversing that signature.

Second is the attributed but absent source, where a claim names a journal and the journal does not have it. Degradation percentages for tesamorelin circulate attributed to the Journal of Pharmaceutical Sciences; a PubMed query restricted to that journal returns no records for the compound. Third is the figure that contradicts a registry it could have been checked against. A molecular weight of 4813.45 circulates for tirzepatide, against 4,813.53 in FDA labelling and 4,813 in PubChem. The discrepancy has no practical consequence; it is logged because it travels with the authority of a sourced value.

Fourth is the number whose primary source exists but does not contain it. Receptor potency triples for retatrutide are widely quoted; the characterisation paper reports the receptor profile in relative terms in its abstract, has no PubMed Central deposit, and its assay table cannot be opened. Fifth is the comparison for which no direct study has reported. Indirect network meta-analyses of separate trials have placed retatrutide above tirzepatide on the outcomes those trials measured, and the comparison is logged here as untraced to any head-to-head result. The head-to-head phase 3 is registered as NCT06662383, estimated 800 adults, listed as active and not recruiting, primary completion estimated November 2026, no results posted.

The search that failed, recorded next to the claim

Vialog currently carries 17 records holding 128 ledger entries, each with a claim, population, route, sample size and source: 50 animal, 47 human, 31 in vitro. Of those 128, 124 cite a PMID; the remaining four cite a regulatory label or a trial registration, because that is what exists for them. The reference lists hold 191 entries and 133 distinct PMIDs. Alongside them sit 51 claims marked untraced, each published with a note recording which database was searched, on what terms, and what came back. Thirty-three of the 51 name PubMed explicitly.

One figure in that set deserves flagging by the site's own standard. The untraced entries run at exactly three per record, in all 17. That uniformity is a property of how many are surfaced during review, not a measured constant of the subject matter. It is the count the current process produces, not the count that exists, and it should move as records are revisited. Three per record is a floor, not a measurement.

The argument for publishing rather than deleting is narrow. A deleted claim leaves no evidence that it was ever examined, and the figure will still be met elsewhere, arriving with the fluency four prior appearances have already given it. A recorded negative result does what a deletion cannot. It dates the attempt, and it leaves the next reader a failed search with terms attached rather than a number alone.

A failed search is not a disproof, and the notes say so. Material sits behind paywalls, in patents, in industry files never deposited, and in journals no index covers; a query can also be built badly. What a note establishes is that one search on one date did not find a source, and that the claim was carried forward marked rather than laundered. The untraceable number stays on the record here; it stays on it as an untraceable number.