A citation looks like the end of an argument. A number, a journal, a year — the claim has been checked, and you are being shown the receipt. In practice a citation is the beginning of a check, and there are four separate things that can be wrong with one, each of which requires a different look.
The paper can fail to exist. It can exist and not say what it is cited for. It can have said what it is cited for and since been retracted or corrected. Or it can be entirely sound and simply never have been tested by anyone else. Vendor copy, forum threads and AI-generated summaries all produce citations that fail in one of these ways, and they do not announce which. This piece is about doing that check quickly, and about what the identifiers you are handed actually guarantee.
What a PMID, DOI or NCT number actually proves
Each identifier answers exactly one question, and it is a narrower question than most readers assume.
A PMID means a record exists in PubMed. That is not a quality signal. PubMed contains more than 41 million records, including over 67,000 preprints, which have not been peer reviewed at all. The curated subset is MEDLINE, and the set of journals currently indexed for MEDLINE numbers 5,218 — a selective list maintained by the National Library of Medicine. A record can sit in PubMed via a PubMed Central deposit without its journal ever being MEDLINE-indexed. "It's on PubMed" and "it passed editorial selection" are different statements.
A DOI is a persistent pointer. It guarantees that a registered object exists and that the link will keep resolving. It says nothing whatever about whether the contents are correct, and — importantly — it keeps resolving after retraction.
A PMCID means the full text is freely readable in PubMed Central. That is a convenience, not a credential, and it is the difference between reading a paper and reading its abstract. Abstracts routinely omit the sample size caveat, the species, and the fact that the primary endpoint was missed.
An NCT number is a trial registration, which records an intention rather than a result. That distinction, and the batch of fabricated registry records this site verified in August, is covered separately in trial registry literacy.
Retraction is common, and it is quiet
PubMed currently tags 34,414 records with the publication type "Retracted Publication," plus a further 3,887 carrying an Expression of Concern. Of papers published in 2023, 2,308 have since been retracted; for 2024, 2,258 so far. These are not exotic events confined to obvious fraud.
They also reach this catalog directly. Four examples, all verifiable in PubMed today:
- A randomized human trial of tirzepatide with exercise training, published in Clinical Hemorheology and Microcirculation (2024;87:465-480), was retracted on 23 November 2025 per the Crossref record. It has accumulated 19 citations.
- A BPC-157 rat study in Current Neuropharmacology now appears with "Withdrawn:" prefixed to its own title, withdrawn at the authors' request.
- A MGF paper on nucleus pulposus cell apoptosis in Bioscience Reports (2019) was retracted carrying 12 citations.
- A MOTS-c bone fracture healing paper in European Review for Medical and Pharmacological Sciences was retracted.
Two features of that list matter more than the individual papers. First, the lag: roughly nineteen months separated publication and retraction for the tirzepatide trial, and citations accrued normally throughout. Second, the silence: the retraction notice for that trial is indexed under the title "Retraction notice." with no abstract at all. The machine-readable flag exists; the reason often does not. A retracted paper does not vanish, does not stop resolving, and does not stop being quoted.
Where to check, in about a minute
The Retraction Watch Database was acquired by Crossref in September 2023 and is now openly available. It appears in the Crossref REST API in the update-to field, filterable by update-type:retraction, and as a CSV dataset updated every working day. The CSV carries fields most readers would want and rarely see: RetractionNature (retraction, correction, expression of concern, or reinstatement), RetractionDate, and a controlled Reason vocabulary.
No single source catches everything. Comparing databases across 131 high-impact medical journals, a 2026 study in Accountability in Research (2026;33:2484555) found 878 retracted publications in total, of which the Retraction Watch Database retrieved 92.8%, PubMed 86.3% and Web of Science 83.6%, with per-discipline coverage as low as 52% for PubMed. Checking one database is much better than checking none, and is still not exhaustive.
For a specific paper, the fastest pass is: open the PubMed record and look for a "Retracted Publication" tag or a RetractionIn link, then resolve the DOI and see whether the publisher page carries a banner. Those two checks take under a minute and catch the large majority of cases.
The journal is a prior, not a verdict
Journal-level patterns are informative. European Review for Medical and Pharmacological Sciences, which published the retracted MOTS-c paper above, carries 420 retracted publications across roughly 13,200 PubMed records. That is not evidence that any particular paper in it is wrong. It is a reason to want an independent source before treating a single finding from it as settled.
The mirror-image error is treating every correction as a scandal. Corrections are ordinary science. The bioregulator literature review published here on 4 September covered two 2025 corrections to Epithalon papers — replaced figures, conclusions unchanged, no misconduct implied. The practical consequence of a correction is not distrust; it is that you should cite the corrected version, because "published" is a state that can be updated.
The failure mode no database catches
The most common problem with a peptide citation is not that the paper is retracted. It is that the paper is real, sound, and does not support the sentence attached to it.
This is a documented mechanism, not an impression. Greenberg's analysis in BMJ (2009;339:b2680) reconstructed the complete citation network behind a single biomedical belief: 242 papers, 675 citations, and 220,553 citation paths supporting the claim. He identified citation bias (papers refuting the belief go uncited), amplification (the belief system expands through papers presenting no data addressing it), and invention — described as the conversion of hypothesis into fact through citation alone. The same distorted network appeared inside NIH grant text obtained under the Freedom of Information Act.
That is precisely how a preclinical hypothesis becomes a product-page mechanism. This site has documented the pattern repeatedly with specific numbers: the widely circulated potency multiple for Dermorphin that belongs to a different analog; the long half-life quoted for IGF-1 LR3 that runs directionally backwards from the only direct measurement; the sequence discrepancy in the MGF literature. In each case, the citations behind the claim exist. They just do not say it.
A five-step check
- Resolve the identifier. If a DOI or PMID does not resolve, stop. A plausible-looking citation that resolves to nothing is the single strongest signal available.
- Check the retraction status. PubMed publication type plus a Crossref or Retraction Watch lookup.
- Read past the abstract for the claimed detail. Species, sample size, whether the endpoint was prespecified, whether the compound tested is the compound named. Fragments, analogs and stabilized derivatives are routinely cited as their parents.
- Check the author affiliations, not the lead institution's country. A paper can look independent and still share authors with the originating group — the distinction that separated genuine replication from lineage extension in the bioregulator literature.
- Ask whether anyone else has found it. A single unreplicated result is a hypothesis regardless of how good the journal is. Where a claim sits on that ladder is the subject of the evidence hierarchy primer.
None of this requires database access, subscriptions or specialist training. It requires treating a citation as a claim to be opened rather than a credential to be counted.
FAQ
Does a retraction mean the finding was fraudulent? No, and assuming so is its own error. The controlled reason vocabulary spans honest error, unreliable data, image duplication, authorship disputes, plagiarism, and withdrawal at the authors' own request. The Current Neuropharmacology BPC-157 example was withdrawn at author request with no substantive reason given in the notice. What a retraction reliably tells you is that the paper should no longer be used as support, not why.
If a paper is still cited after retraction, is that dishonest? Usually not. It is mostly a timing artifact — the tirzepatide trial spent about nineteen months in normal circulation before the retraction landed, and secondary sources written in that window are frozen at the earlier state. This is why the check has to happen at the point of use, not once.
Why does this belong under foundations rather than quality control? Because it is upstream of everything else. A certificate of analysis answers questions about the material in a vial; the literature answers questions about what the molecule is understood to do. Both are documents making claims, and both are read the same way — by asking what specific question the document actually answers, and declining to extend it further than that. Browse the compound library for how the research profile for each compound is summarized.
This article is educational and for the laboratory research community. Trulogic Labs products are sold for laboratory and research use only and are not for human consumption.