Audit your own paper's references before you submit it

A paper carries forty to eighty references. A thesis carries five hundred. So the reasoning goes: at this length I would notice.

Except the failures in a paper are not smaller versions of the failures in a thesis. They are different failures, produced by three things a thesis mostly does not have — several people writing into one list, a reference manager doing the formatting, and a review cycle long enough for the literature to move underneath you.

Nobody owns the reference list

In a thesis one person adds every entry. In a paper, three authors each add sources from their own libraries, and the list is merged at the end by whoever is assembling the manuscript. Merging is where the duplicates come from, and they are not the obvious kind.

From author A's library
  Okonkwo, C. B., & Ferreira, L. (2022). Workflow variance in
  regional centres. Health Systems Review, 14(3), 201–218.

From author C's library
  Okonkwo CB, Ferreira L. Workflow variance in regional
  centres. Health Syst Rev. 2022;14(3):201-18. doi:10.xxxx/hsr.2022

Merged list        both survive — different styles, same paper
In the text        cited once, so one entry is never referenced

The merged list now has an orphan: an entry nobody cites, because the citation points at the other copy. It looks like sloppiness to a reviewer and it is really an artefact of two reference managers exporting the same DOI in two house styles. Neither author did anything wrong; the merge did it.

Most reference defects in a co-authored paper are produced at the moment of assembly, by people who were each individually correct.

The mirror image happens too. Two genuinely different papers by the same group in the same year collapse into one entry during the merge, and every citation to both now points at whichever one survived. That one is invisible to a reviewer and to you, because the list looks shorter and tidier than it should.

Your manager exported what it had, not what is true

Reference managers are excellent at formatting and indifferent to currency. The record you imported in 2023 is the record you will export in 2026, unchanged, because nothing in the tool goes back to check whether the world moved.

Three things move. A preprint you cited gets published, so the correct citation is now a journal article with a different year, volume and DOI. A DOI is reassigned or the publisher restructures, so the link resolves somewhere else — or nowhere. And occasionally a paper is retracted, which is the one that genuinely matters, because citing a retracted result without noting it changes what your own claim is standing on.

At submission                    At acceptance, 8 months later

Preprint, bioRxiv 2024      →    Published, J. Mol. Biol. 2025
                                 different year, volume, DOI

doi:10.xxxx/old-host/1183   →    resolves to a landing page
                                 the article moved hosts

Alvarez et al. (2021)       →    RETRACTED, notice issued Feb 2025
                                 still cited as a supporting result

None of this is caught by reading your own paper, because your paper did not change. The reference decayed while sitting still. We wrote about how to judge which older sources actually need replacing — the same reasoning applies here, compressed into the months a manuscript spends in review.

The audit, in the order that pays

This takes under an hour on a paper-length list, and the order matters because each step makes the next one cheaper.

One: de-duplicate before anything else. Sort the list by first author surname and read down it looking only for the same paper twice, in two styles. Do this first because every later check is wasted effort on an entry that should not exist.

Two: tally in both directions. Every citation in the text must reach an entry; every entry must be cited somewhere. In a paper the second direction catches more than the first — orphans are the merge artefact, and they are what a copy-editor will query. What a healthy result looks like is one or two of each, not zero.

Three: re-resolve every identifier. Take each DOI, PMID or URL and confirm it still points at the paper you think it does. This is the step people skip and the step that catches the preprint that grew up, the moved host, and the retraction. If you do only one thing from this list, do this one.

Four: check the entries you inherited. References that came from a co-author's library, or from a previous paper of your own, deserve more scrutiny than the ones you added last month — not because colleagues are careless, but because you have never personally verified them and you are the one submitting.

Five: check what you are citing for. A correctly formatted reference to a real paper can still be attached to a claim that paper does not make. No tool can check this and no reviewer will thank you for it, but it is the failure with the largest consequences, and it is worth one deliberate pass.

Then do it again at revision

The audit you run before first submission is not the one that matters most. Revision is where references are added under time pressure, in response to a reviewer, often to literature you have only just read — which is precisely the condition that produces the cited-but-not-listed gap.

It is also the moment the decay above has actually happened. A reference list that was current in March is not necessarily current in November, and the version that goes to production is the November one.

A practical rule: re-run the identifier check at revision, always, even if you added nothing. It is the cheapest step and the only one whose result changes without you touching the document. There is a fuller guide to refreshing a list that has aged if the gap has been longer than a few months.

What this is and is not

An audit of your own work is a quality step, and most of what it returns is exactly that: duplicate entries, stale DOIs and orphaned references are what happens when several competent people assemble a document over a year. They are ordinary, fixable in an hour, and invisible to re-reading your own prose.

The exception is worth naming plainly, because it is the one with consequences beyond a corrigendum. A reference that resolves to nothing — no record, no DOI, no such paper — is not a formatting problem. Citing work that does not exist is academic misconduct, and a journal that finds it after publication does not issue a correction. You are far better placed to find it now than anyone is later.

The point of doing it yourself, before anyone else does, is simply that you are the only person who can still fix it cheaply. After submission it becomes a copy-editor's query list; after publication it becomes a correction.