citation-hygiene-checklist

Citation Hygiene: Catching Retracted, Corrected and Phantom References Before You Submit

Every reference list contains assumptions. When you cite a paper, you are telling the reader that it exists, that it says what you claim it says, and that it is still considered part of the scholarly record. Each of those assumptions can fail. Papers are retracted, corrected or flagged with expressions of concern. Citations get copied from other reference lists with errors intact. And with the spread of AI writing tools, a new failure has appeared: references that look plausible but do not exist at all. This guide describes a practical routine for checking citations before submission, and how a source-linked workbench such as yila.ai reduces the risk from the start.

Three kinds of citation problem

Retractions and corrections. Thousands of papers are retracted every year, for reasons ranging from honest error to fabrication. Retracted papers continue to be cited long after the retraction notice, often because authors copied the reference from an older review. Corrections and expressions of concern are less dramatic but can change the meaning of a result.

Misattribution. The paper exists, but it does not say what you claim. This happens when authors cite from memory, cite a secondary source's summary rather than the original, or cite the abstract without reading the results. Misattribution is the most common citation problem and the hardest to detect automatically.

Phantom references. The reference has a real-looking author list, title, journal and year, but no such paper exists. Phantom references used to be rare and usually came from typos or garbled metadata. They are now a well-documented side effect of using general-purpose chatbots to draft literature summaries, which can generate fluent text with invented citations.

Each problem needs a different check.

Step 1: Verify that every reference exists

Start with existence, because it is the easiest to test and the most embarrassing to get wrong. For each reference:

• Look up the DOI if one is given and confirm it resolves to the stated title and authors.
• If there is no DOI, search the exact title in a major index such as Crossref, OpenAlex, PubMed, Scopus or Google Scholar.
• Check that the journal, volume, year and page range match the record.

Reference managers can automate much of this by re-importing metadata from DOIs. Any reference that cannot be found in two independent indexes should be treated as suspect until you hold a copy of the document in your hands.

If you used an AI tool at any point while drafting, do this check for every single reference, not a sample. Phantom citations are not distributed randomly; they cluster in passages where the tool was asked to "support" a claim.

Step 2: Check retraction and correction status

Retraction Watch maintains a large public database of retractions, which is now integrated into Crossref metadata. Many reference managers can flag retracted items automatically when this data is available. Publisher websites display retraction and correction notices on the article page, and PubMed links notices to the original records.

For each flagged item, read the notice. Not every retraction invalidates every use of a paper: you may legitimately cite a retracted article when discussing the retraction itself. But you must say so explicitly, and you should never cite a retracted finding as evidence.

Corrections deserve the same attention. A corrected table or a revised effect size may change whether the paper supports your point. Expressions of concern signal that the publisher has doubts but has not completed an investigation; cite such papers cautiously and note the concern.

Step 3: Confirm that each citation supports its claim

This step cannot be fully automated, but it can be organised. Go through your manuscript claim by claim and, for each cited statement, open the source at the relevant location. Ask:

• Does the source report this finding, or merely mention it in the introduction while citing someone else?
• Does the source's population, setting and method match the context in which you cite it?
• Are you citing the strength of the finding accurately, or have "associated with" and "causes" drifted?

Where a source turns out to be secondary, trace the claim to the original and cite that instead. This is often where you discover that a widely repeated "fact" rests on a single small study, or on no study at all.

A workbench that keeps each claim linked to the exact sentence or table it came from makes this step dramatically faster. Instead of searching a PDF for a half-remembered number, you click through from the claim to the passage. Yila was designed around this idea: every extracted statement carries its origin, so the check is built into the drafting process rather than bolted on at the end.

Step 4: Watch for citation chains

Many errors propagate through chains of citations. Review A misreports a finding from study X; review B cites review A; your draft cites review B. Each step looks reasonable in isolation. The fix is simple but tedious: for any central claim, read the primary study yourself.

Citation-chasing tools that show who cites whom can help identify where a claim first appeared. If you find that all the citations for an important claim lead back to a single source, say so; readers deserve to know that the evidence is thinner than the citation count suggests.

Step 5: Clean the metadata

Finally, fix the mechanical details. Inconsistent author names, missing DOIs, wrong years and broken URLs make references harder to verify for reviewers and readers, and they are a common reason for copy-editing delays. Use a reference manager, re-pull metadata from DOIs, and spot-check the formatted output against the journal's style.

Building the routine into your workflow

Doing all five steps at the end of a project is painful. Spreading them out is much easier:

• When you add a source, confirm it exists and record its DOI.
• When you extract a claim, record the exact location in the source.
• When you draft a paragraph, cite from your extracted claims rather than from memory.
• Before circulating a draft, run the retraction and existence checks.
• Before submission, do the claim-support check for every central statement.

This pattern also makes it much safer to use AI assistance. If the assistant can only draw on sources you have actually collected, and every statement it drafts comes with a link to its origin, phantom references simply cannot appear. That is the design principle behind the Yila research agent: no claim without a traceable source.

What reviewers and editors look for

Editors increasingly run automated checks for retracted citations and unresolvable DOIs. Peer reviewers in your field will often spot misattributed claims about well-known studies. A manuscript with clean citations signals care and makes the reviewer's job easier, which rarely hurts.

A compact checklist

• Every reference resolves in at least one major index, with matching metadata.
• No retracted paper is cited as evidence; any retraction discussed is labelled.
• Corrections and expressions of concern have been read and accounted for.
• Every central claim has been checked against the primary source at the cited location.
• Secondary citations have been replaced with primary ones where possible.
• Metadata is complete and consistent.

Final thoughts

Citation errors are rarely dishonest, but they erode trust in a piece of work out of proportion to their number. A short routine, spread across the life of a project, catches almost all of them. Combine that routine with tools that keep every claim attached to its evidence, and your reference list becomes something you can defend line by line.