
Reference software can format a record it cannot verify. A reliable audit checks whether the source is real, supports the sentence and appears correctly in both places.
Audit meaning before punctuation
A reference manager can turn the wrong record into perfect Vancouver style. It cannot determine whether that paper supports the sentence, whether a correction changed the result or whether an editorial has been cited as primary evidence.
The ICMJE Recommendations state that authors are responsible for citing references accurately and should be able to attest that each reference supports the associated statement. They say references should be verified against a bibliographic source such as PubMed or against the original sources. They also advise citing original research where possible, avoiding abstracts when full reports exist, clearly identifying preprints and checking that cited articles have not been retracted except when the retraction itself is relevant.
A reference list audit therefore has two layers:
| Intellectual audit | Mechanical audit |
|---|---|
| Does the source support the exact claim? | Does every citation have one matching record? |
| Is it the appropriate evidence type? | Is metadata accurate and complete? |
| Is its status current? | Does formatting follow journal style? |
| Is the citation balanced and fair? | Do DOIs resolve, and are pages and author details correct? |
Run them in that order. Formatting first creates polished noise and has to be repeated after weak sources are replaced.
Freeze the draft and assign responsibility
Choose a stable manuscript version. Ask co-authors to pause citation edits while the audit is underway or track every change in a shared log. Otherwise one person reconciles citation 24 while another deletes it.
Assign sections to authors familiar with the claims, but appoint one person to run the final whole-manuscript reconciliation. For each reference, record:
- Where it is cited.
- Which proposition it supports.
- Who opened and verified it.
- Date the status was checked.
- Any correction, version or access note.
Do not divide responsibility so finely that nobody sees duplicate records, inconsistent use of one source or imbalance across the paper.
Verify the claim against the original source
Read the sentence containing the citation, then open the full source. Ask:
- Is the cited population the population named in the sentence?
- Does the study design permit the type of claim?
- Is the stated outcome actually reported?
- Are direction, magnitude and uncertainty represented accurately?
- Does the source support the whole sentence or only one clause?
- Is the claim coming from the article's data or from its discussion of another source?
Citation distortion often occurs through small expansions:
| Source reports | Manuscript incorrectly says |
|---|---|
| Association in one cohort | X causes Y |
| Improvement in surrogate marker | X improves clinical outcomes |
| No threshold-crossing difference | X and Y are equivalent |
| Findings in adults at one centre | Findings apply to all patients |
| Authors hypothesise a mechanism | The mechanism is established |
Repair the manuscript claim, replace the source or add a source for the unsupported clause. Do not stack several citations after a broad sentence unless each genuinely supports it.
If you learned a fact through a review but the statement concerns a particular experiment, inspect and cite the primary report where appropriate. Do not cite a paper that cites a paper that cites the result. Secondary sources remain useful for synthesis and context when that is the claim being made.
Our guide to scientific writing for medical journals explains why the Introduction should cite only directly pertinent references; the same claim-level discipline applies throughout.
Check publication and integrity status
Article status can change after it enters a library. For each reference, search the DOI or exact title on the publisher page and in a trusted bibliographic database.
Look for:
- Correction or erratum.
- Expression of concern.
- Retraction or withdrawal.
- Updated guideline or living review.
- Preprint replaced by a journal article.
- Duplicate publication.
PubMed links many corrections, expressions of concern and retraction notices to indexed records, and the ICMJE treats PubMed as the authoritative source on retractions for MEDLINE-indexed journals. Crossmark, when available on a publisher page or PDF, can indicate whether updates exist. Neither should be treated as universally complete; check the publisher record too.
A retracted paper should generally not support a scientific claim. If the paper is discussed because it was retracted, say so in the text and cite the retraction notice as the journal style requires. An expression of concern does not automatically settle validity, but it is material context; reconsider whether the claim needs that source and disclose status when relevant.
Corrections require reading. A corrected author affiliation may not affect the claim; a corrected denominator or figure might. Use the corrected version and ensure your wording still holds.
Run status checks close to submission and again during major revision if months have passed.
Select the right source, not merely a source
An accurate citation can still be a poor choice. Match source type to proposition:
| Claim | Usually strongest source type |
|---|---|
| Effect of an intervention | Relevant systematic review or comparative primary study |
| Current professional recommendation | Current official guideline |
| Disease frequency | Appropriate surveillance or population study |
| Reporting requirement | Original reporting guideline or journal instruction |
| Biological mechanism | Direct primary evidence, with model identified |
| Historical fact | Contemporaneous or authoritative historical source |
“Strongest” does not always mean newest or highest in an evidence hierarchy. A recent review may not apply to the same population. An official guideline may be authoritative for a recommendation but not primary evidence of treatment effect.
Review currency. Replace superseded reporting guidelines and old policy documents when the claim is current. Do not automatically delete older landmark studies when they establish discovery or method history.
Review balance. Search the Discussion for phrases such as “consistent with previous studies” and check whether contrary evidence was omitted. Citation count is not a vote. Include relevant disagreement and explain methodological differences.
Also inspect self-citation and citation concentration. Authors' prior work may be directly relevant, but it should not crowd out independent evidence. Repeatedly citing one research group can create a distorted field map.
Handle special source types correctly
Preprints
Label preprints clearly. Before submission, search for a peer-reviewed version. Do not swap records automatically: compare authors, methods, results and conclusions because versions can differ. Cite the version actually used and follow journal policy.
Conference abstracts
Search for a full report. Abstracts provide limited methods and may contain preliminary results. If no full report exists and the abstract is essential, identify it as an abstract and moderate the claim.
Guidelines and web pages
Record the issuing organisation, title, version or update date, URL and access date if the style requires it. Link to the specific document, not an organisation's home page. Save a copy or persistent archive where policy and permissions allow.
Data, software and code
Cite research objects as first-class outputs when they informed the study. Include creators, title, repository or publisher, year, version and persistent identifier. Our guide to citing data, software and code gives detailed patterns.
Reconcile citations in both directions
After intellectual changes are complete, match manuscript and list.
Forward check: every in-text citation points to exactly one reference-list record.
Reverse check: every reference-list record is cited at least once for a substantive reason.
Look for:
- Orphan references left after deleted paragraphs.
- Citation placeholders such as “REF,” question marks or comments.
- Duplicate records imported with abbreviated and full journal titles.
- Author–date citations with the same author and year but missing suffixes.
- Numbered citations that changed order after tracked edits.
- Citations in tables, figure legends and supplements omitted from the main library.
- References typed manually and therefore invisible to software.
Do not edit numbered citations by hand. Refresh fields through the reference manager after accepting or rejecting tracked changes, then inspect the rendered document. Back up first because field refreshes can expose broken links.
Search the entire submission package, not only the main text. Some journals require one continuous numbering sequence; others treat supplements separately.
Validate metadata against the source
Open the publisher record or trusted index and check:
- Author names and order.
- Article or chapter title.
- Journal or book title.
- Publication year.
- Volume and issue where required.
- Page range or article number.
- DOI or other persistent identifier.
- Publication type and version.
Import errors are common with group authors, particles in surnames, online-first dates, e-locations and article numbers. A DOI is not decoration: test that it resolves to the intended object. Remove duplicated DOI prefixes and tracking parameters.
For journal titles, follow the target style. ICMJE points authors using MEDLINE style to the NLM Catalog for journal abbreviations, but the target journal's instructions control the final form.
Do not “correct” an article title into preferred spelling; preserve the published title except where the style has a defined capitalisation rule. Retain accents and diacritics accurately.
Check citation ethics and permissions
References should acknowledge intellectual sources, not serve as currency. Remove decorative citations that do not support a claim. Do not add references solely because a reviewer, editor or target journal requested them; assess relevance and explain respectfully when a requested citation is not appropriate.
Watch for:
- Citation of work not read.
- Excessive author self-citation.
- Coercive clusters from one journal.
- Citation of predatory or pseudo-journal content as evidence.
- Failure to cite the original method, instrument or dataset.
- Citation of AI-generated text as a primary source.
Citation is separate from permission. Citing a published figure or table does not authorise reproduction or adaptation. Check the licence and publisher terms, obtain permission where needed, and use the wording required in the legend. The guide to reusing published figures covers that workflow.
Quotation also needs care. Verify wording and page location, use quotation marks, and keep quotation within licence and fair-dealing or fair-use limits applicable to the submission.
Apply journal style last
Download the journal's current instructions and reference style. Check:
- Numbered or author–date system.
- Citation order and placement relative to punctuation.
- Number of authors before “et al.”
- Journal title abbreviation.
- DOI and URL treatment.
- Access dates.
- Formats for preprints, datasets, software and web documents.
- Reference limits for the article type.
Install the official style where available, but inspect at least one example of each source type manually. Styles and library records can both contain errors.
Do not meet a reference limit by removing inconvenient contrary evidence. Tighten background, remove redundant citations and retain sources essential to claims and balance.
Run the final reference list audit
Use one checklist on the locked submission copy:
- Every substantive citation was opened and verified
- Source population, design and outcome match the claim
- Causal and certainty language does not exceed the source
- Primary or authoritative sources are used where appropriate
- Current guidance has replaced superseded versions
- Corrections, concerns and retractions were checked
- Preprints and abstracts are clearly identified
- Contrary evidence and independent groups are represented fairly
- Every in-text citation has one list entry
- Every list entry has a genuine in-text citation
- Tables, figures and supplements were included in reconciliation
- Metadata and identifiers were checked against original records
- Permissions were obtained separately from citation
- Target journal style and reference limit are met
- Citation fields were refreshed and the clean export inspected
Archive the verified library, manuscript version and audit log together. When a reviewer asks for a changed claim or source, update the log and rerun the affected checks rather than assuming the list remains valid.
A good reference list audit is invisible in the published article. Readers simply reach the right source, find that it says what the manuscript claims and encounter no avoidable surprises about status, version or attribution.
Frequently asked questions
What is a reference list audit?
It is a structured check of both citation meaning and mechanics. Authors verify that each source exists, has the correct publication status, supports the precise nearby claim, is appropriate and current, and has accurate metadata. They then reconcile every in-text citation with the reference list and apply the target journal's style.
Do authors need to open every cited source?
At least one accountable author should open and verify every source, especially those supporting substantive claims. Relying on abstracts, reference-manager imports or another paper's description risks citation distortion.
How can I check whether a cited paper has been retracted?
Search the article title or DOI in the publisher record and a trusted bibliographic database, and look for linked retraction, expression-of-concern or correction notices. PubMed labels and links many notices, while Crossmark may expose current status on publisher pages. Check close to submission because status can change.
Should a preprint be cited in a medical manuscript?
A relevant preprint can be cited when the journal permits it, but identify it clearly as a preprint and do not present it as peer reviewed. Check whether a journal version now exists and compare versions before replacing the record. Follow the journal's reference format.
Can reference-management software complete the audit?
No. It can detect duplicates, insert citations and apply a style, but it cannot reliably decide whether a source supports a claim, whether the best source was selected or whether a retracted article is being used appropriately. Imported metadata can also be incomplete. Human verification remains necessary.