A finished paper can read well and still carry a reference that fails the first time a marker clicks on it. One entry was never published. One has the right title and the wrong year. One was withdrawn by its journal after you saved the PDF. One is cited in the text and missing from the list, because its source was cut in the last edit.
To check every reference in your paper before you submit, you need lookups, not judgement: a reference is in the publisher's record or it is not. This guide from EdCitation sets out the four ways references fail and what the research shows about each, then the check by hand, then what came back when we uploaded the site's own sample paper to References from a file, the Pro tool that checks every reference in a paper and matches every citation to the list. The same lookups are free, with no account, in Verify references; the difference is set out near the end.
Why do the references in a finished paper fail?
The references in a finished paper fail in four ways: never published, cited with the wrong details, withdrawn by the journal, or unpaired between the text and the list. Only the first is mainly an AI problem.
A reference nobody published
The fabricated reference is the hardest to spot by eye, because nothing in its format gives it away. Walters and Wilder (2023) had two versions of ChatGPT write 84 short literature reviews on 42 topics, of the kind set in first-year composition courses at US universities, and checked all 636 references. GPT-3.5 had fabricated 55% of its references and GPT-4 18%. Even GPT-4 fabricated 70% of the book chapters it cited.
A fabricated entry can also arrive second-hand, in notes or a classmate's shared list. How each chatbot goes wrong is in ChatGPT fake citations and how to find and fix them.
A real paper with the wrong details
The commoner failure is a real paper cited wrongly. In the same study, 43% of the real references from GPT-3.5 and 24% of those from GPT-4 carried a substantive error, most often a wrong volume, issue, page range or year. People make the same slips unaided. Azadeh and Vaez (2013) checked 704 journal references in PhD theses from two Iranian universities of medical sciences against six elements (authors' names, article title, journal title, year, volume and pages) and found errors in more than half of them at each university. A DOI with one digit wrong either fails to open or opens somebody else's paper.
A paper its journal has withdrawn
A retracted paper can be real, correctly cited and still unusable, and citations keep arriving after the notice. Retraction Watch (2026) lists the most cited retracted papers in Web of Science: a paper on visfatin, published in Science in 2005 and retracted in 2007, had 232 citing articles before its retraction and 1,351 after it.
Retraction Watch (2024) calls its database the most comprehensive source on retractions, while warning that it does not try to be complete for corrections or expressions of concern. What each notice means is in what a retraction is and how to check your reference list for one.
A citation with no entry, or an entry nobody cites
The fourth failure sits between the text and the list. APA's rule is that each work cited in the text appears in the reference list and each work in the list is cited in the text; personal communications are the exception, cited in the text only (Augusta University Libraries, 2026). The pairs come apart in the final week: a paragraph is cut and its source stays in the list, or a year is corrected in the text and not in the list, so a citation to Lee, 2019 points at an entry for Lee, 2020.
| Failure | What it looks like | How to catch it by hand | What References from a file reports |
|---|---|---|---|
| Never published | Plausible authors, title and journal; the DOI fails or opens another paper | Search the exact title in Crossref, PubMed or Google Scholar | Not found |
| Wrong details | The right paper with the wrong year, pages or DOI | Compare every field with the publisher's page | "Check this", when a record is found that the entry does not fully match |
| Retracted | A notice on the publisher's page, none on your saved PDF | The live page, or the Retraction Watch Database | Flagged as retracted |
| Cited but not listed, or listed but not cited | A citation with no entry, or an entry nobody cites | Tick each citation against the list, then each entry against the text | Each mismatch, at the place it happens |
How do I check the references in my paper by hand?
Check the references in two directions: from the text to the list, and from each entry out to its record. For each reference, in this order:
- Pair every citation with an entry. Search the file for each surname in the list and confirm that the spelling and year in the citation agree with the entry. An entry whose surname never appears in the text is uncited; a citation you cannot pair has no entry.
- Follow the DOI. Put it after
https://doi.org/and compare the title and authors on the page that opens with your entry. - Search the exact title when there is no DOI, in quotation marks, in Crossref, PubMed or Google Scholar. Look a book up by its ISBN in a library catalogue.
- Check the paper's status. Look for a notice on the publisher's live page, or search the Retraction Watch Database, which takes a DOI, a PubMed ID or a title (Retraction Watch, 2024).
- Read the passage you cite it for. Baethge and Jergas (2025) pooled 46 studies covering about 32,000 quotations in medicine and found that 16.9% cited a source that did not support the authors' claim, 8.0% seriously so, with no improvement over time.
Steps 1 to 4 are lookups, and EdCitation's References from a file makes all four for every entry from one upload. Step 5 is reading, and no reference checker does it, ours included. By hand, each reference takes a few minutes, so a long list takes an evening; the innocent reasons a DOI can fail are in how to check whether a reference is real.
What does References from a file do with an uploaded paper?
References from a file reads the reference list out of your paper, looks every entry up in the publisher's record, matches every in-text citation to the list, and for a reference that cannot stand recommends published sources in its place. It is part of EdCitation Pro and runs in the dashboard.
It is the best tool for this job because it looks every reference up and never writes one. A chatbot asked whether its own reference is real writes another answer, and Walters and Wilder (2023), drawing on five earlier studies, note that ChatGPT often answers wrongly when asked to confirm the works it cites. A lookup can only report what the record holds, or that an index did not answer, and References from a file never reports the second as the first.
What you upload
The paper itself, as a Word document or a PDF. It is kept in My papers, so when you have fixed a reference and run the check again, the new report shows what you fixed since the last run.
What comes back
- A verdict on every reference: verified, doubtful (shown as "check this"), or not found in any publisher's, registry's or library's record.
- Retractions and corrections: a withdrawn paper is flagged before your marker finds it, and a published correction is linked.
- Citations matched to the list: a source cited and not listed, or listed and never cited, is pointed out where it happens.
- A source for every gap: for a reference that cannot stand, published sources that fit are recommended, each a real record you can cite in a click.
- The finished list: in APA, MLA, Chicago, Harvard, IEEE or Vancouver, ready to paste with italics and hanging indents.
What it costs
Pro is $8 a month, or $72 for a year paid at once, which saves 25%. A free account includes three checks a month of References from a file, and Pro has no limit. The plans are compared on the pricing page.
What came back when we uploaded the sample paper?
Four references came back verified and one not found, all five citations were paired with their entries, and nothing was retracted. We ran the check on 23 September 2026, in APA, on the sample paper behind the "Verify the sample paper" button on Verify references, so you can repeat it. It is a four-page APA student essay on whether nurse-to-patient ratios protect patients, with 221 words of body text and five references.
The five references
| Entry in the sample | Verdict | Record it matched | Place in the file |
|---|---|---|---|
| Aiken et al. (2014), The Lancet | Verified | DOI registry | Paragraph 17 |
| Hall et al. (2016), PLOS ONE | Verified | Crossref | Paragraph 18 |
| Kane et al. (2007), Medical Care | Verified | DOI registry | Paragraph 19 |
| Needleman et al. (2011), New England Journal of Medicine | Verified | Crossref | Paragraph 20 |
| Thompson and Alvarez (2021), Journal of Hospital Workforce Studies | Not found | None | Paragraph 21 |
Each verified entry came with the same reason: "The DOI resolves to this record and the title matches." No entry was marked "check this", retracted or could not check.
The citations, matched to the list
Every citation had an entry, and every entry was cited. The report's summary line read "5 in-text citations · 5 entries · 0 orphan citations · 0 uncited entries", an orphan being a citation with no entry behind it. In a long dissertation, the place given with each finding takes you straight to the paragraph.
The reference that cannot stand
Thompson and Alvarez (2021) looks as complete as its neighbours: two authors, a year, a title, a journal, a volume, an issue and pages. It has no DOI, and the report's reason was plain: "No publisher's or registry's record matches this reference." We then checked by hand, as step 3 says: Crossref's search for the exact title returned only unrelated works, and Crossref's list of journals held no title of that name.
The entry carries the paper's opening claim, that the nursing shortage "is expected to deepen over the coming decade", and is the only reference behind it. Deleting the entry leaves that sentence unsupported and its citation an orphan, so the fix is a real source, not a deletion.
This is where References from a file goes further than a verdict: for a reference that cannot stand, it recommends published sources that fit, so the gap is filled from the record and not from memory. Read one before you cite it; if it says something different about the shortage, the sentence changes. Our run was of the check alone, so the recommendations are described here from the tool's own page rather than shown. To search yourself, EdCitation's free Find sources takes the sentence and returns published work that could support it, and Cite a source builds the new entry from its DOI.
How do I read each verdict?
Each verdict reports what the record showed, and each has its own next step.
| Verdict | What it means | What to do |
|---|---|---|
| Verified | A record matches the entry | Read the source for the claim you cite it for |
| "Check this" | A record was found, but the entry does not fully agree with it | Compare them field by field and correct the entry; Cite a source rebuilds it from the DOI |
| Not found | No publisher's or registry's record matches it, nor, for a book, a library catalogue's | Replace a recent journal article; check a report, thesis or older book at its source |
| Could not check | An index did not answer | Run the check again, or check by hand |
| Retracted | The journal has withdrawn the paper | Replace it, unless the retraction is your subject |
Why "could not check" is kept apart from "not found"
A checker that reports silence as absence frightens students about real sources. When an index does not answer, EdCitation says "could not check", never "not found". Grey literature needs the same care: a government report or a thesis can be missing from the indexes and still exist, so look for it on the organisation's site or in the repository before you remove it.
Verify references or References from a file: which do I need?
Use Verify references for a single check before a deadline, and References from a file for a paper you will revise and check again. Both look every reference up in the publisher's record, flag retracted papers, never show "could not check" as "not found", and, with the whole paper uploaded, match every in-text citation to the list and every entry back to the text.
| Verify references | References from a file | |
|---|---|---|
| Where it runs | On the site, with no account | In the dashboard, with an account |
| What you give it | Your paper, or a pasted reference list | Your paper, as a Word document or a PDF |
| After the check | Fix the entries and upload again | The paper stays in My papers, and the next check shows what you fixed |
| A reference that cannot stand | Search for support in Find sources | Published sources recommended in its place |
| Cost | Free | Pro, $8 a month; three checks a month with a free account |
The rest of Pro checks the rest of the paper: Mechanics QA checks formatting against the APA 7 student paper and finds comments and tracked changes left in the file; Check plagiarism marks passages that match published or web text; Check AI labels each passage AI-written, AI-assisted or human, and no score is proof of anything. Free on the site, Check your paper reads your assignment's instructions into a checklist, including how many sources are required. EdCitation never writes any part of your paper.
Quick questions
Can I upload my paper to check that my references are real?
Yes. EdCitation's free Verify references takes the upload, looks every reference up in the publisher's record and matches the in-text citations to the list. References from a file, part of Pro, also keeps the paper for the next check.
Does a reference checker tell me whether a source supports my sentence?
No. A reference checker confirms that a source exists, is cited correctly and has not been withdrawn; only reading it shows whether it supports your claim. Baethge and Jergas (2025) found that 16.9% of the quotations studied in medicine did not.
What does "not found" mean for a reference?
It means no publisher, registry or library has a record of the entry. For a recent journal article that is strong evidence it does not exist; for a report or a thesis, check at the source first.
Is a retracted paper flagged even when my reference is correct?
Yes. A retraction is a fact about the paper, not about your entry, so a correctly cited paper is still flagged. Replace it unless the retraction is your subject.
How much does References from a file cost?
References from a file is part of Pro, at $8 a month or $72 a year. A free EdCitation account includes three checks a month, and Pro has no limit.
References
- Augusta University Libraries. (2026, May 18). In-text citations. APA 7th citation guide. https://guides.augusta.edu/APA/in-text
- Azadeh, F., & Vaez, R. (2013). The accuracy of references in PhD theses: A case study. Health Information & Libraries Journal, 30(3), 232-240. https://doi.org/10.1111/hir.12026
- Baethge, C., & Jergas, H. (2025). Systematic review and meta-analysis of quotation inaccuracy in medicine. Research Integrity and Peer Review, 10, Article 13. https://doi.org/10.1186/s41073-025-00173-z
- Retraction Watch. (2024, October 23). Retraction Watch Database user guide. The Center for Scientific Integrity. https://retractionwatch.com/retraction-watch-database-user-guide/
- Retraction Watch. (2026, July 27). Top 10 most highly cited retracted papers. The Center for Scientific Integrity. https://retractionwatch.com/the-retraction-watch-leaderboard/top-10-most-highly-cited-retracted-papers/
- Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, Article 14045. https://doi.org/10.1038/s41598-023-41032-5