Bibliography Maker That Checks Sources Exist, Not Just Formats Them

A bibliography maker has two jobs, and most of them only do one. The first job is formatting: putting author, year, title, volume and pages in the order APA, MLA or Chicago wants, with the right italics and the right punctuation. The second job is verification: confirming that the work described in the entry exists, and that the details you typed match the real record. Skip the second and you can hand in a reference list that is immaculate on the surface and still points at a paper nobody ever wrote.

If you are assembling a bibliography right now and want the short version: paste your list into a tool that looks each entry up in a live bibliographic database, not one that only reformats the text you gave it. cytado's bibliography generator is built around that second step, and the sources cited in this article were found through cytado's own corpus rather than from memory.

Formatting correctness tells you nothing about existence

The clearest evidence for treating these as separate problems comes from a study of 636 references produced by ChatGPT across 42 topics. Walters and Wilder (2023) found that 55% of the GPT-3.5 citations and 18% of the GPT-4 citations were fabricated, describing works that do not exist (p. 1). In the same dataset, every single citation was in APA format (p. 5). A reference list can score 100% on style compliance and roughly half of it can still be fiction.

The errors do not stop at invented works. Among the GPT-3.5 citations that pointed to real papers, 43% carried at least one substantive error: wrong author names, wrong titles, wrong dates, wrong journal, or wrong volume, issue and page numbers. More than a third had incorrect volume, issue or page numbers, and 22% had the wrong year (p. 4). Numeric fields are where things break, which is unfortunate, because numeric fields are exactly what a marker checks when they go looking for your quotation on the page you claimed.

There is a second-order risk worth knowing about. Walters and Wilder note that at least two of the fabricated citations in their sample named journals whose publishers have been identified as predatory (p. 6). A fake reference can pull a reader toward a real but disreputable venue.

What an existence check actually does

"Verified" gets used loosely, so it helps to be concrete. A real check is three separate comparisons, run against an authoritative index such as Crossref, PubMed, OpenAlex or DOAJ:

1. Does a record exist for this title and author combination?
2. Does the metadata match what your entry claims: year, journal, volume, issue, page range?
3. Does the identifier resolve to that same record, rather than to some other paper?

Point 2 is the one people skip. An entry can name a real paper and still get the year wrong, and a checker that stops at "yes, something with this title exists" will wave it through.

The pattern is starting to appear in reference management software. Jin et al. (2026) reviewed conventional tools against newer AI-based ones and describe validation systems that cross-check citation metadata against an authoritative database, flagging both non-existent references and inconsistencies in details such as publication year or author information, with alerts raised at the moment the reference is entered. They characterise this verification layer as absent from conventional reference management software (p. 5), which has historically optimised for library organisation and for formatting across thousands of styles (p. 2).

There is an architectural point underneath this. In a system for AI-assisted citation presented in 2025, Szeider describes bypassing the language model at the export step entirely, fetching each entry straight from the source database so that bibliographic data reaches the file without passing through the model's context; when a record key is invalid, the fetch fails and reports an error instead of producing a plausible-looking entry (p. 5). Whatever generates your draft, the bytes in your final bibliography should come from a database, not from something reconstructing them.

Li (2018) made a related argument years earlier for LaTeX users: use the DOI itself as the citation key, and pull the BibTeX entry from a DOI-to-BibTeX service, so the only thing you type by hand is an identifier. Manual retyping is where volume numbers get transposed.

Before and after

Here is an entry of the kind an AI assistant will happily produce. It is well-formed APA, and it is wrong.

Before:

Walters, W. H., & Wilder, E. I. (2022). Fabricated citations in AI-generated literature reviews. Journal of Academic Librarianship, 48(4), 102–115. https://doi.org/10.1016/j.acalib.2022.102534

Nothing about that looks suspicious. The authors are real, the journal is real, the formatting is clean, the DOI has the right shape. Every retrievable field is nevertheless wrong.

After (the actual record):

Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, 14045. https://doi.org/10.1038/s41598-023-41032-5

Different title, different journal, different year, different DOI. A style checker sees no difference between the two, because both satisfy APA. A lookup against Crossref separates them in about a second. If you would rather do this by hand for a short list, we wrote up the manual procedure in how to check if a source really exists.

Where the check runs out of road

An existence check is not a truth check, and you should know its two limits before you trust it blindly.

First, the registries themselves contain errors. Massari et al. (2023), describing the construction of OpenCitations Meta, note that Crossref does not double-check the metadata publishers submit, so mistakes are preserved downstream; they cite a record whose metadata claims print publication in 2029 (p. 24). Verification confirms that a record exists and that your entry matches it. It cannot confirm the registry is right.

Second, coverage is uneven by discipline. The same paper cites research finding that close to 90% of publications in the sciences and social sciences carry a DOI, against roughly 50% in the arts and humanities (p. 3). If you are writing on a historical or literary topic, expect a share of your list, particularly older books and primary sources, to sit outside DOI-based checking entirely and need a library catalogue instead.

Neither limit is a reason to skip the check. Both are reasons to read the output rather than assume a green tick settles the question.

A five-minute workflow before you submit

1. Export your reference list as plain text.
2. Run it through a checker that queries live databases, entry by entry.
3. Read the mismatches rather than auto-accepting fixes. A flagged year is usually a genuine error in your entry; occasionally it is the registry.
4. For anything with no DOI, search your library catalogue by title and author, and record the ISBN.
5. Confirm that your in-text page numbers point at pages the record says exist. A quotation on p. 340 of a 212-page book is the kind of thing that gets noticed.

If you are unsure which list your style even wants, bibliography, references or works cited covers the distinction. Get that right second. Get existence right first.