Bibliography, References or Works Cited: Which List Your Style Wants

The three headings sit in the same place, at the end of your paper, and hold the same kind of entries. They differ on one question: may the list contain works you read but never cited? A References list (APA) and a Works Cited list (MLA) hold only the sources you actually cited in the text. A Bibliography, in the Chicago notes-and-bibliography sense and in most thesis regulations, may also hold works you consulted. And you rarely get to pick. The heading arrives with the style your department requires, so the practical question is not which word you prefer but which rule that word switches on.

The question that separates them: did you cite it?

A References or Works Cited list is a one-to-one map. Every entry has a matching in-text citation, and every in-text citation has an entry, with no orphans on either side. A Bibliography relaxes one direction: it may carry background reading that never got an in-text pointer, while still requiring that everything you did cite appears.

This is not a matter of taste that citation software ignores. It is wired into the software. Tezpur University's thesis bibliography style has a section headed "Inclusion of non-cited documents" and tells authors to insert \nocite{<identifier 1>,<identifier 2>,...} somewhere in the thesis to push those works into the list (p. 15). The R package RefManageR exposes exactly the same idea as a function, NoCite(), sitting in the documentation right next to Citet() and Citep() (McLean, 2014, p. 15). Tools ship a dedicated switch for "in the list, not in the text" because the distinction is real and authors need it.

Before — an APA-formatted paper whose author treated the list as a reading log:

Bibliography
Author, A. A. (2019). Title of the article. Journal Name.
Author, B. B. (2021). Title of the book. Publisher.
Author, C. C. (2020). Title of the article. Journal Name.   <- never cited

After — two decisions applied, the heading and the membership rule:

References
Author, A. A. (2019). Title of the article. Journal Name.
Author, B. B. (2021). Title of the book. Publisher.

The third entry is not deleted from your notes. It is deleted from the list, because in APA the list documents citations, not effort.

Which heading your style expects

• APA 7: References
• MLA 9: Works Cited
• Chicago, notes and bibliography: Bibliography
• Chicago, author–date: References
• IEEE and Vancouver (numbered styles): References
• Harvard variants: usually References, though individual university guides differ

One caveat that overrides everything above: many faculties publish their own thesis regulations, and those regulations often say "Bibliography" regardless of the style they otherwise mandate. Follow the department's heading. If you are still choosing between styles rather than following one, how to choose a citation style covers that decision, and the entry-level formatting lives in the APA 7 guide and the MLA guide.

What changes inside the list, not just above it

Picking a list also picks a citation mode, and mixing modes is the error a supervisor spots in seconds. Tezpur's guide sets the two out side by side: author–year gives "Rogers (2010) defined ..." and "It was shown by Sarma and Singh (2014)", while numbering mode gives "Rogers [15] defined ..." for the same sentence (p. 2). Repeat works by one author in author–year mode take a letter appended to the date, so the same author's two 2005 papers become 2005a and 2005b, both in the text and in the list (p. 3).

Numbered styles then compress. The biblatex manual describes a numeric-comp style that prints runs of consecutive numbers as ranges, turning "[8, 3, 1, 7, 2]" into "[1–3, 7, 8]", sorting the citations to get there (p. 75). Ordering inside the list is a separate setting again: RefManageR documents sorting schemes named "nty", "nyt", "nyvt", "ynt" and "ydnt" (McLean, 2014, p. 12), which are name–title–year, name–year–title, year–name–title and so on. A Bibliography that carries consulted works still sorts them by the same rule as the cited ones, so uncited background reading is interleaved alphabetically rather than grouped into a section of its own.

Page numbers have their own conventions inside the list. Tezpur compresses a journal article to <volume>(<issue>):<start>–<end>, giving 26:227 for a single page or 12(6):405–419 for a range, writes non-consecutive pages as 18,25–29,32, and marks an uncertain end point as 57+ (p. 8).

Before: vol. 12, no. 6, pages 405 to 419 After: 12(6):405–419

The failure that actually costs marks

Markers are rarely lost over the heading. They are lost over entries a reader cannot follow back to anything.

Klein et al. (2014) measured this for web links in scholarly articles and concluded that for a large majority of the science, technology and medicine articles in the period they studied, it is impossible to revisit the context created by the web resources those articles reference (p. 26). The same paper summarises earlier work: Dellavalle et al. found link rot in high-impact journals rising from 3.8% three months after publication to 13% after 27 months (p. 6). That finding is about links to the open web rather than journal articles with a DOI, which is the practical lesson: prefer the identifier over the URL when a source has one.

The newer failure is entries typed by a language model. Szeider (2025) collects the measurements: Agrawal et al. tested GPT-3, ChatGPT and GPT-4 and saw citation error rates fall from over 70% to under 50% across those models, while Kim et al. found fabrication rates above 80% for certain publication categories, producing plausible but nonexistent authors, venues and DOIs (p. 2). Szeider's own system, which fetches entries straight from a bibliographic database instead of letting the model retype them, reached an 82.7% perfect match rate against 28.2% for standard web search, with zero cases of metadata corruption (p. 3). The architecture, not the prompt, is what makes the difference: an entry copied from a database is right, and an entry recalled by a model is a guess.

That is the same principle behind our bibliography generator, which builds the whole list from real identifiers, and the citation generator for a single entry. The academic sources cited in this article were themselves found through cytado's own corpus, with the page numbers taken from the printed pages.