BibTeX, RIS, and CSL: Exporting References Without Silently Losing Data

Every reference tool exports to BibTeX, RIS, and CSL-JSON, which suggests the formats are interchangeable. They are not. Each has a different data model, and round-tripping between them loses fields that a bibliography depends on. Knowing what each format preserves prevents the familiar discovery, three days before submission, that half the author lists are truncated or the page ranges have vanished.

What each format is for

BibTeX is a LaTeX-native format designed around entry types (article, inproceedings, book) and a fixed field vocabulary. It handles math in titles with braces and ties citations to \cite{} keys. It has no standard place for arbitrary metadata, which is why tools stuff extra data into the note or annote field.

RIS is a line-based format used by reference managers and many publisher sites. It is more permissive than BibTeX and maps cleanly onto book chapters, conference papers, and reports. Its TY type codes are extensive, though support varies by importer.

CSL-JSON is the structured format behind citation styles. It models names as objects, dates as structured parts, and supports rich types. It is the best format for programmatic processing and the one Zotero, Pandoc, and many APIs use internally.

Fields that commonly disappear, and verifying the export

Author name particles and suffixes. "van der Berg, J." and "Smith, Jr., J." can lose particles or suffixes when converters flatten names. BibTeX's and-separated name syntax has specific brace rules for corporate authors that many exporters ignore.

Page ranges. BibTeX stores pages = {123--145}; RIS uses SP and EP. Converters sometimes collapse to a single page or drop the en dash format.

Dates. BibTeX has year, month, and sometimes date. RIS has PY, DA, and Y1. CSL-JSON has structured issued parts. Online-first versus issue dates are a frequent casualty.

DOIs and URLs. RIS has DO and UR; BibTeX needs doi and url fields, which older .bst styles ignore. A bibliography that silently omits DOIs is hard to audit.

Abstracts and keywords. Often dropped by BibTeX exporters that target traditional styles, even though the data exists.

Do not trust the exporter's success message. Check three things.

First, count entries in the exported file and compare with the selection. A mismatch usually means malformed entries were skipped.

Second, scan for empty required fields: author, title, year, and container. A converter that fails on a corporate author often leaves author = {} and continues.

Third, spot-check five entries against the source records, including one with a multi-word family name and one with a page range. Those two catch most systematic loss.

For LaTeX workflows, compile the document with the exported .bib and confirm every citation resolves. biber warnings about missing fields are easier to fix before writing than after.

Round-tripping, version control, and a format decision rule

Treat the reference library as the source of truth and exported .bib files as build artifacts. Regenerate them from the library rather than editing them by hand; hand edits are lost on the next export.

If the project is shared, keep the .bib in version control and diff it when the library changes. A diff that shows fifty changed author fields after a minor library update is a signal that the exporter is rewriting names.

Use CSL-JSON for processing and APIs, BibTeX for LaTeX submissions, and RIS for tools that accept nothing else. Store the canonical record in one place, export at the end, and verify the specific fields your target style requires. The format is a transport, not the archive.

Character encoding, and corporate or literal names

BibTeX is not Unicode-native in older toolchains. Accented names, en dashes, and Greek letters in titles must be escaped or handled by a Unicode-aware engine. biber with utf8 input handles most cases; classic bibtex may not.

Test with a name containing a diacritic, a title containing an en dash, and an abstract with a Greek symbol. Those three reveal whether the toolchain will mangle your bibliography at compile time.

A corporate author must be wrapped in braces in BibTeX: author = {{World Health Organization}}. Without braces, BibTeX parses it as a person with first name "World" and last name "Health Organization". CSL-JSON has literal for this; converters sometimes drop it.

Check one corporate author per export. It is the most common silent corruption in a converted file.

Entry types that map badly, and validating with the real style

A dataset, software, or preprint has no first-class BibTeX type. Tools map them to misc, which loses the type information a reader needs. Use @misc with howpublished and a note describing the type, or use biblatex which has richer types.

Conference papers are another mismatch: @inproceedings expects a book title for the proceedings, and many exports leave it empty.

Compile a short document that cites one of each entry type in the target style. Reading the rendered bibliography catches errors no schema validator sees: missing "and" before the last author, wrong page-range separator, a title in title case that should be sentence case.

Do this once per style rather than per entry. The style's quirks are consistent.

Version control, spot checks, and field loss in translation

Commit the .bib file and a generated snapshot of the rendered bibliography with each submission. The snapshot is the evidence of what was submitted; the .bib is what you can rebuild from.

If the library changes after submission, the diff shows exactly what changed. Without both files, a revision request about a citation becomes an archaeology project.

After every export, open the file and search for empty braces, author = {}, and year = {}. These indicate a field the exporter could not map.

Compile one citation per entry type before submitting; the style, not the schema, reveals the remaining gaps.

RIS has no native field for eprint, archiveprefix, or keywords in the way BibTeX does. ArXiv identifiers disappear silently on export.

Round-trip a test record through both formats and diff the fields. The diff is your list of what needs a note or a custom field.

Keep a small golden file of ten entries that cover the awkward cases: a corporate author, a preprint with an eprint field, a paper with a page range and an article number, and a name with a particle. Run every export format against it and diff the result. The golden file turns a class of silent metadata bugs into a visible diff, and it takes an hour to build once.