Reference checker: match every citation to your reference list with literal evidence
Send a PDF or DOCX and see which references are cited and where, which in-text citations have no entry in the list, which entries are never cited, and what needs fixing in the formatting.
Free credits · No credit card · Instant access What changed recently
What this solves
The Reference Checker reads your PDF or DOCX and pairs each entry in the reference list with the place it is cited in the body, showing the citation and the surrounding sentence. The pairing is done by fixed rules, not by a language model, so those matches come out the same every time. Only references the rules cannot resolve go to a conservative AI check.
How it works
-
Choose what you are sending
Pick full document to audit citations against the list, or reference list only to check and convert the formatting without a manuscript.
-
Send the file and pick the style
Upload a PDF or DOCX of up to 20 MB, with at most 300 references, and choose ABNT, APA 7, Vancouver, IEEE, or Chicago author date.
-
Read the report
The report lists cited references with evidence, orphan citations, uncited entries, format issues with a suggested fix, and the complete list rewritten in the chosen style.
Method
What happens to your file, step by step, and where the result stops being a decision made by a model.
How it works inside
PyMuPDF (PDF) or python-docx (DOCX) extracts the text, which is split at the last reference heading among 16 accepted spellings. The body is folded: NFKD accent stripping, lowercase, dashes and apostrophes collapsed to spaces, plus an index map back to the original. Matching is regex over that folded text: author surname and year inside an 80 character window, discarded unless the span has citation shape, or bracketed numbers for Vancouver and IEEE. Each hit is quoted verbatim through the index map.
Where AI is used, and where it is not
No model participates in the core matching. Citation to reference pairing, orphan detection and the uncited list are regex and set arithmetic over folded text, reproducible run to run. GPT-5.6-sol handles style detection, the style conversion and the per-reference issue list, which are model judgment. GPT-5.6-terra sees only the unmatched residue, at most 40 references, and its hit is discarded unless the quoted citation appears literally in the body and contains the year.
What it accepts
- PDF or DOCX up to 20 MB. The cap is checked on the file actually received, not only on the size the upload declares in advance.
- Two modes chosen before upload: full document (body plus reference list) or reference list only.
- Target style: APA 7, ABNT, Vancouver, IEEE or Chicago 17 author-date. Vancouver and IEEE switch the matcher to numeric brackets.
- Cap of 300 references per analysis. Above that the run stops before any charge and asks you to split the file.
- In full-document mode a reference heading must exist. 16 spellings in English and Portuguese are recognized, and the last occurrence wins.
What you get back
- Per cited reference, the report shows the literal in-text citation and the surrounding sentence from the original text. The saved result also records which rule matched and a fixed confidence (1.0 first author with year or numeric bracket, 0.9 institutional acronym or disambiguating co-author, 0.85 fallback co-author, 0.8 adjudicated).
- Four counted lists: references with a citation found, references with none, orphan citations (in the text with no entry in the list), and format issues per reference with a suggested rewrite.
- The whole list converted to the chosen style as plain text with a copy-all button. Numeric styles get the entry number prefixed.
- Written guidance: one summary, up to 6 prioritized recommendations and up to 8 style rules, in the interface language.
- The complete result as JSON, kept in your history and in a temporary file on the server. There is no DOCX or PDF export.
What this tool does not do
- It does not check that a reference exists and does not fill in missing metadata. No DOI is resolved and no catalogue is queried, so a fabricated but well formatted reference passes, and an entry with no volume or pages is reported as incomplete, not completed.
- When two references share the same first author and year, no first-author match is attempted. It resolves only through a co-author unique to one of them, or through the adjudicator, which reads just the first 60,000 characters of the body. Otherwise the reference is listed as uncited.
- PDF reference parsing is heuristic. Wrapped lines are rejoined by numbering and author-signature patterns, and lines under 10 characters are dropped, so an unusual PDF layout can split or merge entries.
- Reference-list-only mode reports no orphan citations and no uncited references. With no manuscript body there is nothing to compare against.
- It does not see the page layout. Claims about italics, bold, hanging indent, font and spacing are stripped from the issue list, because plain text extraction cannot prove them.
- Plan
- Pro
- Cost
- 30 credits
What you get
-
Matching without a language model
The core that links citations to references is deterministic: fixed rules over author, year, and numeric marks. Run the same file twice and these rule-based matches do not change; only the leftover goes to a conservative AI check.
-
Evidence you can check
Every match opens to show the citation exactly as it appears in your manuscript, plus the surrounding text, so you can confirm the finding instead of trusting it.
-
Both directions checked
The report separates citations with no entry in the list from entries that are never cited in the body, which are the two errors examiners find first.
-
The list converted to one style
Each reference is rewritten in the style you chose, with a Copy all button for the complete list. When you send only the reference list, a before and after shows each entry whose text changed.
Changelog
Every line below is a change that actually shipped, dated by the day it went out.
-
Latest
- The upload limit rose to 20 MB, and the page now shows the same limit the server enforces.
- An analysis run inside a project is saved into that project with your other results.
Show 5 earlier updates
-
- Oversized files and bibliographies over 300 references are rejected before any credit is charged.
- When the bibliography is too long, the message points to the reference-list mode instead of splitting the manuscript.
- The Recent analyses list shows whether each run used a full document or only a reference list.
-
- New mode: send a file containing only the reference list, with no manuscript.
- If the file does not match the chosen mode, the tool offers the other mode without charging.
- An analysis the AI never produced stopped consuming credits.
- Long DOIs and links stopped spilling out of the result cards, and errors reach screen readers.
-
- The citation-to-reference match became deterministic: the same document gives the same result, with the passage as evidence.
- The Crossref check was dropped: books, preprints and pre-DOI articles are no longer marked unverified.
- The format audit stopped flagging routine tweaks on every reference and explains the real corrections once, with examples.
- Sections titled Lista de Referências, Literature Cited and other variants stopped being refused as having no references.
-
- References extracted from a PDF stopped breaking apart when one entry spans several lines.
- Vancouver-style surnames such as Ong C. are read correctly, so those references stop being mismatched.
-
- The tool was created on 21 July 2026 behind a release flag: production only shows it when enabled.
- First version: send the manuscript and see citations without a reference and references never cited.
- Headings with accents and numbering (6. REFERÊNCIAS) are recognized, and the upload accepts drag and drop.
- Reference checking against Crossref runs in parallel: 25 references went from about 22 to 5 seconds.
Questions and answers
Is the citation matching accurate?
The matching is deterministic and shows the literal passage for each hit, so you can verify it yourself. It can still miss a citation that the PDF extraction broke apart, which is why the report always shows the evidence.
Does the AI decide anything?
Yes, but the main pairing is rule-based. AI detects the current style, rewrites the references, flags format problems, and writes the guidance. References the rules cannot resolve go to a conservative AI check that only accepts a literal passage containing the year.
What files and limits does it accept?
PDF or DOCX up to 20 MB, with up to 300 references per analysis. Above that, send the reference list on its own and split into parts. Each completed analysis consumes 30 credits.
Is my document used to train AI models?
The extracted text is sent to the AI provider's API for the formatting and guidance steps. We use paid API tiers, whose published policy states that content sent this way is not used to train models.
How to cite this tool
Used it in your research? Here is the reference, already filled in with the version you are looking at and today's access date.
ABNT (NBR 6023)
LESSA, P. W. B. Reference Checker. Versão 2026.08. [S. l.]: Xplore Dados, 2026. Disponível em: https://xploredados.com/en/tool/reference-checker. Acesso em: 21 ago. 2026.
APA 7
Lessa, P. W. B. (2026). Reference Checker (Version 2026.08) [Computer software]. Xplore Dados. https://xploredados.com/en/tool/reference-checker
BibTeX
@software{xploredados_reference_checker_2026,
author = {Lessa, Patrick Wendell Barbosa},
title = {Reference Checker},
organization = {Xplore Dados},
version = {2026.08},
year = {2026},
url = {https://xploredados.com/en/tool/reference-checker},
urldate = {2026-08-21}
}
Start with the free credits
Create an account and test the tools before deciding on a plan.
Create free account