Used by 1,000+ researchers and students
See the sentences your sources don't support
A reference that looks right can still be attached to a claim the paper never made. Grongy compares each sentence against the text of the source it cites and shows you the quote it relied on.

Your sentence is compared against the source's own text
The unit of the check is not a paragraph; it is one sentence and the source it cites. Code builds that pair — which sentence points at which record comes out of the document's structure. The model's job starts after that and stays inside one question: does this sentence hold in the text we pulled from the source?
The pair is built
A sentence is matched with the record it cites. The bibliography itself is left out of the check.
The source's own text is read
The pages closest to the claim are pulled from the source's indexed full text, together with the publication's own abstract.
The verdict rests on a quote
Each finding carries a verbatim quote from the source, with the page or the abstract named beside it. If no quote can be found, the verdict stays unverified.
When a result measured in four languages is written as “all languages”
Transfer learning improves translation quality for all low-resource languages (Zoph et al., 2016).
The source's own sentence, p. 1568:
Using our transfer learning method we improve baseline NMT models by an average of 5.6 BLEU on four low-resource language pairs.The paper measures four language pairs: Hausa, Turkish, Uzbek and Urdu into English. “All low-resource languages” goes past that measurement, and the part of the sentence the source does not carry is exactly those words. It is the kind of sentence the panel marks “Claims more than the source shows”, and the card files it under “Partially supported”.
Zoph, B., Yuret, D., May, J., & Knight, K. (2016). Transfer learning for low-resource neural machine translation. EMNLP, 1568–1575.
A summary never stands in for evidence. An AI-written summary of the source does not count in this check; the comparison runs against the source's own sentences. Where a source has neither indexed text nor an abstract, no verdict is given — the page says it could not be checked.
A correction can be offered for a flagged sentence. The suggested wording cannot be stronger or more certain than the original, and its citations stay as they are.
A run checks up to 120 sentence–source pairs. Beyond that, claims carrying a number, a cause or a trend go first, then a sample spread across the document — and the panel says how many were checked. The citation audit is an AI task and can draw on your usage allowance.
Each card carries one of five verdicts: supported, partly supported, unsupported, contradicted, unverified.
A finding also says what kind of problem it is. Beside each sentence sits the verdict its card falls under:
- Claims more than the source showsPartially supported
- The source says something differentPartially supported
- The source does not support thisUnsupported
- The source says the oppositeContradicted
- Not enough of this source is indexed to checkUnable to verify
What gets flagged on the record itself
This check looks at the record rather than the sentence, and it needs no usage allowance. You start the scan; findings are sorted by how much they matter.
Critical
- The DOI in the reference is not registered
- The DOI is registered but resolves to a different publication
- The paper has been retracted — the card carries a Retracted badge
- A citation in the text has nothing matching it in your library
Major
- The year, authors or title differ from the index entry
- The finding shows you the index's record so you can correct your own
Minor
- The reference was not found in Crossref, Semantic Scholar or arXiv
Not found is only a minor finding, and for books, reports and journals published outside English it is ordinary — those indexes simply do not hold most of them. The card says the record is not visible in these indexes. It does not say the source does not exist.
The retraction check runs through the record's DOI and refreshes weekly in the background when you open a document. A source without a DOI cannot be checked this way.
A scan looks up to 24 records at a time, the ones you have cited first. Findings stay on the card for 90 days.

The AI cites only what is in your library
These are not checks; they are rules that run while text is being generated, and they apply only to text the AI writes. They do not touch sentences you typed yourself.
A citation with no match does not enter the text
Every citation in generated text has to resolve to a record in your library, or to a search result — and a cited search result is added to your library. One that resolves to neither is removed; if the citation is the subject of the sentence, the sentence goes with it.
A page number has to be a page that exists
The AI cannot point at a page outside the source's indexed range. A number that falls outside it is dropped from the citation, though the citation itself stays.
A quotation is looked for in the source
A direct quotation of six words or more that cannot be found in the source's text loses its quotation marks. A sentence that is not a quotation is not presented as one.
Text too close to the source is blocked
A generated sentence that carries ten consecutive words from the source, or overlaps it by sixty percent, does not enter the document.
Bring the draft you already wrote
The check is not limited to text written in Grongy. Import a finished chapter as .docx and its citations and bibliography come into your project.
.docx only
The bibliography is read from the document's structure, so pasted plain text does not carry it.
The bibliography needs a real heading
A bibliography marked with Word's own heading style — References or Bibliography — is read. A line that is merely bold is not a heading.
Each line is looked up
Entries are matched in Crossref and DataCite. Matches enter your library with their full record; the rest are added as written.
An entry added as written is not a verified record. Source integrity still searches the indexes for those cards and flags the ones that cannot be found, carry an unregistered DOI, or resolve elsewhere. Because they have neither an abstract nor indexed text, the citation audit reports them as not checkable.
What this is not
Not found does not mean fabricated
When a source cannot be found in the indexes, nothing here calls it invented. Books, theses, institutional reports and journals published outside English are frequently absent. The card reports that the record is not visible; the judgment is yours.
Not a plagiarism report
Similarity against other publications is not measured, and the report your institution asks for is a separate thing this does not replace.
Not an AI detector
No verdict is produced about whether a text was written by a model, and no feature here makes text harder for a detector to flag — the terms of use prohibit using it that way.
Not a judgment on the source itself
The check asks whether your sentence holds in the source, not whether the source is right. A sentence citing a retracted paper accurately can come back supported — which is why the retraction finding sits on the record instead.
Questions about checking citations
ChatGPT gave me references. How do I check them?
Start with the record: does the DOI resolve, and does the index hold the same year, authors and title? Then check the claim: open the source and find the sentence your text leans on. The first step is mechanical and the second is the one that catches a real reference attached to a claim it never made.
What if my source has no full text?
If it has an abstract, the sentence is compared against that. With neither an abstract nor indexed text, no verdict is given: the panel reports how many citations could not be checked and asks you to add the PDF. Missing text is never filled in by guessing.
Does it catch fabricated sources?
It flags records that cannot be found and DOIs that are not registered, which is how a fabricated reference usually shows up. But it does not label anything a fabrication, because a missing record and an invented one look the same from the outside — and plenty of real books are missing.
Are my own sentences checked too?
The citation audit runs over the whole document, including what you wrote and what you imported. The rules that run while writing apply only to generated text: a page number in a citation you typed is not removed.
Does checking cost anything?
Source integrity does not draw on your usage allowance and works on the free plan. The citation audit is an AI task and can draw on it.
Related guides
- AI Citation Error Rates: Seven Studies in One Table
AI citation error rates run 6% to 91.4%, and that spread is not one scale: seven studies counted seven different kinds of miss, each with its sample.
Start with the draft in front of you
Import a chapter and check its citations. Source integrity runs on the free plan.