CiteTrue vs ZeroGPT

ZeroGPT tries to answer "was this text written by AI?". CiteTrue answers a narrower and much more checkable question: "do these references exist, and do they say what the text claims?" The second question has a right answer you can verify by opening a link; the first does not.

Why our verdicts hold up

Most citation checkers hand the whole judgement to a language model and show you what it says. Models hallucinate — that is the very problem this category exists to solve — so we put the model last, not first.

  1. 01

    1. Parse, don’t guess

    Every reference is split into its parts — authors, title, year, journal, DOI — by a parser built for citation formats, not by asking a model what it thinks the title is.

  2. 02

    2. Look it up in the authorities

    The parsed reference is queried against Crossref and OpenAlex, and against the open web when a reference has no DOI. What comes back are real records with real identifiers.

  3. 03

    3. Score the match deterministically

    Candidate records are scored with a Sørensen-Dice similarity over title tokens. A record counts as a match only if the title score clears 0.7 and the author list agrees — a year mismatch is reported as a notice, never quietly accepted. The same reference always produces the same score.

  4. 04

    4. Only then, ask a model

    A language model is called only when the deterministic pass finds no authentic match — to read the candidates and the source text and decide whether anything fits. It is the fallback, not the judge. That ordering is why a hallucinated verdict has nowhere to enter: by the time a model is involved, the evidence is already on the table.

The practical consequence: a reference marked authentic here matched a real record on title and authors, and you can open that record. It is not a model's impression of whether the citation looks plausible.

How fast it gets through a bibliography

Checking one reference means several lookups across different sources. Doing that one reference at a time is what makes a tool feel slow on a long bibliography, so we run them in parallel and stream the results back.

References checked in parallel
20 at a time in the app, 100 through the API
A 50-reference bibliography
Goes out as three waves of parallel lookups, not fifty queued ones
Results as they land
Streamed per reference — you read the first verdicts while the rest are still running
Bulk work
A supervisor’s stack of theses runs through the API at five times the app’s concurrency

Different jobs, compared

What you are comparingCiteTrueZeroGPT
The question it answersDo these references exist, and do they support the claim?Does this text look AI-written?
Can the answer be checked?Yes — every verdict links to the record it is based onNo — a probability score with nothing behind it to open
False accusationsA failed check names the reference and the reason, so it can be argued withDetectors are known to flag human writing, especially by non-native speakers
Use in an integrity caseEvidence you can put in front of the authorWidely disputed as sole evidence
CostFree for 30 a day, then from $5.99/moFree tier with paid plans
ConveniencePaste the bibliography; results per referencePaste the text; one score for the whole thing

So which should you use?

If you are trying to establish whether a document is trustworthy, fabricated citations are the hard evidence and a detector score is not. Check the references first — that finding stands up in a conversation with the author, and a probability score does not.

Check your references now

30 references a day, free, no card. Paste a bibliography and see for yourself.

Other comparisons