AI Citation Checker
Paste the draft. Every DOI, link and Author (Year) gets resolved, read, and checked against the sentence that cites it.
A citation fails three ways: the source does not exist, it exists but belongs to someone else, or it exists and does not say that. A DOI validator catches the first. This check resolves each reference through Crossref and the live page, then a model reads the source and quotes the passage that supports or contradicts your sentence. Free, once a day, on one model. Paid plans put four models on the same sources and show you where they split.
Single · free · 1 a day — Council and Deep on any paid plan
Looking each one up on Crossref and the live page, then reading the source. Ten to twenty seconds.
Deep takes a minute or two: four pro models answer, two critics review, one mediator writes the final call. This page updates when it lands.
Citation Checker runs the moment you sign in, on the text you just pasted. Nothing to enter twice.
A saved check is kept for 24 hours.
You've used today's one-model runs for this check.
Or come back tomorrow for another one-model run.
Why a DOI That Resolves Is Not Enough
The 2026 pattern is not an invented reference. It is a real identifier attached to a claim the paper never makes. Clicking the link tells you the paper exists. It does not tell you it says that.
Extract every reference
Each DOI and link is pulled from the text, along with any Author (Year) citation, together with the sentence that leans on it.
Resolve it for real
DOIs go to Crossref and links get fetched; an author-year reference is searched by author and year, then matched on title. A DOI nobody registered or a page that 404s is nonexistent. A real paper under the wrong name or year is misattributed. Neither verdict comes from a model.
Read the source against your sentence
For every source that resolves, the model reads the text and quotes the passage that supports or contradicts what you wrote. Supported needs a quote. No quote, no pass.
We built this because the draft that cites a real paper for a claim it never made is the failure a link checker cannot see, and the one that ends up in a correction. On our eight-draft fixture set the lookup alone catches every nonexistent and misattributed reference; the reading step is where the misrepresented ones fall.
This check reads the references. To check the claims between them as well, run the draft through the AI fact checker, which runs the claim check and this one together.
Example Check
Council run, 21 August 2026 · GPT, Claude, Gemini and Grok · 11 seconds · all four agreed on every reference
"Three sources anchor this piece. Vaswani et al. (2017) introduced the Transformer, an architecture based solely on attention mechanisms (https://arxiv.org/abs/1706.03762). A 2023 Nature study found that 80% of coders use AI daily (https://doi.org/10.1038/s41586-023-99999-1). Ioannidis (2005) proved that 80% of medical studies are wrong (doi:10.1371/journal.pmed.0020124)."
One reference is fully supported. One does not exist. One is misrepresented: the draft claims Ioannidis proved 80% of medical studies are wrong, but the paper argues that under certain conditions research findings are more likely false than true. A probabilistic framework, not a proven percentage.
Attention Is All You Need
SOURCE SAYS: “We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely.”
The source proposes the Transformer and describes it as based solely on attention mechanisms. The 2017 attribution matches the record.
Crossref has no record of this DOI. The study and its 80% figure have nothing behind them.
Why Most Published Research Findings Are False
SOURCE SAYS: “It can be proven that most claimed research findings are false. Simulations show that for most study designs and settings, it is more likely for a research claim to be false than true.”
The paper argues that most findings are likely false under certain conditions. It never gives 80%, never restricts itself to medical studies, and "proved" is not what a probabilistic model does.
Frequently Asked Questions
It finds every reference in your text and resolves each one through Crossref and the live page, then has a model read the source and quote the passage that supports or contradicts the sentence citing it. Each reference comes back as supported, misrepresented, misattributed, nonexistent or unverifiable.
It includes one. A DOI validator tells you whether the identifier resolves. That catches the invented reference and misses the more common 2026 failure, a real paper cited for something it does not say. This check resolves the DOI and then reads the paper.
Paste the reference. A DOI goes to Crossref, a link is fetched, an author and year are searched. A DOI nobody registered or a page that 404s is reported as nonexistent. A real paper whose author or year does not match what you wrote is reported as misattributed. Both are decided by the lookup, not by a model's opinion.
When the quote sits in a sentence that cites a source, yes. The model looks for the quoted words in the source text and reports misrepresented when they are not there or say something else. A quote with no source attached has nothing to be checked against; add the link.
The source exists but its text could not be read from here, usually a paywall or a site that blocks automated fetches. We say so instead of guessing. Paste the relevant passage into the draft and the check can read that instead.
It resolves links and DOIs and reads any source page it can fetch, so a brief that links to its authorities works. Case citations in reporter format are not resolved yet; those still need a legal database.
Yes. Paste the answer with links to the policy pages or documents it should draw from, and the check reads each one against the sentence that cites it. Private documents need a reachable URL; a paywalled or internal page comes back unverifiable.
A model asked to vouch for a reference tends to vouch for it. Here the existence and attribution checks are lookups, not opinions, and the reading step only accepts supported when the model can quote the passage. On a paid plan four models from different vendors read the same sources and a reference passes only when they agree.
Every run produces a receipt that is private unless you share it; it holds the text you pasted and the result. Sources are fetched straight from their publishers. Your content is never used for model training.
Powered by TrueStandard
Multi-model AI verification for high-stakes decisions.