A tool that knows where it stops.
CitePure exists because AI assistants made it cheap to produce references that look authoritative and expensive to discover that they are not. The fix is not another confidence score. It is a dated, reproducible record of what was checked against which index, with the places where the method is silent clearly marked.
Evidence over certainty
Every verdict names the index, the matched record, the DOI and the score. A tool that cannot show its work does not belong in a submission packet.
Determinism over fluency
No model decides whether a paper exists. Rules, public metadata and a published formula mean the same input produces the same result and can be described in a Methods section.
Named boundaries
We say what is not checked: AI-written text, plagiarism, semantic support, and literature absent from our two indexes. "Not found" is never presented as "fabricated".
There is no server holding your paper.
CitePure is a static site. Parsing, scoring, reconciliation and report generation run as local JavaScript in your browser. The only network traffic is a read-only lookup request containing a DOI or a title, sent to Crossref or OpenAlex.
What we do not have
No account system, no database of manuscripts, no analytics or advertising script, no cookies, no session identifier, no payment form, no email collection, no training corpus. Your draft text is never transmitted to us because there is no "us" in the request path.
What leaves the browser
A DOI string or a title string, sent directly from your browser to public metadata endpoints. Your browser's normal IP-level metadata is visible to those services under their own policies. Nothing else. The process log lists every request it made.
A report is an aid to review, not a review.
CitePure checks bibliographic metadata. It does not decide whether a source supports the claim that cites it, whether the work is original, or whether the author behaved properly. Those judgements belong to the author, the supervisor, the reviewer and the institution.
Coverage is limited to Crossref and OpenAlex; a clean result is a statement about those two indexes at the time of the run. Rule-based parsing can misread unusual formats. The exported report and the Methods statement both carry these caveats, and we ask that they not be stripped when the report is filed.
Metadata is provided by Crossref and OpenAlex under their respective terms. CitePure is not affiliated with or endorsed by either, nor with any AI assistant, publisher or university.
What comes next
Local-language index adapters
CiNii / J-STAGE / JaLC, SciELO / Dialnet, CNKI / Wanfang. The real coverage gap in reference verification is not language on the page — it is which catalogue the tool can query.
Automatic claim splitting
Turning a full draft into individual claims and checking each against candidate sources, in one run instead of one pasted paragraph at a time.
Team audit trails
Signed, shareable verification records for editorial offices, supervisors and integrity teams, plus the API / MCP tier described on the pricing page.
Localised editions
The site launches in English first. Japanese, Spanish and Portuguese editions will be derived from it afterwards, with translated interior pages — not just a translated homepage. Page language alone changes little; the local catalogues behind them are the real gap.
Why I built CitePure: A 3 AM discovery before thesis submission
Julian Vance
In the autumn of 2024, I was working late on chapter three of my literature review on cognitive scaffolding and algorithmic mediation. Like many graduate students under tight deadlines, I used Claude and ChatGPT to help rephrase dense paragraphs and suggest relevant papers in adjacent subfields. The generated draft sounded articulate, confident, and impeccably academic.
Forty-eight hours before my supervisory committee review, I sat down with a mug of black coffee to click through every single Digital Object Identifier (DOI) in my 34-item bibliography.
- 7 DOIs returned dead 404 errors — the papers literally did not exist in any registry.
- 5 references were "Chimera Citations" — the DOIs were completely real, registered works in PLOS ONE and Nature, but they pointed to papers about deep-sea sediment ecology, not cognitive linguistics. The language model had borrowed real DOIs from its training corpus and stitched them onto fabricated paper titles.
- Had I submitted that draft, my advisor would have spent five minutes spot-checking references and reasonably concluded I had either faked my data or committed academic misconduct.
I spent fourteen uninterrupted hours manually copying and pasting titles into the Crossref search bar at 3 AM to find what actually existed. During that sleepless night, I realized two fundamental things:
- AI detectors (like Turnitin AI or GPTZero) are useless for this problem. They spit out probabilistic guessing percentages ("47% AI-written") that tell you nothing about whether a citation is authentic or where the claim came from.
- Researchers do not need another generative tool promising to "write papers faster." What researchers desperately need is an uncompromising, deterministic verification script: something that queries official metadata registries, checks token similarity with transparent mathematical thresholds, and gives you a clear pass/fail verdict you can prove in a Methods section.
I wrote that verification script in Python the following week. When lab colleagues asked for copies, I stripped out the server backend entirely, rewrote it in lightweight client-side JavaScript, and deployed it as a static web tool. That way, no student or faculty member ever has to upload an unpublished manuscript draft to someone else's cloud server.
CitePure is 100% free, runs in your browser, and will never have generative models deciding whether a paper exists. If an indexed registry like Crossref or OpenAlex doesn't have it, we don't guess. That is our permanent promise.