Reference checker
Check a bibliography against an economics corpus, and what a miss does and does not mean.
The Reference checker takes a bibliography and tells you, entry by entry, whether each reference exists in a large economics corpus and whether it has been retracted. It is aimed at one specific failure mode: AI-generated citations that look plausible and are not real.
Who built the corpus
The checker is powered by Eddy's Papers, built by Eduard Brüll (ZEW). The corpus is roughly 455,000 economics papers drawn from RePEc and EconLit. None of that is our work. He gave us access while his own service is still in preview, and every check you run here costs him compute.
His tools, which are worth using directly:
- https://econpapers.eduard-bruell.de — paper search
- https://econpeople.eduard-bruell.de — economist and author lookup
- https://agenticsearch.eduard-bruell.de — agentic search over the corpus
- https://github.com/edubruell/eddyspapers — the R package
What a miss actually means
Not found does not mean fabricated. It means unverified against this corpus, and nothing more. Do not present a miss as proof that a citation was invented.
There are several ordinary reasons a genuine reference will not be found:
- It is not economics. The corpus is RePEc and EconLit. A real computer-science, sociology or medicine paper is simply out of scope.
- It predates 1995. Coverage starts there, so older work is legitimately absent.
- It is not a journal article or working paper. Books, chapters, reports, datasets and theses are largely outside the corpus.
- The entry is malformed. A mangled author list or a wrong year can defeat the matcher on a paper that is present.
A hit is strong evidence the work exists. A miss is a prompt to check by hand. Treat the two asymmetrically.
Retractions
Where the corpus records a retraction, the checker surfaces it. This is the more valuable signal for most people: a citation can be perfectly real, widely cited, and withdrawn. Retraction data is only as complete as the upstream record, so absence of a retraction flag is not a clean bill of health.
Using it
Paste references one per line, or paste BibTeX. Submissions are capped at
50 entries per request and the endpoint is rate limited, because each check
costs someone else compute. It accepts POST only, so a crawler
cannot trigger a run by following a link.
Deliberately not exposed: his literature-search tool, which spends model tokens per call. We use only the tools that do no inference, so the cost we impose on him stays bounded. If you want full search, go to his site rather than asking us to proxy it.
If the service is unavailable, the page says so rather than silently reporting everything as missing. A checker that fails closed would be worse than useless here, because "not found" is exactly the answer people over-read.