Our data sources
Every reference is queried against all three sources. Each holds a different part of the record: curated notice data, publisher metadata, and biomedical literature indexing. The report states which of them answered.
Primary source for notices
Curated publication events: retractions, expressions of concern, corrections and reinstatements, each with its date and stated reasons. In our tests, the only one of the three that reported a retraction being reversed.
Identification and notice links
Publisher-deposited DOI metadata, used to resolve a reference to a record and to follow the field linking an article to its notices. That field is absent on some deposits.
Biomedical cross-check
Biomedical and life-science literature, queried as an independent check on notice status. It mirrors PubMed and MEDLINE in full, and adds preprints and other records they do not carry. Availability has been intermittent; see the runs recorded below.
Different collections, shown separately. These counts use different units and include overlapping content. Adding them would not give a count of unique papers, and it would not describe how much of the literature is screened.
Why complementary sources matter
In our benchmark, Retraction Watch reported that a trial in JAMA Internal Medicine had been retracted and replaced: the retraction reversed, with a corrected version issued under its own identifier. The Crossref record queried in that test returned no linked update entry at all — so a check resting on Crossref alone would have reported nothing about either half of that, and an author would have been left citing the withdrawn version.
Same reference · observed in our test
update-to entry returnedOne case from our twenty-reference benchmark. It illustrates the value of complementary records; it does not estimate overall accuracy. And a reversal describes publication status, not the scientific validity of a study.
A missing notice, an unavailable source and an uncertain match are three different results, and the report keeps them apart.
Nothing was found in the records returned. This does not establish that no notice exists.
A source did not answer. Its contribution to that check is missing, and the report says so rather than passing over it.
Confirm that the publication found is the one you cited before reading anything into its status.
Every report carries a line such as 2 of 3 databases answered, and names any source that did not. It would be easier to print “three databases” on every report, and it would sometimes be false.
Naming the sources that answered is a weaker claim than naming three, and a true one. Where a source is missing, anything known only to it is missing from that check — which is a different statement from “no notice exists”, and the report does not let the two collapse into each other.
OpenAlex is open, free and large, so leaving it out is a choice rather than an oversight. Two reasons, and the second is the one that decided it.
Its retraction data comes from Crossref, which this tool already queries directly. Adding it would return the same answer twice. A second look that shares a source with the first is not a second opinion, and counting it as one would inflate the confirmation figure on every report.
The deciding reason is what OpenAlex does with that data. A published analysis found that it records publication status as a single yes-or-no value, is_retracted, where Crossref distinguishes between the kinds of notice. Because any notice at all sets that value, an article carrying nothing worse than a correction is presented as retracted. The authors traced misclassified records in the data released between December 2023 and March 2024. A second study, building a dataset from Retraction Watch and the OpenAlex API, found the error runs the other way too: about 7.5% of the articles Retraction Watch records as retracted were not marked retracted in OpenAlex.
Read as a screening test, that is a field with both false positives and false negatives, and no way to tell from the value which you are looking at. For an author checking a reference list, the two mistakes have opposite costs — one casts doubt on a sound paper, the other lets a retracted one through unflagged — and neither can be told apart from a correct answer by looking at the value.
Keeping those apart is what this tool is for. A correction is not a retraction, a reinstated article is not a retracted one, and a reference that is itself a notice is neither. A source that cannot make those distinctions cannot be used to confirm them.
Evidence for the above. Neither is a database this tool queries.
Hauschke C, Nazarovets S. (Non-)retracted academic papers in OpenAlex. Journal of Information Science. 2025. 10.1177/01655515251322478 ↗
Fletcher AHA, Stevenson M. Predicting retracted research: a dataset and machine learning approaches. Research Integrity and Peer Review. 2025;10:9. 10.1186/s41073-025-00168-w ↗
Retraction Watch: 69,504 events in our database snapshot of 11 September 2026. This describes our snapshot, not a live count of the full database, and the events have not been independently recounted for this page. The database is refreshed on working days.
Crossref: nearly 180 million DOI metadata records in the 2026 public data release, announced 17 March 2026. The collection spans many research output types and disciplines.
Europe PMC: 48.6 million publications, preprints and other documents, as described in the search-index snapshot of its About page, accessed 12 September 2026. Full-text holdings are not counted separately.
These figures describe collections at different dates. They are not a shared denominator, not a count of records examined in one check, and not a measure of how well notices are detected.
Crossref acquired the Retraction Watch database in 2023 and distributes it openly. This tool does not read that copy — it reads the notice metadata publishers register with Crossref themselves, which is a separate deposit. That is why agreement between the two counts as independent confirmation here, and why Crossref can hold no record of a reversal that Retraction Watch reports, as it does in the case above.
Europe PMC is a third path again: it receives retraction notices from MEDLINE, as supplied by publishers, and links them to the article. The three sources do hold overlapping content, and a notice reaching all three usually came from the same publisher in the first place — but they are reached by different routes, and the routes fail independently.
Retraction Watch is the primary curated notice source, with more comprehensive coverage of retractions than of other event types. Publisher deposits, record matching and source availability all affect what can be retrieved. An absent linking field on one Crossref record does not establish that no notice exists.
Europe PMC mirrors PubMed, so the notice data above already reaches this tool through it. PubMed is also queried directly, but for a different question: which article a reference points at.
Identification comes first. Nothing can be said about what has happened to an article until it is established which article it is, and Crossref — the source used for that — does not hold every paper. Older work, regional journals and titles from publishers that never registered DOIs are the usual gaps. Where Crossref returns nothing firm, the reference is searched in PubMed by title. A reference identified that way carries a PMID rather than a DOI, and the report says so on the entry.
It is not counted as a fourth source of notices, and does not raise the confirmation figure on any entry. PubMed and Europe PMC both take their retraction notices from MEDLINE, as supplied by publishers, so agreement between them is one record reaching us by two routes rather than two independent findings. The three sources named above are the ones that answer the status question; PubMed answers the identity question that comes before it.
The example concerns reference 8 of our September 2026 benchmark: a trial in JAMA Internal Medicine, 10.1001/jamainternmed.2024.5726. Retraction Watch was the source that reported its reversal in those tests, and the reasons it gives include retract and replace — which is why the report names the replacement rather than telling the reader the version they cited now stands. This describes a recorded test outcome, not a fresh query of the article’s current metadata.
Across the completed comparison runs, no competing tool named the reversal. Those outputs do not establish which sources those tools use, or why the event was missed. The benchmark used twenty selected references and a single assessor; it cannot establish population-wide accuracy or permanent superiority.
Europe PMC is asked once for a whole batch of references rather than once per reference, so a single failed request removes the source from that batch entirely. The service is occasionally unavailable for short periods and answers with a server error rather than data.
A request that fails this way is retried twice, with a short gap, before the source is treated as absent. Most brief outages are covered by that. A longer one is not, and the report says so rather than quietly returning fewer sources.
When a source does not respond, that run is incomplete: the report names the databases that answered, excludes the absent one from the count, and says so on the page and in every export. Anything held only by the missing source is absent from those results, which is not the same as no notice existing.
No uptime figure is given. Availability is not measured here, and a number drawn from our own runs would describe our usage rather than the service.
Publication status is one part of critical appraisal.