Citations
Every work in OpenAlex knows which works it cites and which works cite it. This page covers how those links are built, and the citation metric derived from them.
Citations and references
Every work carries a list of the works it cites (referenced_works) and a count of the works that cite it (cited_by_count). Both are built from the same process: extracting each work’s reference list, then matching those references to other works already in OpenAlex.
When OpenAlex creates a work record, it pulls the reference list from the record’s source (Crossref, PubMed, and similar). When the work is open access, OpenAlex can also extract references directly from the PDF, which fills gaps for works whose source records omit references. Each extracted reference is then matched to an existing work. A reference that prints a DOI, PubMed ID or arXiv ID is matched by that ID. A reference with no ID is matched by its text: since November 2026 a reference matcher finds the works whose metadata is closest in meaning to the printed reference, then a model reads the reference next to each candidate and links it only when it is confident, so a reference it can’t place is left unlinked rather than guessed. New reference lists are matched every night as they arrive, so a newly added paper’s references are linked the night its list is read. On a random sample of about 500 of these text matches, about 97% were judged right. A successful match counts as a reference in the citing work and a citation of the cited work.
On a sample of about 110,000 printed references from about 3,500 papers across OpenAlex (November 2026; AI judges checked every reference, and we checked a sample by hand), OpenAlex links about 84% of the references whose cited work is in OpenAlex, and about 92% of the citations it serves are estimated as right. Most of the remaining wrong ones are older links inherited from Microsoft Academic Graph, which we plan to replace with our own matches.
Because citations are built by matching, a work’s referenced_works can be shorter than the reference list printed in its PDF. The common reasons: the cited work isn’t in OpenAlex (references to unknown works are dropped); the source record’s references differ from the final PDF; many Crossref records include no references at all; or a reference has no ID and the matcher wasn’t confident enough to link it.
Field-Weighted Citation Impact
fwci is a snowball metric normalizing a work’s citations for its type, publication year, and subfield. The formula is citations received / citations expected: 1.0 is world average, 2.0 is twice expected, 0.5 is half.
- Citations received (numerator): citations in the publication year plus the three following years.
- Citations expected (denominator): the average of that same 4-year received count over every work with the same year, type, and subfield (articles split journals vs. conference proceedings). Calculation code.
FWCI follows the standard recipe book, so the math matches other databases, but the inputs differ: OpenAlex is more comprehensive (many uncited works pull the expected value down, raising FWCI for cited works); we classify each work into a single subfield from its own text, not its journal; and our publication year is typically the first-online date. Not every work gets an FWCI — work types that aren’t expected to accrue citations (e.g. paratext) are omitted so they don’t distort institutional averages. As of mid-2026 about 218M of 322M works (68%) carry one.
Last updated
View as Markdown