1 · What this page covers
CiteLink publishes two independent things, and keeping them apart is the point.
The CiteLink Value (CLV) scores a journal’s publishing practice from journal-level evidence, with no citation input at all. It is described on the evaluation criteria page, and no journal has one yet.
The citation metrics on this page — CNI, percentiles, C1–C4 quartiles, self-citation rates — are computed for the whole catalogue from open data, whether or not a journal has ever applied to us. This page is about those.
Current edition
What changed in v1.1-rc1 (loaded 3 Sep 2026; Decision 20, 2 Sep 2026): the retraction filter now applies to the citing side as well as the cited side, and OpenAlex records verified as one journal — a twin pair or a renamed title — are scored as one journal, every member record showing the same figures. Snapshot and metric year are unchanged. A quartile from one edition is not comparable with a quartile from another. Quote the edition with the number.
The version above covers every figure this page defines — CNI, percentiles, C1–C4 and self-citation. It does not cover the article and author tables, which are restated in their own runs and are not v1.1 yet: articles carries no version stamp at all and cl_author_metrics is v1.0-rc2. This box said v1.0-rc2 until 3 September 2026 and v1.0-rc1 until the promotion later that day; cl_metrics and cl_quartiles held 56,086 and 47,482 v1.0-rc1 rows before it. The corpus line above said 7,167,968 articles until 19 September 2026: the article and author layers grew when the site moved to its own server, and nothing the metric version covers moved with them — same edition, same 2026-07 snapshot, same 47,481 quartile rows.
2 · What these numbers are not
We claim parity with nothing. Specifically:
The reason is not modesty. Each of those metrics has a defining component we cannot reproduce: the Journal Impact Factor depends on Web of Science’s editorial classification of what counts as a “citable item”, which is not published; SJR2 — SCImago’s indicator, computed on Scopus data — depends on a co-citation matrix whose convergence tolerance and dangling-node handling have never been published; Clarivate does not publish the formula behind “Impact Relative to World” at all.
A number we cannot verify is a number we will not publish under someone else’s name. Ours are computed from formulas printed below and reproducible from public data.
3 · The citation window
Every article is counted over its own publication year plus the three following — a fixed four-year window per article, not a fixed calendar range.
This is a deliberate departure from the Journal Impact Factor and CiteScore, and it fixes something both of them get wrong: in a two-year window a paper published in January is observed for nearly twice as long as one published in December. Here every article gets the same observation period, which is also what makes a frozen, citable edition possible.
The cost is latency: an article is not fully measured until four years after publication. We think an honest number that arrives late beats a fast number that measures the calendar.
Retracted work is outside the corpus on both sides of every citation. A retracted article is neither counted nor allowed to count: the citations it received and the citations it made are both dropped before anything is normalized. Through v1.0 only the cited side was filtered (errata E64); the citing side has been filtered since v1.1-rc1 (Decision 20, 2 Sep 2026).
4 · Normalization, and the evidence against our choice
A citation in mathematics and a citation in immunology are not the same event. Comparing raw counts across fields is the single most common misuse of citation data, so every impact figure here is normalized against the articles that share its subfield, publication year and document type. CNI = 1.00 means “exactly the average of that peer group”.
The peer group is the OpenAlex subfield, and the pipeline groups by its id: 246 pools across 239 distinct names, because seven names appear twice under different ids and are normalized separately. This section said “239 subfields” until 30 Aug 2026, which is the name count and not the number of groups anything is normalized against.
Two things are worth knowing about that finding before drawing a conclusion from it. Its gold standard was ChatGPT scores as the primary reference and REF2021 departmental averages as the secondary, neither of which is article-level expert judgement. And a broader level buries exactly what a journal-level index needs to separate: at domain level, a specialist journal and a mega-journal in the same domain normalize against the same pool.
Our reason for choosing subfield is operational, not statistical, and we state it as such. We do not ignore the finding: a parallel quartile at field level is on the roadmap, and every published row already carries which level it was normalized at.
The same paper found something else we publish rather than bury: raw citation counts competed with almost every normalized indicator. So our claim is not that averages are bad. It is narrower — that as a single journal-level ranking driver, a top-percentile share is more defensible than a mean, because it is far less sensitive to one runaway paper. We publish both.
5 · The top-10% threshold
The quartiles rank a journal by the share of its articles that fall in the world’s most-cited 10% for their peer group. That share is computed with fractional assignment at the threshold (Waltman & Schreiber, 2013) rather than a hard cut.
A hard cut cannot produce exactly 10% in any real cell, because citation counts are integers and many articles tie precisely at the boundary. Waltman and Schreiber measured that tied share at 0.4% of publications in biochemistry and 3.6% in mathematics — so the error a hard threshold introduces is itself field-dependent, which defeats the purpose of normalizing. The fractional method returns exactly 10.00% in every cell by construction.
Journals tied on the same share are promoted together to the better band. The four bands therefore do not hold a quarter each: in the current edition the fills are C1 31.7%, C2 29.3%, C3 29.9%, C4 9.0%.
6 · No English filter, no core-journal filter
The CWTS Leiden core-publication definition requires English and drops roughly a sixth of Web of Science. We do not apply it, and this is a deliberate deviation we want on the record.
A filter like that would quietly delete most of the scholarship published in Turkish, Indonesian, Portuguese, Russian and Spanish from the denominator, and then report the survivors as though they were the field. An index whose stated purpose is that visibility should not follow language cannot start by filtering on language.
It has a cost and we accept it: our peer groups are noisier than Leiden’s, and comparisons with CWTS figures are therefore not like-for-like.
7 · Self-citation is shown, not removed
Every journal carries its self-citation rate beside its metrics, and a self-citation-free variant alongside the headline figure. We do not suppress, penalise or silently adjust.
The baseline the published figure should be read against is 6.19% — the journal self-citation rate in the same universe scr_j is computed in (indexed article corpus on both the citing and the cited side, citing window Y−3…Y). This sentence gave 5.12% until 22 Aug 2026, attributed to 2.95 billion edges. Errata E10/K2 re-measured exactly that denominator at 137,119,708 / 2,950,598,362 = 4.65%, and 5.12% did not reproduce — so the number was both wrong and drawn from a wider population than the metric beside it. An editor at 5.8% is unremarkable on the journal page and still below this baseline — 12,171 of the 54,675 journals that carry a rate (22.3%) sit above 6.19%, so 5.8% is in the lower three quarters. This sentence said “above baseline” until 3 September 2026, which reversed the very comparison the paragraph exists to make. A journal well above the baseline is visible in the table to anyone who looks. That is the whole mechanism: an outlier gives itself away, and we do not have to run a private tribunal to decide what counts as too much.
8 · What we will not publish
This list exists because the absence of a metric is usually invisible, and a reader deserves to know what was considered and refused.
| Not published | Why |
|---|---|
| Any author or institution score derived from journal quartiles | A direct violation of DORA and of the Leiden Manifesto. Technically trivial, deliberately absent. |
| “Hot papers” and velocity indicators | OpenAlex citation edges carry no event timestamp, so a two-month accumulation window cannot be computed honestly. |
| JIF / CiteScore / SJR / CNCI parity figures | Each depends on a component that has never been published. See section 2. |
| “Impact Relative to World” | Clarivate does not publish the formula. Publishing an unverifiable number under that name would not be honest. |
| An ESI Highly Cited analogue under that name | The 22-field single-assignment scheme is proprietary and our corpus constraint differs. |
| h-index as a headline metric | Not normalized, proportional to career length, confounded by size. We compute and display it; we do not lead with it. |
OpenAlex’s own citation_normalized_percentile | We examined it and found it unsound. We do not inherit it. |
| Suppression-style sanctions | Hidden thresholds are incompatible with publishing our methodology. Delisting is public and reasoned. |
9 · Known defects
We keep a public errata. It currently runs to 69 entries and records every defect found in this methodology, including the ones still open, with the measurements that found them.
Two are worth naming here because they affect numbers on screen today:
- A catch-all subfield. Our source files articles under one engineering subfield as a fallback, many with no real connection to it. Normalizing against that group produces impact figures we know to be wrong — the worst we have measured reaches 25,556 — and it reaches 308 author profiles. Affected pages now say so on the page. A fix is pending. The size we publish for it is measured as the largest single subfield in the article table, and it now reads 4,091,690 articles — 3.9% of the corpus. Until 19 September 2026 it read 443,233, or 6.2% of the 7,167,968 articles the site then held, and that was the fallback subfield itself. When the article layer grew with the move to our own server the largest subfield became Molecular Biology, which is not the subfield this defect is about — so the figure now measures the biggest subfield rather than the catch-all, and the two have to be told apart again before the fix lands. The 308 profiles are counted directly, from author metrics above CNI-A 100, and are unaffected by that.
- Only part of a journal is ranked. 23,164 of 47,481 quartile rows (48.8%) belong to a journal that published fewer than 60% of its 2022–2024 articles in subfields large enough to rank; the remainder of that journal’s output is ranked nowhere on CiteLink. The figure is journal-wide, so the same percentage is shown on every subfield row for that journal — measured: of 5,944 journals holding more than one quartile row, not one has a varying figure. Affected rows are marked with the percentage. This item previously described the same number as the share of articles “matched into the citation window” because of missing reference lists; that was a different, never-implemented measure. Errata E24.
Neither was found by a reader complaining. Both were found by measurement and written down before anyone asked.
The errata is published as an Atom feed, so you can be told when we find the next one rather than having to come back and check. It carries the 58 entries whose own heading records a date. Eight carry none — E1–E5, E7, E8 and E25 — and the feed leaves them out rather than inventing a date: the repository can only date them to the day the file was first committed, which the entries themselves contradict. This said “the eight oldest” until 3 September 2026; seven of them are, and E25 is not. Entries are in Turkish, like the record. It is listed with our other machine-readable routes on the data page.
10 · Responsible use
DORA Recommendation 1: do not use journal-based metrics, such as Journal Impact Factors, as a surrogate measure of the quality of individual research articles, to assess an individual scientist’s contributions, or in hiring, promotion, or funding decisions.
We publish journal-level numbers, so this applies to ours. A quartile describes a journal. It says nothing about any single article in it, and nothing whatever about the person who wrote it.
The CiteLink Value is separate again: it measures publishing practice, carries no citation input, and must not be read as an impact metric or used to evaluate a researcher.
Everything here is open. The formulas are published, the source data is CC0, and the errata records where we were wrong. If a number on this site looks wrong to you, tell us — that is the correction channel, and we would rather hear it from you than not at all.