Where citation data actually lives: 18 scholarly platforms measured

A reference is only as good as the data behind it. Eighteen platforms a researcher would plausibly use were asked what they declare about their own articles. Eleven answered; the most complete answer did not come from an article page at all.

3 August 2026 · 18 platforms, one pass · raw data

What each platform declares

Platform	Authors	Year	DOI	Pages	ISSN	s
PubMed	2	2017	✓	—	✓	0.52
PMC	4	2020	✓	✓	—	0.46
Europe PMC	Remote end closed connection w					55.49
Springer	5	2013	✓	✓	✓	1.01
PLOS	2	2023	✓	✓	✓	1.45
BMC	6	2023	✓	✓	✓	0.91
Nature	13	2023	✓	✓	✓	0.98
Frontiers	The server answered 404 Not Found.					0.68
arXiv	8	2017	—	—	—	0.21
bioRxiv	The only title the page declares i					0.71
Zenodo	1	2023	✓	—	—	0.51
SSOAR	The page looks like an error messa					0.33
DOAJ	The server answered 403 Forbidden.					0.14
Wiley	0	2022	✓	✓	✓	0.41
MDPI	The server answered 403 Forbidden.					0.19
ScienceDirect	The server answered 403 Forbidden.					0.24
DOI-Resolver	2	2016	✓	✓	✓	0.41
Wikipedia	1	2006	—	—	—	0.18

The finding worth acting on

Resolving the DOI beat visiting the article page. The same work that Wiley's own page serves without page numbers came back complete through doi.org — authors, year, journal, volume, pages and ISSN — in 0.4 seconds, from a publisher whose article pages refuse server-side readers outright.

So the practical rule for anyone assembling a bibliography: if you have the DOI, use https://doi.org/…, not the link your search engine gave you. It is faster, more complete, and it works where the publisher's own page does not.

What this is good for

Keeping a source that will not survive the term

Web pages cited in student work vanish — we have measured how often. A capture keeps the page as it looked, with the retrieval time down to the second and its time zone, a checksum of the image data, and the citation record beside it. When the marker asks what the page said in August, the answer is a file rather than a memory.

Turning a reading list into records

Given a list of addresses, the endpoint returns RIS entries that import into Citavi, Zotero or EndNote without retyping — and names the ones it cannot reach, so those can be opened in a browser instead of being silently dropped.

Sources behind a login

A university licence, a library proxy, a paywalled journal: no server-side reader can follow you there, and three of the publishers measured here refuse them outright. A capture extension runs in your own session with your own access, which is why the two approaches belong together rather than competing.

Feeding a long page to a language model

One continuous sheet with real, selectable text — taken from the document rather than recognised from pixels — and page breaks that fall between lines instead of through them. What the model reads is what the page said.

What it is not good for

• Replacing a publisher's own export. Where an article page offers RIS or BibTeX, that file is authoritative and this is not.
• Proving what a page contained. A screen capture is not a qualified electronic document under eIDAS. It records what a browser displayed at a stated time, nothing more — the checksum covers the file, not the truthfulness of the page.
• Filling gaps. Where a platform declares no page numbers — PubMed's abstract pages, for instance — the field stays empty rather than being fetched from somewhere else and presented as if the page had said it.

Limits of this measurement

• Eighteen platforms is a survey of the common ones, not a census.
• Europe PMC timed out after 55 seconds. That is counted as no result here, not as a fault of the platform.
• Three refusals (MDPI, ScienceDirect, DOAJ) are publisher policy against server-side readers, measured on one afternoon and liable to change.