Methods & Sources

Every factual claim on this site points to a numbered source, and every source sits in one registry. We rank sources by how they were checked, confirm that each DOI resolves to the paper we meant, and look for corrections and retractions. It matters: in one case study of articles retracted in 2014, the vast majority of later citations were positive despite clear retraction notices (level C)¹. This page shows what the process does, what is planned and what it misses.

Peer-reviewed work comes first

Peer-reviewed research articles and reviews are the first tier. They carry the measurements, the methods and the uncertainty that a claim needs. A review is weighed by its methods, not its label. A systematic review with a documented search and inclusion rules counts for more than a narrative or critical review, and any review counts only for the question it examined.

Reports from public agencies, scientific academies and standards bodies form the second tier. They often summarize evidence for policy, as the European academies' SAPEA review did in 2019², or document method performance, as the JRC/BAM interlaboratory study did for PET in water³.

Reputable press is the third tier. We use it for dates, events and statements that have no peer-reviewed record yet. When a paper reports the same result, we cite the paper.

Preprints are always labeled as preprints. bioRxiv states that its articles are posted without peer review, editing or typesetting, after only basic screening⁴. EarthArXiv, a server for the earth sciences, says it does not evaluate the scientific quality of submissions⁵.

Marketing material, product claims, unsourced blogs and AI-generated summaries are never cited as evidence that something is true or works. When an organization's own statement is itself the fact being reported, such as a nonprofit's removal totals, we cite it as that organization's statement and label it self-reported. An AI tool may help us find a paper; it is never the thing we cite.

How we searched

This is a curated review, not a systematic review. The search covers the literature from 2004, the year Thompson and colleagues reported in Science that microscopic plastic fragments and fibers are widespread in the oceans⁶, to the snapshot date of 4 November 2026. Work from 2024 to 2026 was searched first, and older landmark papers were added where later work still depends on them.

The tools were general web search, publisher and journal pages, agency and standards-body websites, and the Crossref metadata service⁷. Each page was researched separately, in English, against a written brief. The sources found for single pages were merged into the shared registry on 4 November 2026⁸.

Exact search queries were not logged for every page, so the selection cannot be repeated query by query⁹. A dated search record for each page is planned.

For the shared registry, every item dated 2025 or 2026 was reconfirmed by web search or Crossref on 3 November 2026. Recent records change most often: early-view papers get page numbers, and preprints get published or withdrawn.

How each source is checked

Checking happens in four steps. The first three apply to each source, the fourth to each claim.

1. DOI resolution and match. Each DOI in the shared registry was looked up through the public Crossref REST API, which needs no sign-up and returns the metadata that publishers deposit⁷. The returned title had to match the paper we intended; title, authors, year and journal were then copied from Crossref. Sources first found for single pages had their DOIs resolved and titles matched through Crossref again when they were merged on 4 November 2026.
2. Registration agency. Some DOIs are registered outside Crossref. The SAPEA review is registered with mEDRA² and the JRC/BAM report with the EU Publications Office³. Both resolve correctly, and we mark them as not Crossref-verified.
3. Corrections and retractions. Crossref links correction and retraction notices to the original record¹⁰. Since September 2023 it also holds the Retraction Watch database as an open resource, updated every working day^11,12. We queried these links for every Crossref DOI when the shared registry was first built, and again on 4 November 2026 for every Crossref DOI in the merged registry⁹. A corrigendum to the 2020 reporting guidelines by Cowger and colleagues surfaced this way¹³.
4. Claim-level verification. A verifier opens each cited source and records one verdict per claim: confirmed, contradicted, unsupported or unknown. Every claim starts as unverified. A reviewer from a second model family then reads each group of pages for overclaiming and missing work. For this first release, both steps ran on all 26 pages in November 2026⁹.

Every contradicted, unsupported or unknown verdict then goes to a fix round. There the text is corrected, the claim is removed, or the reason for keeping it is recorded in the page's fix log. Publishing the verdicts and fix logs is planned.

How citations work on this site

Footnotes appear as numbers that link to the source list at the foot of the page. Each entry there gives the full reference with a DOI or stable link, and preprints carry the tag "Preprint · not peer reviewed".

In the site's files, each source has a registry id, such as S-029, that stands for the same document on every page. Sources found for one page carried temporary ids starting with N until the merge of 4 November 2026; the pages now cite registry ids only.

Behind each page sits a claims ledger. Every number, comparison or strong statement is meant to have a line there, with its value, scope, source, year and, once checked, the verifier's verdict. Sentences found without a line get one in the fix round, or are removed.

Key scientific findings carry an evidence level from A to E. Records that are checked rather than weighed, such as a publication date, a DOI's registration agency or a quoted guideline, carry the category P: provenance checked against the primary record. The rubric is on Evidence levels; for the numbers themselves, see How to read the numbers.

Funding and conflicts of interest in the literature

Who paid for a study does not decide whether it is right. It can decide which questions were asked and how results were framed, so we note it when it is material to a claim.

When authors are affiliated with an organization whose work they evaluate, our rule is to say so next to the number. The 2018 estimate of plastic in the Great Pacific Garbage Patch by Lebreton and colleagues is an example. Fourteen of its 16 authors, including the first, list The Ocean Cleanup as an affiliation; the acknowledgments thank its donors for helping fund the research; the paper declares no competing interests¹⁴.

Figures that an organization reports about its own results, such as The Ocean Cleanup's annual removal totals, are labeled self-reported¹⁵. Where a finding is publicly challenged, both sides are cited, as with the 2025 exchange over plastics in human brain tissue^16,17. We do not make allegations about funding or motives that no published source supports.

The registry

All sources live in one registry file that the site is built from⁸. It currently holds 429 entries. Each entry records type, authors, year, publisher, DOI or URL, open-access status where known, the systems and topics it covers, and whether its DOI was verified through Crossref.

The registry is published as a download, sources.json, copied from the registry file at each build; the copy in this release dates from 4 November 2026⁸. The claims ledgers and their verifier verdicts are not yet public; a download is planned. Until it exists, the claim-level record readers can see is the footnote on each page.

The two graphics below show the shared registry, the same file the footnotes come from, as it stood on 4 November 2026, after the merge. The entry count above and the data table under each graphic are computed from the registry at each build; redrawing the graphics automatically at each release is planned.

Shared registry by source type, as of 4 November 2026

Data table

label	value	unit	crossref_verified_doi
journal-article	208	entries	207
review	49	entries	49
report	29	entries	6
standard	8	entries	1
dataset	5	entries	0
press	27	entries	0
preprint	3	entries	3
web	94	entries	1
site-record	6	entries	0
total	429	entries	267

Shared registry by publication year, as of 4 November 2026

Data table

label	value	unit	journal_article	web	report	review	standard	press	dataset	preprint	site_record
1922	1	entries	1
2001	1	entries		1
2002	2	entries		2
2004	1	entries	1
2007	2	entries		1	1
2008	1	entries	1
2009	4	entries	2		1	1
2010	1	entries					1
2011	2	entries	2
2012	3	entries	2			1
2014	2	entries	1					1
2015	5	entries	2	1	1	1
2016	12	entries	8	1	2	1
2017	14	entries	10	2		2
2018	11	entries	7			4
2019	25	entries	17	1	2	2	1	2
2020	42	entries	25	4	1	8	3	1
2021	23	entries	13		3	7
2022	30	entries	19	2	4	4		1
2023	20	entries	13	3		3	1
2024	36	entries	20	3	3	6		2	2
2025	42	entries	23	6	2	6	1	3	1
2026	137	entries	41	55	9	3	1	17	2	3	6
no year recorded	12	entries		12

Known limits

The search was run in English. Studies published only in Chinese, Japanese, Korean, Spanish or other languages are underrepresented, and we have not measured by how much.

Some papers sit behind paywalls. Where only the abstract was available to us, the claim rests on the abstract, and the verifier records that.

Indexes lag behind publication. Papers from the weeks before 4 November 2026 may be missing, and Crossref records for very recent articles sometimes lack corrections that the publisher has already posted.

AI-assisted search finds what is well linked and well cited. It can miss small journals, conference proceedings and data repositories. Several official websites also block automated checks, so their pages were confirmed through secondary listings rather than fetched directly.

If you find an error or a missing study, Corrections & contact explains how to report it. The process that produces these pages is described on How we work, and the model behind the home counters on Flows & model.