How to fact-check an article, step by step
How to fact check a claim, as a repeatable process: isolate the claims, trace each to a primary source, verify figures and quotes, and record what you checked.
Fact-checking is not reading an article and deciding whether it seems right. It is breaking the piece into claims and checking each one.
Between 1 July and 6 October 2026, reviewers working this way on CitePep accepted 1,753 corrections across 1,700 articles from 8 publishers, over 202 topics. The shape of those corrections is what this process is built around: of the 1,249 that carry a category, 738 were a plainly wrong fact, 529 were a source that did not carry the claim citing it, and 123 were a figure that had been right once and was quoted forward until the date came off it.
Step 1: Isolate the claims
You check claims, not articles. Read the piece and pull out every statement of fact. Opinions and clearly framed predictions are not factual claims, though the facts underneath them usually are.
Claims to pull out
- Numbers and statistics, and what they measure
- Dates, names, titles, and affiliations
- Direct quotes and who said them
- Causal claims (“X causes Y”, “X led to Y”)
- Claims presented as established fact
Step 2: Find primary sources
For each claim, find the source closest to it: the original study rather than a summary of it, the official record rather than a blog repeating it, the dataset rather than a chart built from it. A page restating a claim is not a source, only another place the claim appears.
A link that resolves is not a source that supports. This is the most common finding in practice: 529 of 1,249 categorised corrections were a real citation pointing at something that said less than the sentence did. One reviewer cut a quote credited to an anonymous “cloud engineering hiring manager at a mid-size fintech company”, not because it was false but because nothing verifiable stood behind it.
Where a claim traces through a chain of citations, follow it to the end. Claims narrow as you walk back up them. One article carried a formal numbered reference whose source was a Reddit post titled “How I went from 65% to 88% without dumps”, which is a headline rather than evidence.
Step 3: Verify figures and quotes
Figures and quotes are where errors concentrate, because both are easy to copy and easy to bend. For a statistic: who published it, what it counted, over what period, and whether the article uses it that way. Redo the arithmetic, which catches more than it should: one article reported a rise from about 1 billion dollars to 9.7 billion as a 970% overrun when it is 870%, and another put Vivatech's attendance at 165,000 when the 2025 edition drew more than 180,000 from 171 countries.
| Claim type | What to check |
|---|---|
| Statistic | Origin, what it measured, time period, and whether it is used in context |
| Quote | Where it was said, the surrounding context, and whether it was trimmed |
| Date or name | The primary or official record, not a secondary retelling |
| Causal claim | Whether the source shows causation or only correlation |
Step 4: Handle what you cannot verify
Some claims will not check out, either because a reliable source contradicts them or because no reliable source exists. A contradicted claim needs a correction. An untraceable one is unsupported, not true-by-default, and the move is to flag it with your reason for removing or qualifying it. One article stated the 150-coconut-deaths-per-year figure as established fact; it is a myth with no rigorous basis, and the correction said so.
Never fill a gap with an invented source or a number that sounds about right. Fabricating support is worse than the original error, and it is rare for a reason: across those 1,249 corrections a fabricated quote came up 3 times, an invented threshold 3 times, a finding reported backwards 3 times.
Step 5: Record what you checked
A fact-check is worth what its record is worth. Note which claims you checked, what you found, and the sources used. 918 of the categorised corrections were a single clean error; the other 331 arrived tangled, usually a wrong fact sitting in the same sentence as a source too weak to hold it, which is exactly the detail a publisher needs to act on.
On CitePep, a fact-check becomes a public record: the claims checked, what was corrected, and the sources used, tied to the named person who did the work. That is what turns a fact-check into something a reader can verify.
Claims, not articles. Primary sources, not pages that resolve. And a record of what you checked, or the check is unverifiable too.
Common questions
What is the first step in fact-checking an article?
Isolate the checkable claims: figures, dates, names, quotes, causal statements. An article cannot be checked as a whole.
What counts as a reliable source?
The source closest to the claim: the original study, the official record, the person quoted, the dataset. A page that resolves is not evidence that it supports the sentence citing it, and 529 of 1,249 corrections were exactly that failure.
How do I check a statistic or figure?
Trace it to its origin and redo the arithmetic: who published it, what it counted, over what period. 123 of 1,249 corrections were a figure that had gone stale rather than one that was never right.
What do I do when a claim cannot be verified?
Treat it as unsupported and say so. The correction proposes removing or qualifying it, and gives the reason.
Should I record what I checked?
Yes, and the record is most of the value: which claims, what was found, which sources. Without it a reader is taking the check on faith, which is the problem it was meant to solve.