Only accepted work is recordedAboutGuidesContactSign in
CitePepCitePep
Guides/Accuracy/How to review an article for accuracy
Accuracy6 min read · Intermediate · Updated Sep 2026

How to review an article for accuracy

The four-step method, the judgement calls it depends on, and the secondary-location sweep that accounted for a third of our accepted corrections.

The method is four steps: isolate each checkable claim, trace it to a source, judge whether the source carries it, record what you found. The difficulty is in the judgement calls, and in one failure mode most checklists leave out.

Between 1 July and 6 October 2026, 1,753 corrections were accepted across 1,700 articles on CitePep. Of the 1,249 that carry a category, 738 were a plainly wrong fact and 529 a source that did not support the sentence citing it. More usefully for a reviewer: 595 of the 1,753 involved a second copy of the error somewhere other than the body, which is the subject of the section below on secondary locations.

Start by isolating every checkable claim

First pull out every statement that can be true or false: facts, figures, dates, names, direct quotes, cause-and-effect claims. A claim you never isolated is one you never checked.

The skill is separating the claim from the framing around it. “A bold move” carries no factual burden; “raised costs 12% in its first year” does. Most sentences blend the two.

Carries no factual burden

“The new policy was unpopular and poorly designed.”

Checkable, and the only part you verify

“The policy raised costs by 12% in its first year.” Check the 12%, the base it is taken from, and the year.

A typical article yields ten to thirty discrete claims. The list is the deliverable: it is what lets you report which claims you checked instead of how the piece read.

Trace each claim to its primary source

One question per claim: where did this come from, and how far is that from the original? Each hop is somewhere a number got rounded, a caveat dropped or a finding widened. The primary source is the study, the dataset, the official document, not the news piece about it.

A real case from the correction log shows how far a chain can drift. An article cited a 2016 University of Bari meta-analysis reporting an effect size of 0.55. Neither the meta-analysis nor the effect size could be verified, and the article's own reference list already contained a genuine source for the same point, Nesbit and Adesope (2006), which the reviewer substituted. The fabricated citation looked more precise than the real one, which is exactly why it survived earlier reads.

The DOI is often the fastest way to settle an attribution. One correction found Lyubomirsky, Sheldon and Schkade (2005) cited to Psychological Review, while the DOI on the same line, 10.1037/1089-2680.9.2.111, carries the prefix 1089-2680, which belongs to Review of General Psychology. The identifier contradicted the journal name in the same sentence.

i

Count the independent sources, not the citations. Six pages carrying a figure that all trace to one press release is one source. Where everything else re-quotes a single origin, the claim rests on that origin, and saying so is more accurate than listing six links.

Verify figures for accuracy and currency

Numbers carry two risks: being wrong, and having been right once. 123 of the 1,249 categorised corrections were the second kind, a figure accurate at publication and quoted forward until the date came off it. Confirm the figure matches its source and that nothing newer supersedes it, then put the date next to the number.

Redo the arithmetic rather than reading past it. One article reported a rise from about 1 billion dollars to 9.7 billion as a 970% overrun; it is 870%. Another put an event's attendance at 165,000 where the organiser reported more than 180,000 from 171 countries.

Watch for a precise statistic resting on an approximate one. An article stated that an organiser “declared 71% of 130,000 attendees in 2025 as decision-makers”, implying about 92,000 qualified buyers. The organiser's own reporting put attendance near 127,000, and no published source supported the 71% at all: a specific-sounding percentage had been layered onto a roughly-correct headline number. A surprising figure is a prompt to recheck, not a finding to publish.

Check quotes word for word, and in context

Two things have to hold for a direct quote: the words are exact, and the context fits the use. A quote can be word-perfect and still misleading.

Check it against the transcript or document rather than against another article quoting it. The failures here are mostly chains: one correction removed a quote grafted onto a named researcher, another removed an unverified line attributed to Walter Pauk, and a third de-attributed an entire invented persona, “Maria Gonzalez”, presented as a practitioner source.

Attribution mistakes hide here, and they are commoner than invention. Across the 1,249 categorised corrections a fabricated quote appears about 3 times, while 529 were sources that did not carry their sentence. One reviewer removed 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. Another de-attributed a line grafted onto a named researcher who had not said it.

The judgment calls: conflicting and unconfirmable claims

Two cases do not resolve cleanly, and both have a right answer that is not “pass it”.

When sources conflict: show what each one says rather than silently picking a winner. “Estimates range from X to Y depending on method” is more accurate than false precision, and more useful.

When a claim cannot be confirmed either way: mark it unconfirmed with the reason. “No primary source found” names the sentence that needs a source, a hedge, or a cut.

Scoping beats deleting where part of the claim holds. One correction narrowed an article's JWST framing for overstating the implications and left the surrounding astronomy untouched, because the rest was already right. Another softened two unverifiable vendor statistics to qualitative language and removed the unfindable report titles from the reference list, rather than cutting the paragraph.

Check the secondary locations

This is the step most accuracy checklists omit, and it accounted for 595 of the 1,753 accepted corrections here: the body gets fixed and a second copy of the same claim, somewhere else on the page, does not.

A worked example. An article had been corrected so the body correctly said Flavell introduced a term in a 1976 paper and elaborated it in 1979. A “Key definitions” callout further down the same page still said the term was coined in 1979, and so did the FAQ block in the page's JSON-LD. The body was right and the page was still wrong in two places, one of them invisible to a reader but served to search engines.

The Lyubomirsky case ran the same way: body and reference list both carried the corrected journal, while the bibliography entry still read “Psychological Review, 12(1)” alongside a DOI that contradicted it, and an FAQ answer still named the “2005 Psychological Review paper”.

Where the second copy hides

  • The bibliography or reference list, which often keeps its own copy of the citation details
  • FAQ answers, which restate body claims in shorter form
  • Structured data (JSON-LD), where an FAQ or article summary duplicates the text
  • Callouts, key-definition boxes and pull quotes lifted from an earlier draft
  • The excerpt or meta description

Search the page for the specific wrong figure, name or date rather than re-reading it. A string search finds the stale copy in the bibliography that a careful read slides past, and it is the only reliable way to catch the one inside structured data.

Record each claim’s status

The output is a verdict per claim: verified, corrected, or to-confirm. That is what a publisher can act on and what a reader can inspect later. “Looks good” is not a review.

918 of the categorised corrections here were a single clean error. The other 331 arrived tangled, most often a wrong fact sitting in the same sentence as a source too weak to hold it, which is why the verdict belongs to the claim rather than the page.

i

An accepted review on CitePep becomes a public record: the claims checked, the category of each error, the reason, and the text before and after. It lands on the page and on your profile, where it feeds a per-topic score rather than a single overall one.

How to keep it efficient

Order the work by stakes. Statistics and load-bearing assertions first, because an error there costs most, and a time-boxed review then covers what matters.

Pre-sourced claims are faster but not free: 529 of 1,249 corrections were a link that resolved to something saying less than its sentence. Opening the link starts the check.

One pass at the end catches the cheapest category of all. Search the page for each wrong figure or name you corrected, because 595 of the 1,753 accepted corrections involved a second stale copy elsewhere on the page.

Quick checklist

  • Every checkable claim pulled out of its framing, ten to thirty on a typical article
  • Each claim traced to a primary source, and that source actually read
  • Figures checked twice, for being right and for still being right, with the date next to the number
  • Quotes verified against the transcript, not against another article quoting it
  • Genuine disagreement between sources shown as a range rather than resolved by picking one
  • Unconfirmable claims marked with the reason, which names the sentence needing work
  • The page searched for each corrected figure, since a third of corrections here found a second stale copy
  • A verdict stored per claim, because 331 corrections carried more than one error in one sentence

Common questions

What makes a source reliable?

The primary source: the original study, dataset or document, with nothing summarising between it and you. Check the identifier too, since a DOI prefix settles a journal attribution faster than the citation text does.

How is reviewing for accuracy different from editing?

Editing works on clarity and structure. An accuracy review asks whether each claim is true and sourced. Different passes, usually different people: a strong editor still misses a wrong statistic in an unfamiliar field. See how to run an editorial review.

What do I do with a claim I genuinely can’t verify?

Flag it unconfirmed with the reason. From there it gets a source, gets softened to what is known, or gets cut. Softening is often the right answer: several corrections here reduced unverifiable vendor statistics to qualitative language rather than removing the point.

How long should reviewing an article take?

With the number and stakes of claims, not the word count. A short piece dense with statistics takes longer than a long, lightly-claimed one. Budget time for the secondary-location sweep as well, since that was a third of the corrections here.

Can I review a topic outside my expertise?

Anything traceable to a clear primary source, yes. Field-specific nuance needs someone with authority in the subject. Name which claims you checked and which sit outside your scope, see how to check a page for accuracy.

Was this guide useful?