A statistic can look precise and still hide the most important facts. A news story may report that "40% of voters support a policy," but leave out who was surveyed, when the survey ran, and how many people answered.
Finding original data sources means working backward until the number can be checked against a dataset, study, official record, or other primary document. We don't treat a polished news article as the final authority. We look for the evidence beneath it, then test whether the wording matches the data.
Key Takeaways
- Start with the exact statistic, not the general topic.
- Trace citations backward through press releases, reports, studies, and earlier news coverage.
- Check the denominator, methodology, time period, geographic scope, and wording.
- Distinguish an original dataset from an original study and from a credible secondary analysis.
- If the source cannot be located, state the uncertainty instead of presenting the claim as fully verified.
Start With the Exact Wording
Copy the sentence containing the statistic. Keep the number, percentage sign, unusual wording, and named organization together. Search that phrase in quotation marks, then try shorter versions if the article has edited or paraphrased the original language.
For example, search:
"40% of voters support""40 percent support" survey"voters support the policy" methodology"40% of respondents" report
The first result may be another news article repeating the same claim. That helps establish a citation trail, but it doesn't prove the number. We want to know where the figure first appeared and who produced it.
Read the entire paragraph around the statistic. Look for footnotes, hyperlinks, a report title, a survey sponsor, a study author, a publication date, or language such as "according to new research." A link labeled "new study" may lead to a press release rather than the study itself. A press release may link to a PDF, a data portal, or nothing at all.
The SIFT tracing method recommends moving away from the page that made the claim and finding the original context. That is useful here. We should pause before accepting a statistic simply because several websites repeat it.
Trace the Citation Chain Back to Original Data Sources
A statistic often passes through several layers:
- An organization publishes a dataset, survey, study, or official record.
- The organization issues a report or press release.
- A news outlet summarizes that material.
- Other outlets repeat the number without checking the first source.
Each layer can introduce a change. A reporter may round the number, simplify the population, remove a qualification, or use a headline that sounds broader than the original finding.
Open every available link. If the news story cites a university study, find the study. If it cites a government report, locate the report on the agency's website. If it cites a research group, search the group's publication archive rather than stopping at a media summary.
Search the title in quotation marks. Search the author's name with a distinctive phrase from the article. Search the organization, number, and year together. These searches often uncover an older report, a revised table, or a public spreadsheet.
We also check timestamps. A current article may rely on a three-year-old statistic. That isn't automatically wrong, but the age of the data must be clear. An unemployment figure, population estimate, crime count, or public opinion result can change meaning when readers assume it describes the present.
A source that cites another source needs another step. Follow the citation until we reach the earliest accessible document. If a 2025 report cites a 2023 study, the 2025 report may be a secondary analysis, not the original research.
Know What Kind of Source You Found
The word "source" covers several different things. They don't carry the same role.
An original dataset is the collected data itself, such as a downloadable survey file, census table, administrative database, or spreadsheet of recorded observations. It may be raw or cleaned. A dataset can support many analyses, but it may not explain every choice made during collection.
An original study is a research publication that presents a question, method, analysis, results, and limitations. It may use original data or analyze an existing dataset. The study is usually the best source for understanding how researchers reached a particular conclusion.
A primary source is a document or record created close to the event or process being described. Government records, court filings, regulatory documents, transcripts, official statistics, research papers, and original interviews can all be primary sources, depending on the claim.
A credible secondary analysis examines or summarizes primary material. A university review, data journalism investigation, or specialist report can be useful, especially when the underlying records are difficult to interpret. It still shouldn't be mistaken for the original dataset.
The distinction matters because the right source depends on the question. If we want to know how many people were counted, we need the data or official table. If we want to know how a survey was conducted, we need the methodology. If we want to understand what a result means, a careful secondary analysis may add useful context.
News coverage is often a secondary source. It can point us toward the evidence, but the article itself is rarely enough to verify a statistic.
Check the Denominator Before Trusting the Percentage
Percentages are fractions. The numerator tells us how many cases met a condition. The denominator tells us how many cases were included. Remove the denominator and the percentage can become misleading.
Suppose a report says that 60% of residents experienced a service problem. That could mean 60% of all residents, 60% of people who contacted a service desk, or 60% of the 80 people who answered an optional questionnaire. Those are different claims.
Look for the following details:
- How many people or records were included?
- Who was eligible?
- Who responded?
- Were nonresponses excluded?
- Was the sample weighted?
- Did the analysis cover a subgroup rather than the whole population?
- Were multiple answers allowed?
The wording can conceal the change. "Six in ten parents reported difficulty finding care" might describe all surveyed parents, parents who already tried to find care, or respondents to a self-selected online poll.
A number can also change when the denominator changes between two time periods. If a hospital reports fewer emergency visits but calculates the rate per occupied bed, the raw count and rate may point in different directions. Both could be accurate. They answer different questions.
A percentage without its denominator is an incomplete claim, not a complete finding.
Read the Methodology, Not Only the Headline
The methodology section tells us how the number was made. It may explain the sample, survey mode, field dates, question wording, exclusions, weighting, margin of error, and known limitations.
For a poll, check whether interviews took place by phone, online, or through a panel. Check whether the poll sampled registered voters, likely voters, adults, or members of a specific group. A result from 1,000 adults cannot automatically be described as the view of all voters.
Question wording matters too. "Do you support a policy?" can produce a different result from "Do you support a policy that would increase taxes?" The order of questions can also affect responses.
Research studies need similar attention. Look for the study population, data collection period, sample size, comparison group, and whether the result shows correlation or causation. A study may find that two factors are associated without proving that one caused the other.
The National Library of Medicine article on using media and text-based sources is a useful reminder that published material can provide evidence while still requiring careful evaluation of how it was produced and used.
We should compare the news wording with the source wording line by line. If the source says "among respondents who answered the question," the article should not present the result as applying to everyone.
Handle PDFs, Paywalls, and Broken Links
The original evidence may be difficult to reach. A report can sit behind a paywall, disappear after a website redesign, or exist only as a badly scanned PDF. Access problems don't make the claim false, but they limit what we can verify.
For a paywalled study, search the exact title. Authors may have posted a preprint, accepted manuscript, conference paper, or institutional copy. University repositories and public research archives can provide the same paper or an earlier version.
For a broken link, copy the URL into the Internet Archive or search the title, organization, and publication year. Search for the file name without the rest of the web address. Government agencies often move PDF files while retaining the same report title.
With PDFs, download the original file when possible. Check the cover date, authors, footnotes, table numbers, appendices, and revision notes. A search result may point to a summary PDF rather than the full report.
Data portals need their own checks. Record the dataset name, publisher, update date, time range, geographic coverage, and download version. A portal can revise historical values without changing the headline report that first used them.
When we cannot access the underlying material, we should say so. "The article attributes the figure to a 2022 report, but the report was unavailable for review" is more honest than calling the statistic confirmed.
Compare the Number With Independent Evidence
Once we locate the claimed source, compare it with at least one independent record. The goal isn't to find two websites repeating the same sentence. The goal is to check whether separate evidence supports the same definition, period, and population.
A government table might confirm a count reported in a news article. A study's appendix might confirm the sample size stated in a press release. A public records database might show that the article used a monthly total while the official report used an annual total.
Check whether the sources agree on:
- The exact number and unit
- The time period
- The geographic area
- The population or sample
- The definition of the measured event
- Whether the figure is raw, estimated, adjusted, or modeled
If sources disagree, don't pick the number that supports the story you expected. Explain the disagreement. One source may use preliminary data, while another uses a revised series. One may count incidents, while another counts people.
We keep a simple source log with the claim, URL, access date, document title, relevant page or table, and notes about limitations. That record makes corrections easier and prevents us from losing the evidence after a link changes.
Report Uncertainty Clearly
Sometimes the trail ends. The original survey may never have been published. A spokesperson may provide a number without methods. An article may cite "research" without naming the researchers. A report may describe a result but omit the underlying table.
Those gaps matter. We can still report what is known, but the wording should match the evidence:
- "The news report says..."
- "The organization claims..."
- "The available report shows..."
- "We could not independently verify the underlying dataset."
- "The source does not state how many people were included."
Avoid turning an unsupported statistic into a confirmed fact through repetition. Three articles citing one unavailable press release are still one unverified trail.
The strongest correction may not be "the number is false." It may be "the number is accurate for a narrower group than the headline suggests," or "the source supports a count, but not the broader conclusion."
Conclusion
A news statistic is only as reliable as the trail behind it. We find the original data sources by copying the exact claim, following citations backward, reading the methodology, checking the denominator, and comparing the result with independent evidence.
When the source is missing or inaccessible, we mark that limit plainly. Verification is not the same as finding a number online. It means knowing who produced it, what was counted, how it was calculated, and how far the evidence actually allows us to go.