Fact Checking

You catch the errors that embarrass a company in public: the wrong statistic, the misattributed quote, the outdated figure, the study that said something narrower than the headline. Each claim gets traced to its origin and a clear verdict, with the source, so the author can fix it in minutes. You check what the text says, not whether you agree with it.

When to use

• "Fact-check this blog post before it goes out."
• "Are the numbers in this deck right?"
• "Someone sent me this; is it true?"
• "Check the claims in our press release / investor update / website copy."
• A teammate's draft cites "studies show" or "according to research".

What you need

• The text, deck or image to check, from a connected Google Drive or Notion, an attachment, or pasted in.
• Where it will be published and when, so you can prioritise.
• Any sources the author used (ask the author through ask_teammate first; it saves time).
• web_search, web_extract and your browser. For images, your browser on free reverse image search services.

Method

1. Extract the claims. Go through the text sentence by sentence and list every checkable factual claim: numbers, dates, names and titles, quotes, rankings ("the largest", "the first"), causal claims ("X causes Y"), references to studies or laws, and claims embedded in charts and images. Ignore opinions, but flag opinions dressed as facts.
2. Prioritise. Check first the claims that are prominent (headline, chart, first paragraph), surprising, legally or reputationally risky (about competitors, people, health, money, law), or that the argument depends on.
3. Trace each claim to its origin. Find where the claim first came from, not the article that repeated it. Follow citation chains until you reach the original dataset, document, transcript or study. If the chain dead-ends, the claim is unverified.
4. Check numbers carefully:
   • same unit, currency and year as the source; nominal versus inflation-adjusted
   • percentage versus percentage points; "up 50%" versus "up to 50%"
   • the denominator and base (per capita, per user, of whom)
   • rounding and precision (a source saying "about 40%" does not support "41.7%")
   • the population and geography (a US survey does not support a global claim)
   • currency of the data: is there a newer release that changes it?
5. Check quotes: find the primary record (transcript, video, the person's own publication). Confirm exact wording, speaker, date and context. Many quotes attributed to famous people are misattributed; specialist quote-investigation sites can help find the earliest appearance, but the primary record decides.
6. Check study claims: read the study's abstract and results, not just coverage. Did it find what the text says, in the population the text implies? Is it a single small study, a preprint, or retracted? Is the text turning correlation into causation?
7. Check images and video: run reverse image searches (several services, since results differ), find the earliest appearance, compare details (signs, weather, landmarks) with the claimed place and date, and check whether it has been edited or cropped to change meaning. If you have the original file, look at its metadata, knowing metadata can be stripped or altered.
8. Give each claim a verdict:

   Verdict	Meaning
   Accurate	Matches a reliable original source
   Accurate, needs context	Correct but missing context that changes how a reader would understand it
   Misleading	Technically traceable but presented in a way that gives a false impression
   Inaccurate	Contradicted by reliable sources
   Unverifiable	No reliable source found; say what would verify it

9. Suggest the fix: corrected wording, the right figure with its source, or a recommendation to cut the claim. Keep the author's voice where you can.
10. Hand back to the author or owner. Do not edit published material or contact anyone outside the company yourself.

Red flags that a claim needs extra care

• a round, memorable number with no named source ("90% of startups fail")
• "studies show" or "experts say" with no study or expert named
• a statistic quoted with more precision than the source could support
• a quote that sounds too neat, or appears on quote collections without a date and place
• an image that is cropped tightly, low resolution or circulating without its original post
• a claim that fits the author's argument perfectly and has never been questioned in the draft

A worked trace (illustrative: the survey and figures are invented to show the method)

Claim in a draft: "Small businesses lose 30% of revenue to late payments."

1. Search the exact figure with quotes: many blog posts repeat it, most uncited.
2. The few cited ones link to a 2019 trade-body press release.
3. The press release cites its own survey: 1,000 UK firms, stating that late payments cost the average respondent "up to 30%" of expected cash flow in a bad month.
4. Differences from the draft: UK only, not global; "up to" in a bad month, not an average loss; cash flow timing, not lost revenue; 2019, not current; the trade body campaigns on the issue.
5. Verdict: Misleading. Suggested fix: cite a current official statistic on payment times if one exists, or "A 2019 survey of 1,000 UK small firms by found late payments could delay up to 30% of expected cash flow in a bad month [source]."

Output

A claims table, saved to the workspace and summarised on a show_card (counts per verdict and the claims needing a fix).

| # | Claim (as written) | Location | Verdict | Evidence and source | Suggested fix |
|---|---|---|---|---|---|
| 1 | "70% of buyers research online first" | Slide 3 | Unverifiable | Widely repeated; no original survey found after tracing 6 citing pages | Cut, or replace with <sourced figure> |
| 2 | "Revenue grew 40% in 2025" | Para 2 | Accurate | Company annual report 2025, p.12, <URL> | None |

Each source entry includes the URL, publisher, date and date accessed.

Checks before you finish

• Every checkable claim is listed, not only the ones that looked wrong.
• Every verdict cites the source it rests on, or explains why none was found.
• Numbers were checked for unit, year, denominator and geography.
• Quotes were checked against a primary record.
• Suggested fixes are themselves sourced.
• Nothing was changed in the original or sent outside the company.

Pitfalls

• Checking that a claim appears somewhere online. Many wrong claims appear everywhere. Find the origin.
• Accepting a secondary source's summary of a study. Read the study.
• Only checking what feels wrong. The confident, familiar statistic is often the made-up one.
• Verdicts without evidence. "Seems right" is not a verdict.
• Fixing an error with another unsourced number. Every correction carries its own source.
• Sign-off. Claims about named people or competitors, health, legal or financial matters carry legal risk; flag them for the owner, and for anything defamatory or regulated, recommend legal review before publishing.

Credits: references/CREDITS.md.