Where did that number come from?

A figure lands in your draft by way of a vendor deck, a LinkedIn post or a chatbot: India's global capability centres generated about $64.6 billion in revenue. It's precise, so it feels sourced. By the time a number reaches you it has usually been copied a few times, and every copy can lose the year, the scope or the caveat. This page shows how to trace a statistic back to where it came from with CiteJury, using two real checks: one where the number was right but the year was wrong, and one where the check couldn't find the number at all, which turned out to be useful too.

How numbers drift

Statistics are rarely invented from nothing. They drift. A fiscal-year figure gets quoted as this year's. An estimate becomes a fact. A finding about survey respondents becomes a finding about "people". A range across countries is replaced by the one figure that suits the paragraph. Each step is small, and after three of them the sentence no longer matches any source.

AI assistants add their own version of this. In the EBU and BBC study of more than 3,000 AI answers published in October 2025, 20% contained major accuracy issues, including hallucinated details and outdated information. A model that learned a figure two years ago can repeat it as current.

Regulators treat a number in a claim as something you must be able to back up. In April 2025 the US Federal Trade Commission said Workado had promoted its AI content detector as "98 percent" accurate, while independent testing found 53% accuracy on general-purpose content. The order the FTC finalized that August bars such claims without competent and reliable evidence. Marketing teams have their own guide to that side of it.

How the check works

Paste the sentence the way it will appear in your work, with the number, the period and the source it's attributed to. CiteJury splits it into parts, so the figure, the year and the attribution can each get their own verdict. It searches several independent search indexes, favors primary sources such as the publisher's own pages, official statistics and filings, and looks for evidence that contradicts the figure as well as evidence that repeats it. It reads web pages and PDFs, including long reports, and skips social media.

The best passages go into a sealed evidence pack. Claude, ChatGPT and Grok read it separately, without web access, and every quote they cite is checked word for word against its source. A final review writes the verdict (Correct, Incorrect, Sources disagree or Can't confirm) with a headline and a safe way to say it.

Here's a real run on the figure from the top of this page, stage by stage:

Two real checks

Right number, wrong year

The claim put the $64.6 billion figure in FY2026. CiteJury split it into two parts, the figure and the year. It searched 79 results, which turned up Nasscom's and Zinnov's own pages alongside coverage in the Economic Times and ThePrint, and locked 12 sources into the pack. A post on X and a Moneycontrol page that couldn't be read were left out.

All three AIs said Incorrect. $64.6 billion is real, but it's the FY2024 figure from the Nasscom and Zinnov GCC report. For FY2026, the 2026 edition of the report puts revenue at $98.4 billion. The number was right, but attached to the wrong year it understated the current figure by about a third. The suggested wording keeps both figures and names the source:

India's GCCs generated $64.6B in FY2024, rising to an estimated $98.4B in FY2026 (Nasscom-Zinnov).

The check took under two minutes on the Standard tier.

The figure no source carried

The second check was on a widely quoted survey about trust in AI, the kind of figure that ends up on a slide about adoption.

The study and the sample checked out against KPMG's own pages. The 46% didn't: none of the seven sources the check used stated it, and the only related figure was a range of 25% to 79% across countries. Claude and ChatGPT said Can't confirm, and Grok called it Incorrect. The final verdict was Can't confirm, which fits the evidence: the pack neither showed the figure nor contradicted it.

The safe wording then did its job. It kept the parts that checked out and said to "source the 46% figure from the report itself before you publish it." We did. KPMG's press release for the study says "only 46% of people globally are willing to trust AI systems." So the figure is real, and now it has a primary link.

That's what Can't confirm is for. It didn't mean false. It meant the evidence in front of the AIs didn't show the number, and it pointed at the exact sentence that needed a source. A tool that filled the gap with its best guess would have been right this time and wrong the next, and you couldn't tell which.

Check a statistic yourself

1. Paste the full sentence. Include the number, the unit, the period and the named source. "Revenue grew 40%" can't be checked. "Company X's revenue grew 40% in 2025, according to its annual report" can.
2. Use Check a claim. Or type it into Ask anything, which works out the kind of check and asks if it isn't sure.
3. Pick a tier. Standard, about 8 credits, is the right default for anything you'll publish. Deep, about 32 credits, searches wider and helps when a figure comes from a niche industry report. Quick, about 2, uses one AI and suits a first look.
4. Read the parts. A figure can be right while its year is wrong, as in the GCC example. The parts show which piece failed.
5. Open the source that backs it up. Confirm the definition as well as the digits: revenue or exports, global or one country, estimate or actual.
6. Use the safe wording, or go to the publisher. If the verdict is Can't confirm, look at the sources listed and the pages CiteJury couldn't open, and get the figure from the original report.

What to look for in a statistic

• Fiscal years. India's FY2024 ran from April 2023 to March 2024. A fiscal-year figure quoted as a calendar year is off by months. Quoted as the current year, it can be off by years.
• Estimates and forecasts. "An estimated $98.4 billion" and "$98.4 billion" aren't the same claim. Keep the hedge the source used.
• Who was asked, and when. A survey finding applies to the people surveyed in the months it ran. The KPMG fieldwork ran from November 2024 to January 2025.
• Ranges collapsed to a point. If a source gives 25% to 79% by country, a single headline figure needs its own source.
• Market sizes. Research firms often publish very different numbers for the same market. Name the firm and the year, not just "analysts".
• Circular sourcing. Ten articles quoting one figure are one source, not ten. Find the first.

Reporters checking figures in copy can find more in the guide for reporters and editors. If the number came from a chatbot, see how to fact-check an AI answer.

What it won't do

• Read paywalled reports. Many market-research reports sit behind paywalls or logins, and some sites block automated access. Those pages can't be read, but CiteJury lists them so you know where to look.
• Read charts or scans. It checks text, not images, so a figure that appears only in a chart isn't checked. Scanned PDFs without a text layer can't be read, because there's no OCR.
• Do your calculations. It checks what sources state. It won't derive a figure the sources don't give, or audit your spreadsheet.
• Cover fresh releases well. A statistic published this week may have little coverage yet, which often means Can't confirm.
• Guarantee the answer. A verdict is only as good as the sources found, and it can be wrong or incomplete. It isn't financial advice.

A statistic is a claim with a source attached, even when the source has fallen off along the way. Putting it back takes a couple of minutes, and the sentence you end up with, carrying its number, year and publisher, is much harder to pick apart.

Questions people ask