How to fact-check numbers in an AI-generated report

An AI-generated report reads the same whether its numbers are right or wrong. The prose is fluent, the figures are specific, and there is often a citation next to each one. That is exactly why numbers in these reports deserve a deliberate check before they go into a deck, a memo or a client email.

The good news is that most errors fall into a small number of patterns, and a methodical pass catches nearly all of them. This guide covers what to check, in what order, and what to do when a number doesn’t hold up.

What usually goes wrong

In practice, problems with numbers in AI-written research rarely look like pure invention. More often, a real figure has been used in the wrong way:

• Right number, wrong year. A 2022 estimate is presented as the current figure, or a forecast is presented as an actual.
• Right number, wrong definition. A figure for global shipments appears in a sentence about US revenue.
• Right number, wrong source. The figure is real but comes from somewhere other than the cited page.
• Derived numbers that don’t compute. Growth rates, shares and totals that are calculated from correct inputs, but calculated wrongly.
• Citations that don’t support the claim. The source exists and mentions the topic, but says something weaker or different.
• Figures with no traceable source at all. These do happen, and they are the easiest to catch if you look.

Each step below targets one or more of these.

Triage first: find the load-bearing numbers

You don’t need to check every figure with the same intensity. Start by marking the numbers that carry weight:

• figures in the title, summary or conclusions;
• figures that appear in charts or tables;
• figures that a recommendation depends on;
• anything surprising, unusually precise, or unusually round.

Check these fully. For the rest, a quick look at the source is usually enough.

Step 1: Trace each figure to a source you can open

For every load-bearing number:

1. Open the cited source. Does the link work? Is it the page the citation describes?
2. Find the number on the page. Search for it, and try the formats it might appear in: “1.2 billion”, “1,200 million”, “1.2bn”, or “1,2 Mrd.” in a German source.
3. Read the surrounding paragraph. Make sure the number refers to what the report says it refers to.
4. Go upstream if needed. If the source is a news article or blog quoting someone else’s research, find the original publication and check the figure there. Secondary sources round, truncate and occasionally mistype.

If you can’t find the number in the cited source after a reasonable search, treat it as unverified until you find it somewhere you can cite.

Step 2: Check the date

There are two dates to check, and they are easy to confuse:

• When the source was published. An analysis from several years ago may have been overtaken by events.
• What period the number describes. “The market will reach X in 2026” in an older report is a forecast, not a measurement, even if 2026 has now arrived.

Also check fiscal versus calendar years. Many companies and governments report on fiscal years that don’t start in January, so “2025 revenue” can mean different twelve-month periods for different organizations in the same report.

Step 3: Check the definition

This is where the most subtle errors live. Before you accept a figure, establish exactly what it counts:

• Revenue, units, or installed base? Sales in a year are not the same as total products in use.
• Which geography? Global, a region, a single country, or “major markets”?
• What’s included? Hardware plus software? Services? Gross or net of returns and discounts?
• Which currency, and nominal or real? A figure in nominal local currency and one in inflation-adjusted dollars cannot be compared directly.
• Which segment? A figure for “electric vehicles” may or may not include plug-in hybrids, depending on the source.

When a report compares two numbers, check that both use the same definition. Comparing them when they don’t is one of the most common errors in any research, AI-assisted or not.

Step 4: Check units and scale

Quick checks that catch a surprising number of mistakes:

• Millions versus billions, and thousands separators versus decimal commas. In many European sources, “1.234” means one thousand two hundred and thirty-four.
• Percent versus percentage points. A rate rising from 10% to 12% has risen by two percentage points, which is a 20% increase.
• Per-capita versus total figures.
• Units in other languages and systems, for example Chinese 万 (ten thousand) and 亿 (one hundred million), or Indian lakh and crore.

Step 5: Recompute anything derived

If the report calculates something, calculate it again:

• Shares. Do the parts add up to the whole, or to 100%?
• Growth rates. Compound annual growth rate is (end value ÷ start value)^(1 ÷ number of years) − 1. Check the number of years. It’s a common off-by-one.
• Totals. Do sums in tables match the totals shown?
• Charts versus text. Does the chart show the same values the text describes?

A spreadsheet takes a few minutes and settles the question for good.

Step 6: Cross-check against an independent source

Find a second source for each key figure, and make sure it is genuinely independent. Two news articles that both quote the same analyst firm are one source, not two. Good independent checks include official statistics, company filings, regulator data, and a different research firm using its own method.

Different sources will often disagree a little because their definitions and methods differ. That’s normal. What you want to know is whether the disagreement is explainable. If it isn’t, find out why before you use either number.

Step 7: Check the claim, not just the number

Finally, re-read the sentence the number sits in and ask whether the source supports that sentence. Watch for:

• correlation presented as causation;
• a single example presented as a trend;
• a forecast presented as a certainty;
• a qualified statement (“may”, “in some cases”) turned into an unqualified one.

When a number fails

You have four options, roughly in order of preference:

1. Correct it with the right figure and a source you have checked.
2. Replace it with a better-supported figure that makes the same point.
3. Qualify it: give the definition, the date or the range, so the reader can judge for themselves.
4. Remove it. If a claim depends on a number you can’t verify, the claim probably shouldn’t be in the report.

Whatever you do, note it in your source log so the next reader doesn’t have to repeat the check.

The checklist

Copy this into your own notes:

• Load-bearing numbers identified and marked
• Each one found on the cited page, or in the original source upstream
• Publication date and reference period checked; forecasts labeled as forecasts
• Definition matches the sentence: what is counted, where, and in what currency
• Units, scale and number formats checked
• Derived figures recomputed
• Key figures confirmed by an independent source, or the gap explained
• Each claim supported by its source as written, not just its number
• Fixes recorded in the source log

Making the check faster

This process is much quicker when every number arrives with its source attached, so you start from Step 1 instead of hunting for where a figure came from.

That’s how Divrite is designed. Figures in its reports are cited to the pages they came from, numbers in charts can be traced back to the cited text, and when you export a presentation, each slide’s sources go into its speaker notes. It doesn’t remove the need to check. The steps above still apply to anything you’re going to put your name to. It just means the checking starts with a link rather than a search.