In partnership with
Are your sales foundations ready to scale?
You’re growing from 10 reps to 100. Your agents can only work with the data and processes you give them, and the foundations you set now decide whether your systems hold.
Attio’s agentic sales audit looks at where you are today and shows you where to start, whether that’s your data foundation or a fuller agentic sales process.
iPrompt
DEEP DIVE · iPROMPT #160 COMPANION
4 NOVEMBER 2026
How to trace an AI statistic
By the end you’ll have a way to sort every AI-supplied number in a document into four verdicts, fix or cut each one, and keep a record of what you checked and what you couldn’t.
5 MIN READ · FICTIONAL WORKED EXAMPLE AND CHECKLIST
TRACING AI STATISTICS / WORKED EXAMPLE
Three quotes that weren’t there
At eleven on a Sunday night, the account director at Callow & Finch pasted the 20 statistics from a pitch deck into an AI assistant and asked it to trace each one. By midnight she had 20 verdicts, each backed by a quote from a source page. Then she did the part most people skip: she searched each page for its quote.
Three weren’t there. One was a paraphrase. Two stitched together sentences from different pages. None of the three changed a verdict this time, but each was a hop nobody had actually made.
Checkers miss in both directions. Trakkr’s own AI judge, working alone, rated 23% of statistics sourced; hand checks on 150 trails lifted the estimate to 40%, because people found originals the model had missed. So the AI drafts the trail, and you find every quote on its page before a verdict counts.
Callow & Finch, a nine-person B2B marketing agency, is fictional, and so are its figures. Its 31-slide deck for a logistics software firm was drafted with AI. Statistics follow Trakkr’s definition: a share, rate, average, count or cost stated as fact. Prices and the agency’s own results were checked separately.
Before tracing, she wrote the rule. A number stays only if it reaches original research that matches on five points: what was measured, the year, the source, who was asked, the place. A dead page or a missing number isn’t a verdict until the archive and the named report have been tried; an uncited number starts from a search for the exact claim. Agreement between assistants counts for nothing. That gives four verdicts.
Verdict | What it means | What you do | How it reads |
Sourced | A study says the same thing on all five points | Cite the study, not the page that quoted it | ‘In [source]’s [year] survey of [n] [who] in [place], [figure]’ |
Changed | A study exists but differs on the figure or any of the five points | Use the original’s figure and context, or cut | As above, with the original’s details |
Can’t check | Paywall, form, dead page, no number or the five-hop limit, after trying the archive and the publisher | Attribute only a side point whose publisher’s own statement you can see. Anything else, and anything carrying a headline or recommendation: cut | ‘[Publisher] reports [figure]’ |
Nowhere | The trail ends at pages that credit no study | Cut. Use your own data if you have it, checked like any other | It doesn’t |
The one everyone agreed on. Slide 4’s headline said winning a customer costs 5 to 25 times more than keeping one. Every assistant the team tried gave that range, and a dozen blogs repeated it. The trail ran through roundups quoting roundups and ended without a study, which is where Trakkr’s ended too: it’s one of four figures all five assistants quoted with no study found. Nowhere: cut. The client’s own retention figures went in instead, checked against its CRM export, and the claim changed with them. Retention data can’t say what winning a customer costs against keeping one; it can say how long this client keeps its own. So the headline now says that.
The one that looked fine. Slide 9: ‘78% of B2B buyers research on their own before speaking to sales.’ The cited page had the number. So did the study behind it, a 2019 vendor survey of 212 US software buyers. Three of the five points had drifted: the year, who was asked and the place. Changed: rewritten with its date, sample and country, and moved from the headline to a footnote.
The three that stayed out of reach. Two figures came from paywalled analyst reports. Both were side points, and each publisher’s own summary page stated the number, so they went in as ‘[publisher] reports…’. The third came from a dead page the Wayback Machine had never saved, and it sat under a recommendation on slide 22. Cut.
What the trace can’t tell you. Sourced means a study says the same thing. It doesn’t mean the study is any good: 15 people, a vendor surveying its own customers, a leading question. Tracing finds the original; judging it is a second job. Two of the eight sourced figures were vendor surveys, and the slides now say so.
The full audit log: about six hours
Step | Time | Result |
1 List every number with its sentence, slide and cited link | 25 minutes | 27 numbers; 20 statistics once 4 prices and 3 own-client results were set aside |
2 Run the prompt on each statistic | 45 minutes | First hop: 16 of the 20 cited pages contained the figure; 4 didn’t |
3 Find every quote the AI gave on the page it named | 2 hours | 3 quotes weren’t there: 1 paraphrase, 2 stitched from different pages. All 3 re-traced by hand |
4 First sort | 20 minutes | 7 sourced, 3 changed, 7 can’t check, 3 nowhere |
5 Work the can’t-checks: archive, then publisher | 1 hour | Wayback Machine recovered 2 dead pages; 2 gated reports downloaded through their forms. Re-sorted: 2 sourced, 1 changed, 1 nowhere. 3 still out of reach |
6 Second pair of eyes on 5 random traces | 20 minutes | 4 agreed. 1 sourced figure moved to changed: the study surveyed US firms only |
7 Final sort | 10 minutes | 8 sourced, 5 changed, 3 can’t check, 4 nowhere |
8 Fix the deck | 1 hour | 8 re-cited to the original; 5 rewritten to the original’s figure and context; 2 attributed to their publisher; 5 cut, 1 replaced with the client’s own data under a narrower claim |
The record. ‘Statistics check, logistics pitch deck, 1–2 November. 20 statistics traced. 8 sourced, cited to the original; 5 changed, corrected to the original’s figure and context; 3 can’t check, of which 2 side points attributed to the publisher and 1 cut; 4 nowhere, cut. 1 cut figure replaced with the client’s retention data, checked against its CRM export, with the slide’s claim narrowed to match. 2 sourced figures are vendor surveys, labelled. 3 AI quotes failed and were re-traced. Second check on 5 by the managing director.’
TRACING AI STATISTICS / THE CHECKLIST
Carry it over
A link isn’t a source. The cited page had the number 82% of the time in Trakkr’s test. That proves someone repeated it. Keep following until you reach a method, then match all five points.
Agreement isn’t a check. Five assistants quoting the same figure can mean five assistants reading the same roundup.
Check the checker. Find every quote an AI tracer gives on the page it names. A quote you can’t find is a hop you haven’t made.
Sourced isn’t sound. Look at the sample and who paid for it, and say so when it’s a vendor surveying its own customers.
The strongest objection
‘Everyone uses these numbers, and nobody checks.’ Mostly true, and six hours for 20 statistics is real time in a pitch week. So trace the numbers that carry an argument, the ones in a headline or behind a recommendation, and attribute or cut the decoration. What tips it: the one reader who does check finds one zombie, and stops believing the other nineteen.
YOUR MOVE
Take one document going out this week. List every number an AI supplied, write your five-point rule, and trace the three that carry the argument. Find every quote on its page, sort each number into the four verdicts and fix the document to match. Reply with one line: how many you traced and how they sorted. Leave out client names.
THIS WEEK’S FORECAST
By 30 June 2027, at least one of ChatGPT, Perplexity, Gemini, Claude or Google’s AI Overviews will show, by default, whether each cited source is original research or a page repeating someone else’s figure. Only 21% of the pages these assistants cited in Trakkr’s test were original research, and users can’t see which. I expect someone to label it.
Counting. It counts if the product’s official help pages, changelog or announcement describe a label or marker that tells original research or primary sources apart from secondary pages, shown by default next to citations, in a release before the deadline. A setting users must switch on, or a separate research mode, doesn’t count. Void if such a default label already existed before 4 November 2026.
A number should travel with its sample.
New to iPrompt? Every Wednesday, the week’s AI news turned into one thing you can do. Subscribe free
Disclosure: the original draft used Claude, made by Anthropic. Claude is one of the assistants Trakkr tested.
iPROMPT / 160 /