Hi Reader,
Two weeks ago I sent you a number.
Publishers with an OpenAI licensing deal earn 48% more citations in ChatGPT. 129 million citations analyzed. It was the cleanest data any of us had seen on whether those deals actually move anything.
This week that whole category of research went under the knife.
Muck Rack published a public defense of its "What Is AI Reading?" study after a critic argued the PR industry's favorite AI stat rests on a definition of "earned media" so broad it stops meaning much. Muck Rack says 82 to 89% of AI citations are earned. The critique says that only holds if you count nearly every non-owned, non-paid page on the internet as earned media.
Both sides are arguing in the open. Both brought data.
And somewhere this week, an agency is putting one of those numbers on slide 14 of a new business deck without having read either one.
That's the actual risk. Not that the stat is wrong. That you can't defend it when a smart client asks where it came from.
ππ½ Here's the prompt:
π― Expert Mode: The "Stat Defensibility Check"
Scenario: There's a number in your deck. You didn't generate it. Somebody on the team found it, it supports the recommendation, and it has been copied forward through three presentations since. Now you're about to say it out loud to a client who reads more carefully than you'd like. You need to know whether it survives one follow-up question.
Instructions:
1οΈβ£ Pull every stat out of the deck or one-pager you're about to present. Not just the ones you're unsure about. All of them. Most teams find between six and fifteen and are surprised by how many arrive with no source attached.
2οΈβ£ For each one, write down what you actually know: the number, the claim it supports, and where you think it came from. "LinkedIn post" and "somebody sent it to me" are legitimate answers here. Being honest at this step is the whole exercise.
3οΈβ£ Paste that list into Claude along with your core recommendation, and run this:
Act as a research director stress-testing the evidence in a client presentation before it goes out. Below is a list of statistics we plan to use, what each one supports, and whatever we know about where it came from.
β
For each stat, find the primary source. Not the article that quoted it, the original study, survey, or dataset. Then give me: who funded or published it, the sample size and method, the date the data was collected rather than published, any public critique of the methodology, and whether we are using the stat the way its authors intended.
β
Then sort every stat into three buckets. Defensible, meaning you would stand behind it in a room. Defensible with a caveat, and give me the exact caveat sentence to say out loud. And do not use the reason.
β
For anything in the third bucket, suggest a replacement stat that supports the same point and holds up, or tell me the point should be made without a number.
β
If you cannot locate a primary source, say so plainly and mark the stat unverifiable. Do not fill the gap with a plausible-sounding source.
β
[PASTE STAT LIST AND CORE RECOMMENDATION]
4οΈβ£ Read the "defensible with a caveat" bucket first. That's where the value sits. Most stats aren't fake. They're older, narrower, or more vendor-funded than the way you've been using them implies, and one honest clause fixes it.
5οΈβ£ Cut everything in bucket three before the deck ships. Then keep the file. The next deck reuses half of these numbers, and you've already done the work.
Suggested Tool: Claude (it will tell you it couldn't find a source instead of inventing a citation that looks right)
Why You'll Love It
- It turns "where did that come from" from the question you dread into the question you want, because the answer is already sitting in a file.
- It catches the vendor-funded stat before your client does, which is the difference between a footnote and a credibility problem.
- It builds a reusable evidence file, so the numbers you've already vetted stop being relitigated every quarter.
Our industry is about to spend a year arguing about AI measurement stats. You don't need to win that argument. You just need to be the person in the room who read the methodology.
Talk soon,
Chris
β‘ The Cowork Version
Same check, except Claude reads the actual decks in your folder, does the source hunting itself, and leaves a scored evidence file behind.
Setup (one time):
- Make a folder for this and drop in the decks, one-pagers, or proposals you're about to send. PowerPoint, Word, PDF, and Google exports all work.
- Add a short text file naming the client or pitch and the core recommendation that the numbers support.
- Optional: drop in past decks too, so vetted stats carry forward instead of getting checked twice.
- Optional: connect Google Drive or SharePoint if the decks already live there.
Run this Cowork command:
Run a Stat Defensibility Check on the documents in this folder and save the outputs here.
β
1οΈβ£ Read every document and extract each statistic, percentage, dollar figure, and ranking claim. Record the exact wording, the file and slide it came from, and the argument it supports.
β
2οΈβ£ For each stat, search for the primary source. Find the original study, survey, or dataset rather than the article quoting it. Capture the publisher, who funded it, the sample size and method, the date the data was collected, and the URL.
β
3οΈβ£ Search separately for any published critique of that source or its methodology, and note what the critics actually claim.
β
4οΈβ£ Score every stat as Defensible, Defensible with caveat, or Do not use. For anything caveated, write the exact sentence to say out loud. For anything cut, propose a replacement that supports the same point, or say the point should run without a number.
β
5οΈβ£ If no primary source exists, mark the stat unverifiable and say so. Do not substitute a plausible-looking source you did not actually find.
β
6οΈβ£ Save the results as Stat Defensibility Check.xlsx with one row per stat, and save Stat Defensibility Check - Fix List.md listing every stat to cut or caveat, grouped by file and slide.
What you'll get:
- A
Stat Defensibility Check.xlsx scoring every number in the deck, with the primary source and funder attached
- A
Stat Defensibility Check - Fix List.md telling you exactly which slides to change before it ships
- Caveat sentences written out, so nobody has to improvise the qualifier in the room
- Replacement stats for the claims that didn't survive
- A growing evidence file, so next quarter's deck starts from numbers you already trust
Want it on autopilot? Ask Claude to save this as a skill and run it against every deck before it leaves the building. Two weeks in, the team stops pasting in numbers they haven't checked, which is the real win.
Here are a few articles to help keep you up to speed:
1. What does 'earned' mean in the age of AI?β
PR Daily
September 3, 2026
Muck Rack's Matt Dzugan and Linda Zebian respond in public to criticism of their "What Is AI Reading?" research, defending the finding that 82 to 89% of AI citations are earned rather than owned or paid. Their evidence includes an analysis of 489 S&P 500 Wikipedia pages showing 60% of references trace back to journalism. Read it next to the critique before you put either number in front of a client, because this is the stat your competitors are already quoting.
2. Trump administration says a New York Times win in OpenAI lawsuit would "threaten national security" and hurt small newsroomsβ
Nieman Lab
September 2, 2026
The DOJ filed a statement of interest backing OpenAI, arguing that requiring AI companies to pay for training data would hand legacy media an unearned subsidy and hurt small publishers. The filing even cites the Times' own newsroom AI use as evidence against it. If your job involves earning coverage, the economics of who gets paid for that coverage just became a federal argument.
3. EU AI Act Guidance Expands AI Disclosure Rules for Advertisers and PR Teamsβ
Davis+Gilbert LLP
July 31, 2026
Article 50 took effect August 2 with a four-month grace period, which means the real deadline lands in early December. The piece that should get your attention: the EU definition of "deep fake" is far broader than how we use it here, AI-generated press releases and ad copy on matters of public interest need labels unless a human genuinely edited them, and a US agency with no EU office is still in scope if the work reaches EU audiences. Penalties run to 15 million euro or 3% of worldwide turnover.
4. Human-Centric Content Wins on LinkedInβ
CommPRO
September 3, 2026
LinkedIn is testing a button that lets users privately flag posts that seem like AI slop, and more than a million people have already used it. Racepoint's Bill Davies argues the question was never whether comms teams use AI, but whether the output still sounds like a person wrote it. Your executive visibility program now has a crowd-sourced quality filter sitting on top of it.
5. Breaking down complex topics for journalistsβ
PR Daily
September 3, 2026
NARSSA built a local media inquiry portal connecting reporters to 2,000 media-trained Social Security analysts across 47 states, plus a central hub for stats and FAQs. No AI angle, which is the point. If you work in a jargon-heavy category, misinformation fills whatever vacuum you leave, and this is a straightforward playbook for filling it first.
β