AI handed you the report, the spreadsheet, the dashboard, the stat — all without the person who used to check them. Here’s the specific way that bites you.
A friend of mine watched her CEO start pulling his own numbers. He’d query a data set himself, have AI tell him what the numbers meant, and act on it — no analyst in the loop, because he didn’t need one to get an answer anymore. He can now. So he does.
Another colleague watched a senior leader drop a company report into an AI tool, get back a tidy summary studded with statistics, and publish it. The stats came from six-year-old material the company had quietly retired years earlier. Nobody flagged it, because the tool didn’t flag it, because the tool isn’t built to. People came to him afterward asking where the numbers came from. The answer was: a document the company sunset in 2020, still sitting in a public archive, found and served because it was still reachable.
This is the new normal, and most of us are living some version of it. You write the thing you’d have briefed someone else to write. You build the dashboard you’d have requested. You pull the stat you’d have asked an analyst to verify. The friction is gone, the wait is gone, the back-and-forth is gone. It feels like pure upside.
The person who used to sit in that gap wasn’t the inefficiency. They were the verification step wearing a human face. And when you route around them, you don’t just save the time they cost — you also lose the thing they caught. I want to show you exactly what gets through now, using the smallest, most boring example I can find: a single statistic.
The Afternoon I Traced One Number
On June 10, I was tracing the source of a statistic for a client — one of those numbers everyone in the industry quotes, the kind you’ve seen on a hundred slide decks. By the end of the afternoon I’d checked eight of the most-quoted stats in that field. Five of them traced back to exactly two dead sources.
One was a 2010 white paper. The other was a 2016 Gartner article that reported rates “as high as” 98% and 45% while citing nothing more specific than “various sources” — a hedge the industry has been sanding off ever since. More on that one later. The first is where the story really starts.
Here’s the part that stuck with me. The author of that 2010 white paper publicly retired his own number, back in 2022. And it didn’t matter at all.
The Stat That Wouldn’t Die
The number was the famous “98% of text messages get read.” It came from a 2010 white paper by a UK firm called Mobilesquared, which also produced its even more quotable sibling, “90% of texts read within three minutes.” That second one became, by the firm’s own description, the most-quoted statistic in mobile marketing.
In 2022, the firm’s founder, Nick Lane, published a post on his own company’s site walking it back. In his words: it pains him to say it, but neither number reflects how mobile messaging actually works today. Their newer data put the real read rate around 55%. He named the stat, named the year, gave you the corrected figure. This is what an honest retraction looks like. He did everything right.
And then nothing happened.
I pulled vendor stat roundups published in November 2025, March 2026, April 2026 — well over three years after the walk-back. Every one of them still leads with 98%. One was published three weeks before I went looking. The retraction lives on a single page from 2022. The dead stat lives on dozens of pages dated last month.
You don’t even have to leave Mobilesquared to watch the correction lose. The firm’s own LinkedIn page still markets the company as the people behind “90% of SMSs are read within 3 minutes.” The retraction couldn’t outrun the original inside the building where both were written.
The Other Way Stats Die: They Get Promoted
Not every dead stat was retracted. Some never had to be — they just got stronger every time they were copied.
Here’s a favorite from the same audit: “leads contacted within 5 minutes are 21× more likely to convert.” You’ll find it everywhere, usually credited to an MIT study. Two things are wrong with that, and both happened in transit. The original 2007 research didn’t measure conversion — it measured qualification, whether a lead so much as entered the pipeline. “Qualify” quietly became “convert” somewhere along the way, which turns “answered the phone and seemed real” into “bought.” And it wasn’t an MIT study in the way that phrase implies — it was a vendor’s research with an MIT-affiliated co-author, presented at a marketing conference. Real study, real finding, real co-author. Just not the claim that’s now in circulation.
So this one didn’t die from a retraction nobody read. It died from a game of telephone that made it better-credentialed and more impressive at every hop. Nobody lied. The number simply got promoted past what it could back up — and AI, asked for it today, hands you the most decorated version, MIT badge and all.
It’s Not Hallucination. That’s Why It’s Worse.
This is the part people get backwards. We’ve all been trained to worry about hallucination — the model inventing a fake number out of thin air. That’s not what’s happening with either of those stats. The model retrieved real numbers, from real published pages, that real companies really wrote down. One was killed by its own author; one got promoted past its evidence by a decade of copy-paste. Neither was invented. The AI is reporting your industry’s archive back to you faithfully. The archive is just full of corpses, and some of them are wearing medals they didn’t earn.
That’s a different failure than hallucination, and your fact-checking probably doesn’t catch it. If your verification step is “does a real source say this,” the dead stat passes. A real source says it. Dozens do. The source is simply dead, and “dead” isn’t a field your check looks at.
Which is the whole problem with doing it yourself now. When the work routed through a specialist, the check came bundled. The analyst who pulled the number knew the 2010 paper had been walked back, because knowing that was their job. The writer who’d worked the beat for a decade could smell a stat that was too round and too old. You weren’t paying them only for the output. You were paying them for the frown — the half-second where they looked at the source and didn’t like it. AI gives you the output and skips the frown, and unless you know to supply it yourself, nobody does.
The Cruelest Part: Fresh Beats True
You’d hope the machines would eventually sort this out — that newer, better data would surface and the zombie would starve. It works the other way.
AI assistants lean toward fresher content when they decide what to cite. An analysis of seventeen million citations across seven AI platforms found AI assistants cite content that’s meaningfully newer than traditional search serves up — and ChatGPT showed the strongest pull of all, citing pages over a year fresher on average. Freshness sounds like a quality filter. For dead stats, it’s a feeding tube.
The retraction sits on one page, last touched in 2022 — old, stale, sinking. The dead stat gets relaunched every January in a fresh “SMS Statistics for 2026” roundup, exactly the signal the machines reward. The correction ages out; the corpse gets a birthday every year. Freshness preference doesn’t rescue the truth — it selects the most recently laundered copy of the lie.
So the honest author who corrected the record in 2022 is now outranked, in the systems that matter most, by people who never got the memo and republish the original number annually.
And it’s not only retracted stats that get this treatment — sometimes the freshest, most confident version of a number wins precisely because the careful original got locked away. Remember that 2016 Gartner article, the second of my two dead sources? Its actual wording was hedged twice in a single breath: open and response rates “as high as” 98% and 45%, per “various sources.” That’s a sentence explicitly refusing to vouch for anything. Watch what a decade did to it. The hedge got sanded off one copy at a time until you now find the same figure published as a flat “Gartner reports a 98% open rate” — qualifier gone, and Gartner quietly promoted from quoting the number to being the source for it. The one party in the chain that wouldn’t stand behind the stat is now the authority everyone credits for it. And the kicker: Gartner’s hedged original is gated behind a client login. The stripped, confident, wrong version is free, fresh, and everywhere. Of course that’s the one the machines serve.
The Version With a Villain — and the One Without
I’ve written before about AI Recommendation Poisoning — companies deliberately injecting hidden instructions so AI assistants recommend them. That’s the answer pipeline being attacked on purpose, by someone who wants to win.
Dead data is the same pipeline failing with nobody attacking it. No villain. No injection. Just an archive nobody emptied, served back at speed and with confidence. In a way it’s the more uncomfortable version, because there’s nothing to patch. You can defend against an attacker. You can’t defend against your own filing cabinet — especially when you’ve just removed the one person who knew what was in it.
Go back to that CEO pulling his own numbers, and the leader publishing the sunset stats. Neither was sabotaged. Neither used a broken tool. They did exactly what the technology now lets them do — the work, themselves, fast — and the failure walked in through the gap where a second set of eyes used to stand. The bottleneck that got removed wasn’t slowness. It was the person who asked where the number came from.
What This Costs You, Specifically
If you publish marketing content, the stats in your live creative are a liability surface you’ve probably never audited. Some of them are dead. You inherited them the way everyone did — copied from a competitor who copied them from a roundup that copied them from a 2010 white paper nobody read. The difference now is that AI search is grading your content partly on whether the data in it holds up, and “we’ve always cited 98%” is not a defense when the author of 98% retired it on the public record.
Here’s the move, and it’s the rare one where defense and offense are the same action.
Pull the statistics in your current creative — site, ads, decks, all of it. For each one, find the actual origin, not the page you copied it from. Click through until you hit a primary source or hit a wall. The ones that dead-end at a vendor blog citing another vendor blog, or at a white paper old enough to drive, get retired. Not softened — retired.
Then replace them with something nobody can walk back: your own data. First-party numbers from your own platform, your own campaigns, your own results. You can’t be debunked on a number you measured yourself — and, the offense half, original verifiable data is exactly what AI search engines prefer to cite. The same audit that de-risks your content makes it more visible. Cleaning the archive and winning the citation turn out to be one motion, not two.
Doing your own work isn’t the mistake. The mistake is assuming the tool that handed you the output also handed you the judgment that used to come with it. It didn’t. That part is still yours — it just stopped being automatic.
Just because you can do it all yourself now doesn’t mean the checking did itself too.
Sources and Further Reading
- Nick Lane, “SMS engagement dropping, still outperforms other channels,” Mobilesquared, 2022 (linked via Internet Archive — the live page was intermittently unreachable at publication; archived capture used for a stable citation)
- Gartner, “Tap Into the Marketing Power of SMS” (Chris Pemberton, drawing on the Gartner 2016 Digital Channel Survey) — full article gated to Gartner clients
- Ryan Law, “Do AI Assistants Prefer to Cite Fresh Content?,” Ahrefs (17-million-citation freshness analysis)
Leave a Reply