Cimin Ahmadi Cohen, founder and CEO of Idea Peddler and adjunct professor at Texas State, hired an investigative journalist to figure out why ROAS kept climbing while business outcomes stayed flat. The answer has a name now: the attribution illusion.
A few years ago, Cimin Ahmadi Cohen noticed something that didn’t add up. Return on ad spend kept climbing. Business outcomes didn’t. As she put it, “the maths weren’t mathing.” That gap between a metric going up and up while revenue stayed flat is what sent her and her team at Idea Peddler down a rabbit hole deep enough that they eventually brought in a New York Times investigative journalist to help make sense of it. What they found has a name now, the attribution illusion, and it’s a problem far bigger than any one campaign or platform.
What is the “attribution illusion” and why is ROAS decoupling from business outcomes?
Ahmadi Cohen traces the disconnect to a confluence of events happening at once: shrinking panel data, collapsing match rates, and mounting state-by-state privacy legislation that quietly gutted one-to-one audience matching. The moment that connected the dots for her was WARC’s publication of “The Multiplier Effect,” which pulled together 30-plus years of ad effectiveness research and found that audience segmentation, the thing the entire attribution industry is built around, is actually one of the smallest drivers of brand growth and profitability, dwarfed by channel allocation, geography, and creative.
“We noticed a couple years ago there started, or stopped, rather, being a correlation between ROAS and the business outcomes… The maths weren’t mathing. We couldn’t make sense of how return on ad spend could go up and not have the business revenues truly go up.”
— Cimin Ahmadi Cohen, Founder & CEO, Idea Peddler
The scale of the data decay is what makes this more than a theoretical concern. Using CM360 as their benchmark, Idea Peddler’s first-party research found match rates fall from 75% to 25% in just two years.
“If Campaign Manager 360 is having that hard of a time with match rates when they have all the data in the world at their fingertips, that there are many, many providers out there who, when they say they’re doing a precision targeting segment, you should ask deeper questions about where they’re getting that data.”
— Cimin Ahmadi Cohen, Founder & CEO, Idea Peddler
Why are so many “precision audiences” actually filled with junk data?
This is the part of the conversation that should make any brand rethink what they’re buying when they buy a segment. Ahmadi Cohen calls it a “junk food problem.” When a targeted audience can’t be fully matched to real, addressable people, platforms don’t tell you that. They just fill the gap with lookalike and filler data to hit the promised audience size.
David connected this directly to a mechanism BDEX has built technology specifically to catch: Apple’s App Tracking Transparency and similar “do not track” signals cause a user’s mobile ID to refresh repeatedly, generating many IDs tied to one real person, none of them addressable. Those orphaned IDs still get folded into audience segments, inflating the apparent size of an audience that a platform can actually reach only a fraction of.
“You end up having many IDs linked to a single consumer, but none of them are addressable… you have this audience that looks like it’s 10 million people that you can target, but the reality is there’s a very small percentage of them that actually allowed the app to track… So what do they do? They start filling it in with other, more generic users.”
— David Finkelstein, Founder & CEO, BDEX
Ahmadi Cohen backed that up with a concrete number from her own client work: when analyzing one audience-syndication client’s data, her team found 75% of the audience wasn’t addressable at all, due to ATT and Do Not Track opt-outs.
“75% of their audiences were not addressable. So what happens in that case? Someone says they want to use this audience, 75% of it isn’t addressable. So what’s the platform going to do? They’re going to fill it with garbage inventory.”
— Cimin Ahmadi Cohen, Founder & CEO, Idea Peddler
Can AI actually fix bad targeting data, or does it just scale the problem faster?
Ahmadi Cohen is openly skeptical that AI is a fix for any of this. Her concern is the opposite: agentic AI layered on top of dirty data just scales the dirt faster and with more confidence.
“I remain a little bit suspicious until proven otherwise that AI is going to be a panacea for ad targeting problems. We all know the rule of garbage in, garbage out… Agentic AI being brought in to scale bad data feels like the most likely outcome over the next 24 to 36 months.”
— Cimin Ahmadi Cohen, Founder & CEO, Idea Peddler
David’s framing here is a useful distinction: most of today’s ad tech isn’t AI-first, it’s AI-as-plugin, bolted onto existing infrastructure rather than rebuilt around it. He estimated the industry is still years away from platforms that are genuinely AI-native, where targeting could learn from behavior in real time without needing individual-level identity at all, which would actually be more privacy-compliant, not less.
“Today AI is like a plug-in for everything… when it’s AI first, it can be 100% privacy compliant. The AI can actually learn what are you clicking on, and it doesn’t really need to know who you are.”
— David Finkelstein, Founder & CEO, BDEX
Both agreed on why the incentives cut the wrong way: platforms make money on volume, and cleaning bot traffic and click farms out of inventory shrinks the numbers that get platforms paid. Ahmadi Cohen’s team maintains a blacklist of mobile IDs, hashed emails, and IP addresses tied to bots and click farms, and it grows by the hundreds of millions of entries every month.
“That’s how they’re driving billions of dollars in click fraud, ad fraud. That gives you an idea of the scale of the problem.”
— Cimin Ahmadi Cohen, Founder & CEO, Idea Peddler
What actually counts as “decision-grade” data versus just directional data?
Ahmadi Cohen’s rule here is simple and non-negotiable: decision-grade data requires at least two independent sources, and at least one of them has to come from a party that doesn’t sell media, because any platform grading its own homework will always give itself an A.
“Decision grade data, we stand firmly that should be at least two sources. Everyone grading their own homework will obviously give themselves an A. So having some unbiased source… that can be an MMM, but it can also just be qual or quant studies.”
— Cimin Ahmadi Cohen, Founder & CEO, Idea Peddler
For lookalike audiences specifically, which she flagged as a black box whose inputs you can’t inspect, her approach is to build in hypotheses you can actually test, like comparing a Texas lookalike against a California lookalike to isolate what’s actually driving performance rather than accepting the platform’s self-reported win.
David added a warning of his own about one specific default: Meta’s audience extension, which is turned on by campaigns automatically unless a marketer knows to disable it.
“There are a lot of flaws in Meta’s lookalike audiences, their audience extension… we were able to tell clients exactly who they were targeting, and it was not the audience they were trying to reach at all. And that’s the other thing, it’s turned on by default.”
— David Finkelstein, Founder & CEO, BDEX
The bigger picture
What makes the attribution illusion worth taking seriously isn’t any single data point. It’s how many independent failures are pointing the same direction at once: shrinking panels, collapsing match rates, privacy law fragmenting what can legally be tracked, and platforms with zero financial incentive to clean up the inventory sitting behind their own numbers. Layer agentic AI on top of that stack before fixing the underlying data, and you’re not solving the problem. You’re just running the same bad decisions faster and with more confidence behind them.
Ahmadi Cohen’s closing advice is the right note to end on: this isn’t a problem any single marketer fixes overnight, and it doesn’t need to be. Acknowledge that the illusion is real, insist on a second, unbiased source before calling any number decision-grade, and chip away at the rest bit by bit. That’s a far more honest starting point than trusting a dashboard that’s quietly been telling you a story that stopped being true a while ago.
Connect with Cimin Ahmadi Cohen on LinkedIn or visit ideapeddler.com to learn more.
Watch the full episode on YouTube.
This article was adapted from an episode of Deconstructing Data, BDEX’s weekly podcast on data-driven marketing. Tune in live every Thursday at 4:15 PM Eastern on LinkedIn.
About BDEX: For companies that need clean identity data to power their products, BDEX offers unmatched quality and execution. Visit bdex.com and click “Talk to an Expert” to get started.
