Here is the promise the industry has been selling for fifteen years: pay us more, and we will waste less. Precision. Targeting. Identity. Supply-path optimization. Layer on the data, buy the audience, bolt on the graph, and every dollar lands on exactly the right eyeball at exactly the right moment.

It is a beautiful pitch. It is also, increasingly, a math problem the pitch loses.

The uncomfortable finding of the last decade of actual research, the peer-reviewed kind, the regulator kind, the kind where someone turned the ads off and measured what happened, is that a lot of "sophistication" in ad tech is a cost you pay for the privilege of reaching fewer people, less efficiently, and then crediting the machine for sales that were going to happen anyway.

You do not need an econometrics degree to see it. You need one equation and the willingness to run it. Let us do that.

Start with the least controversial fact in the entire supply chain, because a government spent a year proving it.

When the CMA traced a pound of advertiser money through the open-web programmatic chain, and cross-checked it against the ISBA/PwC supply-chain study, the answer was brutal in its simplicity: publishers received roughly 51% of what advertisers paid. Intermediaries kept the other 49%. And about 15% of the total could not be attributed to any named party at all. The ISBA auditors gave that missing money a wonderfully honest name: the "unknown delta." Money went in. Nobody could say where it went. (The CMA's own more conservative estimate still put the take at around 35%. Pick your number; both are a tax.)

So before your campaign buys a single impression, somewhere between a third and a half of the budget has evaporated into DSP fees, SSP fees, data fees, verification fees, ad-serving fees, and the aforementioned money that simply vanished.

Now the equation. If you spend a budget S, and only a fraction p survives fees to become working media, and each dollar of working media returns R dollars of revenue, then:

ROI = (R × p) − 1

That is it. That is the whole trick. Watch what p does to it.

Say your underlying media is genuinely good and returns $3 for every $1 of working media.

Same media. Same audience. Same creative. The only thing that changed was how much of the dollar the middlemen kept, and your ROI got cut in less than half.

Here is the part that should end certain sales meetings. To claw that 170% back while surrendering 40% to fees, your fancy stack would have to lift revenue per working-media dollar from $3.00 to $4.50.

READ THIS PART AGAIN: That is a 50% improvement in real performance, demanded just to break even with a simpler, cheaper setup that did nothing clever at all.

The adtech layer does not have to be bad at its job to destroy your ROI. It just has to be not 50% better, while charging like it is. That is a bar almost nothing clears.

"Fine," says the optimizer, "but precision earns its keep. The narrow audience converts better." Sometimes true. Rarely true enough. And there is a clean inequality that tells you exactly how much better it has to be.

Cost to acquire a customer is roughly:

CAC = CPM ÷ (1000 × conversion rate)

For a narrow audience to beat a broad one on CAC, the algebra collapses to one demand:

Required conversion lift ≥ the CPM premium you paid to get narrow.

In plain English: if your precise, data-enriched, audience-stacked segment costs twice the CPM of a broad buy, it has to convert twice as well just to tie. Not to win. To tie.

How often does doubling the CPM actually double conversion in the wild?

Almost never. A peer-reviewed 2024 study (Ahmadi, Abou Nabout, Skiera, Maleki and Fladenhofer, in the International Journal of Research in Marketing) built exactly this break-even model and ran it against real segments on the Spotify ad platform. The result: roughly half of the audience segments would need click-through rate to double to beat simply running untargeted. And the narrower you go, the worse it gets. Segments reaching 5% or less of the audience needed lifts north of 150% to justify themselves. Their phrase for it: a lift "that is likely not attainable."

The narrowness is not the feature. The narrowness is the bill.

Practitioners say it less politely. "The biggest misconception is that precision is free," says a senior vice president at a major marketing company, who spoke on background. "Every targeting condition you stack on a campaign raises the CPM while shrinking the audience. Marketers treat the first as a quality signal and never measure the second at all. Precision has a price tag, and most media plans never show it."

And it compounds with the previous section, because the study found the same thing the fee data did: when data quality dropped (they used Apple's App Tracking Transparency rollout as the natural experiment), the narrow segments got hit hardest, with higher CPMs and lower CTRs than broad. You paid more for precision precisely as the precision degraded.

Numbers make it real. Two campaigns, same $10,000, one broad and one "smart."

The narrow audience has a better conversion rate. It loses anyway. The CPM doubled while conversion only rose 1.6x, so broad buys more impressions, produces more total conversions, and lands a lower cost per customer. Exactly what the inequality predicted: you needed 2.0% conversion to tie, you got 1.6%, you lost.

Now add the real world, where "smart" also means extra data and identity fees. Skim a modest 20% off the narrow budget for the privilege:

Same headline spend. The broad campaign now acquires customers at 64 cents on the narrow campaign's dollar. Three separate drags stacked up and buried the "smart" buy: a higher CPM, less scale to learn from, and a fee haircut on top. This is not a freak result. It is the default result.

There is one more leak, and it is the quietest, because you pay for it whether it works or not.

Precision targeting assumes the machine actually knows who it is talking to. Often it does not.

An academic audit of third-party data providers (Neumann, Tucker and Whitfield, 2019) found that the best provider put ads in front of the intended target market about 70% of the time. The worst managed about 40%. So a "25 to 54 year old in-market auto intender" segment is, on a bad day, wrong most of the time. You are buying a premium audience and receiving a coin flip.

Match rates make it worse. When your targeting depends on identity resolution that only fires on a fraction of impressions, you still pay the enriched CPM and the data fee on all of them. Legacy cookie match rates have been reported in the 31% to 45% range. So a large share of your spend buys impressions the targeting cannot even act on, at a price that assumes it can.

Put the fee math and the match math together and you get the whole grim engine: you pay a premium on 100% of impressions so the machine can be smart on maybe a third of them, about a coin flip of those are even the right person, and then the "unknown delta" quietly eats another slice on the way out.

Judy Shapiro, a longtime marketing-technology CEO and company co-founder, puts the data half of that plainly: "Not all data is created equally. Bad big data is worse than no data." A premium segment built on wrong inputs does not merely fail to help. It actively points your budget at the wrong people while you pay extra for the confidence of having aimed.

Now the study that should be tattooed on the inside of every CMO's eyelids.

In 2015, three economists (Blake, Nosko and Tadelis) published in Econometrica the results of turning eBay's paid search ads off across dozens of US markets and watching what happened. For branded keywords, the measurable short-term benefit was, essentially, zero. When people searched "eBay," they found eBay whether or not eBay paid for the link on top. Organic search was a near-perfect substitute for the thing eBay was buying.

The kicker feeds everything above. The heavy spenders, the frequent users, were the ones least moved by the ads, and they accounted for most of the spend. The ads mostly reached people already walking through the door and then took credit for the walk. Average returns came out negative.

Now the honest counterweight, because a con artist cherry-picks and a journalist does not. Two years later, a similar shut-off experiment at Edmunds.com (Coviello, Gneezy and Götte) found the opposite: more than half the paid traffic genuinely disappeared when the ads went dark. So the real lesson is not "ads never work." It is that incrementality is wildly heterogeneous, and the dashboard cannot tell the difference between a sale you caused and a sale you photobombed. The targeting that looks most impressive, the retargeting and branded and re-engagement stuff, is precisely the targeting most likely to be standing next to a conversion that was going to happen anyway, pointing at it, yelling "I did that."

The SVP quoted earlier draws the line in one breath: "Attribution tells you which touchpoint was standing nearby when the sale happened. Incrementality tells you whether the sale would have happened anyway." Then the part that should chill anyone paying a performance fee: a great deal of what gets sold as optimization is, in this person's words, "teaching an algorithm to stand closer to people who were already going to buy, then charging a performance fee for it."

Here is where honesty requires a change in tone, because this is the part the industry's critics (us included) most want to be true, and the part with the least gold-standard proof.

The theory is clean. If your algorithm optimizes to the cheapest, fastest conversion, it will learn to chase the cheapest, fastest converters: deal-hunters, coupon-clippers, cash-back shoppers, the promo-sensitive. Those cohorts plausibly carry lower margins and higher churn. So the machine that looks like a genius on a cost-per-acquisition dashboard could be quietly stocking your customer base with your least valuable customers.

The mechanism is real and directionally supported. But let us be precise about the evidence: the fee math is documented, the incrementality math is documented, and the "narrow costs more for less" math is documented. The specific claim that targeted-acquisition cohorts have structurally lower lifetime value than broad-reach cohorts is the softest link, better supported by practitioner logic than by a stack of randomized controlled trials. We flag it as a strong hypothesis, not a closed case.

Which brings us to the cleanest available illustration, and a fair one, because the company more or less says the quiet part out loud.

Take tvScientific, a performance CTV platform that sells television advertising on a cost-per-action, "you only pay for outcomes" basis, and integrates tightly with affiliate and partner ecosystems (Partnerize and the coupon/loyalty publisher world). We are not singling them out because they are the worst. We are singling them out because their own public model is the mechanism. When you bolt a CPA-only commercial model onto affiliate and coupon rails and tell the optimizer to chase whatever fires most reliably in the tracking system, it will drift, inevitably, toward the households already deepest in the funnel: the coupon-site users, the cash-back crowd, the people about to buy regardless. The platform then matches a prior TV impression to that conversion and bills you for "TV performance."

It looks spectacular in a case study. Low CPA. High ROAS. And it is very possibly the least incremental, lowest-LTV traffic in the entire chain, dressed up as your best channel.

Nothing in a CPA-only model forces the question that actually matters: what was the lift against a holdout, and what are these customers worth in a year? The measurement floor is being sold as the ceiling.

If everything above sounds like the paranoid arithmetic of one skeptical publisher, consider the tell: the market is now spending real money to agree with it.

On 7 July 2026, the ad-intelligence firm Guideline announced it had extended its dataset, roughly $200 billion in annual media investment across 65 countries, to capture verified, transaction-level advertising activity on the new AI platforms. The pitch, stripped of the press-release varnish, is the most quietly damning sentence in the industry this year: buy our data to see what is actually happening, "not what platforms say is happening." The context is that OpenAI has projected $2.5 billion in ChatGPT ad revenue this year, rising to $11 billion next. Those are the platform's own numbers. Guideline just built the machine to check them.

The same week, Integral Ad Science installed a new CEO, Lidiane Jones, out of Bumble and Slack, whose stated ambition is to make IAS "a very critical part of the trust infrastructure of this AI era." IAS has spent recent months shipping tools that explain campaign performance "without the black box" and that detect and block AI slop before it runs next to your ads. So we now have AI tools policing other AI tools' output before a third AI tool optimizes your budget into it.

Sit with what that whole scramble means. The fastest-growing product category in advertising is independent verification of what the machines already told you. That is not a market that trusts its own instruments. That is a marriage where somebody has hired a private detective.

And here is the turn, the one the verification vendors will not say out loud because they are selling the meter: auditing the invoice is not the same as auditing the incrementality. Guideline can prove, to the penny, that your dollars really flowed into ChatGPT inventory and that OpenAI's revenue is or is not what it claims. Genuinely useful. But it is a completely different question from the one this entire article is about, which is whether those dollars caused anything. You can hold a beautifully verified, transaction-level record of every cent you spent, and still have bought a pile of the exact photobombed, coupon-adjacent, going-to-happen-anyway conversions from Parts five and six. The new industry verifies the receipt. It does not verify the lift.

And there is a subtler trap underneath the whole verification pitch, which is that the readouts themselves are less objective than they look. Judy Shapiro warns that data is "both liberating and a prison of a brand's own making," and that the dashboards carry "a false sense of certainty" when they are really "a reflection of how the data engineers decided to tell the data story." What got measured, how the attribution equation was written, which sources were trusted and in what context: every one of those is a subjective call made by someone you will never meet. A verified number is still a number somebody chose how to build.

Which lands us back where we started, one layer higher. The verification wave is subject to the same law as the targeting wave it is auditing: it is another intermediary, taking another cut, grading adjacent homework, often paid by the very ecosystem it certifies. It is turtles all the way down, and every turtle bills monthly. Verification is worth paying for. Just apply the same cold question to your meter-reader that you should have applied to your targeting all along: what, exactly, does it prove, and who pays it to prove that and not something else.

String it together and the picture is not "ad tech is a scam." It is worse, because it is subtler. Ad tech frequently charges for complexity, while the delivery system underneath leaks value through fees, identity loss, coin-flip data, and credit-grabbing attribution. Every layer is individually defensible. Stacked, they demand a level of real incremental lift that the evidence says almost nothing delivers.

The precision in the pitch is not the precision in the outcome. More targeting buys you a smaller, more expensive audience, resolved by data that is right about half the time, measured by a system that cannot tell causation from coincidence, after a third to a half of your money already left the building.

And when the machine's own referees show up to sell you a second opinion, remember that most of them are auditing the invoice, not the incrementality. The receipt can be true and the lift can still be zero.

Judy Shapiro traces the whole mess back to its funding. The ecosystem got this baroque, she told ADOTAT, because "a lot of VC money was chasing ventures, any venture," each one selling marketers one more SaaS layer, producing a stack ordered so the venture's financial interest comes first, the investor's second, "and brand last, dead last."

Her prescription is Confucian: "Life is simple, but we insist on making it complicated." Marketing, she says, reduces to one job, reach prospects who could actually buy the thing, and the machinery we built to do that is overkill dressed as sophistication.

Or, as the SVP put it, compressing the entire argument into a single line: "Any layer in the supply chain that can't show its lift exceeds its fee isn't technology. It's a tax."

The math is not hidden. It is just inconvenient. Run it before your next planning meeting. You're welcome.

About the author

Pesach Lattin is an investigative journalist and the publisher of ADOTAT. His beat is ad tech fraud, programmatic economics, and holding-company structures. Confidential tips welcome: [email protected]