On Sunday I published that piece about how programmatic advertising might be, and I chose the word carefully, a criminal enterprise. The $100+ billion fraud layer. Agencies benefiting from opacity. RICO. The FTC reads this newsletter. The SEC reads this newsletter. State AGs read this newsletter. Hi again, everyone.

I've been circling this all week. The migraine hasn't gone away. It just moved.

Because the fraud layer, the bots, the MFA sites, the supply chain grift, that's the obvious crime. That's the part where a prosecutor with log-level data and subpoena power puts people in rooms with bad lighting. We'll get there in this series.

But there's a quieter scam running underneath all of it. One that doesn't need bots. Doesn't need fake sites. Doesn't even need anyone to technically lie. It just needs one word.

"Outcomes."

Let me be precise about what happened here, because the language matters.

The industry spent the last several years selling "outcome-based" buying as the next evolution of programmatic. The pitch was gorgeous: stop paying for impressions, start paying for results. Brand lift. Conversions. Store visits. Purchase intent. Let the machine optimize to what actually matters.

And in the last few months, agency after agency started pulling this apart. They all found the same thing.

The machine isn't optimizing for your business. It's optimizing for the easiest possible signal that looks like your business.

Those are two extraordinarily different things.

Not that we don't want results. Obviously we want results. But "outcomes" as the programmatic industry has been selling it isn't actually results. It's fake results. It's the number the machine can produce cheapest. For many people, for many networks, "outcomes" is a lie that looks like a metric.

Here's the simplest way I can explain what went wrong.

A scammer does not want the skeptic. A scammer does not want the person who reads carefully, asks questions, checks references. A scammer wants the person who clicks the link, fills in the form, and doesn't think too hard. The lowest common denominator. The easiest mark. That's the entire business model of fraud: find the person with the least resistance and exploit that.

A DSP optimizing for "outcomes" is running the same playbook. Exactly the same playbook. Just with better PR and a Looker dashboard.

If you distribute a million ads to a broad set of people who are potentially in-market for your product, you get a real, useful cross-section of humans. You can measure brand lift against that population and learn something true about whether your campaign moved the needle. That's good. That's what measurement is supposed to do.

But if you hand a DSP a "brand lift score" as an optimization target and say "go maximize this," the algorithm does what algorithms do. It finds the cheapest path to the number you asked for. It stops looking for people who might actually buy your product. It starts looking for people who will answer a survey positively.

Professional survey-takers. Reward-seekers. People with time on their hands who click "yes" on everything. The algorithm didn't break. You told it to find the lowest common denominator, and it did.

The scammer finds the easiest mark. The DSP finds the easiest survey responder. Neither of them is finding your customer.

This is the part most people don't think about, and it's the key to the entire scam.

People who happily fill out every brand-lift study they encounter are a distinct, identifiable segment. They are not your customers. They are not representative of your audience. They are a specific type of person who over-indexes on reward-seeking and time-rich behaviors. They remember every ad, not because your creative was brilliant, but because noticing and recalling ads is literally what these people do for rewards.

Hand the bidder a "lift score" target and it will learn, very quickly, to chase these users. Not because they'll ever buy your product. Because they'll tell a survey they might.

Without any correction, brand-lift models treat these people as "highly influenced" by your ads. So they get more impressions. Their lookalikes become your "best" audience. The machine builds an entire delivery strategy around people whose only talent is filling out surveys.

So your brand-lift numbers go up. Your CMO puts it in the board deck. And your sales don't move. Your site traffic doesn't change. Your add-to-cart rate is flat. But hey, "awareness" is up 12 points among people who were already aware of every ad on the internet.

The survey-taker problem is the clearest example, but the same pathology infects every type of outcome selling in programmatic. The machine always finds the cheapest version of whatever you asked for.

Pay-for-conversions? The vendor optimizes to users who were going to convert anyway and claims credit via last-touch attribution. You pay for conversions you already owned.

Guaranteed ROAS? The pressure to hit the number pushes buying into garbage inventory where fraud and manufactured signals are cheapest. Your "outcomes" are partly synthetic.

CPC optimization? The bidder finds the cheapest clicks, which live on made-for-advertising sites and accidental-click placements. Your efficiency metric improves while your actual reach collapses into nothing.

View-through conversions? The algorithm learns to serve one cheap impression to someone already on your site and then bill you for the sale. That's not advertising. That's invoicing.

In every single case, the pattern is identical: you give the machine a metric, the machine finds the cheapest way to produce that metric, and the cheapest way is almost never the way that grows your business. The algo isn't broken. The incentive is broken.

Marc Guldimann put it right: paying for outcomes is a great way to get the worst possible version of that outcome. Platforms learn to manufacture easy, low-quality signals and resell that capability. Over time, they get better at producing fake-looking-real outcomes than you are at detecting them. And they use that information asymmetry to set terms, keep margins, and make you feel like you're de-risking spend when you're actually just paying more for less.

"Outcomes" as a buying model is dead. Not because we don't want results. We desperately want results. But because the word got hijacked by a system that is structurally incapable of delivering them honestly without serious guardrails that almost nobody was putting in place.

On Sunday I wrote that I'm going to start treating parts of this ecosystem the way my training tells me I should. Like a possible criminal enterprise. The "outcomes" scam isn't the bots-and-MFA layer. It's something almost worse. It's a system that tells you it's working, shows you the numbers to prove it, and never technically lies, because the numbers are real. They're just meaningless.

The question is: what do you actually do instead?

That's Part 2. And it requires understanding some things about propensity scores, incrementality testing, and why the only honest way to measure advertising is to occasionally stop advertising and see what happens.

Continue reading with Adotat+ →