I've been elbow-deep in probabilistic marketing research, and I hit one of those thoughts so obvious I assumed everyone already knew it. Then Scott McKinley, the CEO of Truthset, sat me down and walked me through how this stuff actually works, hop by hop, and I stopped assuming anybody knew anything.

We are not naming the bad guys today. Not because we're being polite, but because there are too many of them to name in one column, and picking one lets the other forty off the hook.

So call them Company A and Company B. What matters is not their logos. What matters is that their own marketing materials, if you do the arithmetic they're hoping you won't, quietly admit that using their product is worse than not using it at all. That's the story. The call is coming from inside the deck.

Start with the punchline, because it's the part nobody says out loud. A large amount of adtech is sold on numbers that, read correctly, are evidence against the product.

Company A just got bought for a few hundred million dollars, which everyone treated as proof the thing works, because that is how we think now: big check, must be real. Their pitch is genuinely gorgeous. Take a premium, person-level intent signal, the kind that tells you what someone is planning to buy, and pipe it onto the biggest screen in the house through what they call "proprietary, deterministic ID technology." The headline stats are the kind that make a CMO reach for a pen. A 27% lift in outcomes per dollar. A 65% lift in purchases.

Beautiful. Now watch the hand leave the table. Twenty-seven percent more than what? Sixty-five percent more than what? Because if the honest answer to "versus what" is "versus showing the ad to nobody," then they haven't measured their product. They've measured whether advertising exists. And if a dumb, cheap, untargeted campaign would have posted a bigger number from the same starting line, which in the peer-reviewed literature it very often does, then their own slide is a confession that the premium made things worse. They charged extra to lose you points. They should be wiring money to you.

Let me be precise about which lie is worst, because they stack, and the zero baseline is the foundation the whole con is poured on.

There are three things you could compare a targeted campaign against. No advertising at all (the zero baseline). An untargeted, spray-it-everywhere campaign at the same spend (the real competitor). And nothing, which is what most decks give you, because they just hand you the relative-lift number and let your imagination fill in a flattering denominator.

Company A and Company B and the other thirty-eight all reach for door number one, the zero baseline, and they reach for it because it is the single most favorable comparison that exists for any ad product on Earth. When your control group is people who saw no campaign, then every ounce of lift that comes from simply putting a creative in front of a human eyeball gets credited to your targeting layer. The targeting did not earn that. Photons on retinas earned that. You would post the same lift painting the ad on the side of a city bus and driving it past a bus stop.

That is why the zero baseline is the deepest version of the fraud, not the shallowest. A bad relative-lift number at least argues about size. The zero baseline cheats on the identity of the thing being measured. It swaps "did our targeting beat the alternative" for "does advertising do anything," pockets the difference, and hopes you clap.

The comparison that actually decides whether targeting is worth its premium is targeted reach versus untargeted reach, at the same spend. Not versus zero. Versus the cheap option sitting right next to it on the shelf.

That slide does not exist. It never makes the deck, and not out of shyness. It's that dumb, cheap, run-of-network buying frequently posts the bigger number. Your 30% is somebody else's 34% for the campaign that skipped the targeting tax. So the "lift" isn't a selling point. It's a signed admission that the fancy layer subtracted value. Read that sentence back to a CMO slowly and watch the color leave the room.

Here's the first thing Scott made me sit with, and it reframes everything downstream. "Deterministic" has an actual technical meaning, and it is not a mood. It means an authenticated, declared linkage: the same logged-in credential seen at exposure and again at conversion. Same verified human, both times, no guessing.

That is not what Company A is doing. Some of the individual link steps are deterministic-flavored joins (a login tied to an IP at one instant), but the inference riding on top is probabilistic at every layer. That this IP is this household. That this household holds this person. That this person is the one who bought. IPs rotate on carrier-grade NAT and DHCP leases. They're shared across roommates, apartment buildings, the coffee shop downstairs. They decay fast. And when Truthset actually measured it across six providers in the CIMM study, IP-to-household came in around 16% accurate and IP-to-email around 13%.

So sit with it. Calling a chain whose weakest link is roughly 15% accurate "deterministic" is not a terminology slip. It is borrowing the credibility of authenticated matching to describe inferred matching. It is probabilistic wearing a much better suit, and the suit is closing the deal.

This is where I started writing faster. Trace one honest intent-to-television buy, one hop at a time:

One. A logged-in platform ID. This is the single genuinely deterministic link in the whole chain. Enjoy it. It's the last one.

Two. That ID gets hashed and onboarded into an identity graph.

Three. The graph resolves it to a household IP. This is the CIMM-measured step. 13% to 16% accuracy. This is the load-bearing wall, and it's made of wet cardboard.

Four. The IP resolves to a specific TV device, which quietly assumes the IP was even stable at ad time.

Five. The ad plays and gets attributed to the household, not to any specific person inside it.

Six. A conversion fires later on a phone or a laptop and gets resolved back through another IP or a separate join.

Seven. Exposure and conversion get stitched together and stamped, triumphantly, "a conversion."

And here is the math that ruins the afternoon: errors compound multiplicatively. Be generous. Give every non-IP step a friendly 70% to 80%, and give the IP step the very top of the range at 16%. Multiply it through and the odds that the person who saw the ad is the person who bought land in the single digits to low teens.

The system reports a conversion anyway, because a database join happened. Whether that join corresponds to a real human doing a real thing is the unmeasured question. And here's the part Scott hit hardest, so I will too: false matches don't fail in a neutral direction. They inflate the numerator. A wrong household match still occasionally lands on someone who bought the thing on their own, and that accident gets filed as a win. The error doesn't cancel. It pads the stat. The worse the graph performs, the better the report can look. Let that sit.

Here's the internal contradiction that you can't unsee once you've seen it. The entire premium of Company A is person-level intent. The differentiator, the reason for the multiple, is that the signal reflects what a specific person is planning. That is what you're paying up for.

But an IP resolves to a household, at best. So the 26-year-old planning a wedding and her retired father-in-law watching a free streaming app on the same living-room Wi-Fi are, at the moment of delivery, the same target. Indistinguishable. The exquisite person-level signal gets averaged across everyone standing behind that router the second it hits the television. That is not a degradation. It's a category change in what the identifier means. You can honestly sell "households containing at least one high-intent user." You cannot honestly sell person-level precision through a household pipe. The pricing implies the second. The physics deliver the first. The gap between those is, roughly, the business model.

There's a smaller, dumber lie stacked under the big one, and it's arithmetic in a Halloween costume. Absolute lift is the raw gap: favorability moved from 10% to 13%, three points. Relative lift divides those three points by the baseline and hands you a scarier number: 30%. Same campaign, same data. One is forgettable. The other gets the big font.

Google's own documentation admits the metric is rigged this direction, warning that very large relative lifts happen easily for brands with low baseline activity in the control group. Translation: the smaller you start, the more heroic your percentage looks. A brand crawling from 2% to 3% brags about 50% relative lift. That's one lonely point of humanity dressed for the investor call. Executives hear "almost 20% improvement," the real move was six-tenths of a point, and Kahneman and Tversky built entire careers on exactly this reflex. Adtech built a sales motion on it.

Now the structural problem. Company A sells the audience, the media, and the measurement. When one entity controls all three, the comparison group is chosen by the party being graded. So when a 27% and a 65% land in front of you, the first question isn't "is it true," it's "who defined the counterfactual, and how weak was the baseline they picked for themselves?"

Scott's checklist, which I'm stealing whole: watch for comparison against a deliberately weak baseline campaign. Watch for selection effects, because the high-intent crowd was already likelier to buy, so part of the "lift" is just fishing in a stocked pond. Watch for survivorship, where the "early tests" you're shown are the winners of an unreported distribution. Watch for footnoted methodology with no methodology released. And watch the metric definition, because "outcomes per hundred dollars" is a platform-defined composite, not a standard anyone outside the building can audit.

None of it means the numbers are fabricated. It means they're unaudited claims from the entity with the strongest possible incentive to maximize them, scored through the very graph in question. The referee cannot also be the team. And when one of these companies proudly rolls out a "trust but verify" measurement program, remember: the only reason "verify" is in the sentence is that "trust" was never earned.

I want to be fair, because the vendors will push here and they're not wrong to. A properly randomized incrementality test genuinely controls for a lot. Randomize exposure at the household or geo level before the graph touches anything, and the aggregate lift between exposed and control is a valid causal estimate even with a garbage graph. Real thing. Give it its due.

But two problems survive, and Scott made me hold both. First, attenuation, not inflation: if the graph misassigns who was actually exposed, a clean test dilutes measured lift toward zero, which is a big reason platforms don't rush to run the clean version. Second, and this is the killer, inherited error when the graph defines the cells: if the same broken IP graph builds both your exposed and control groups (matched-market or synthetic-control designs instead of true pre-graph randomization), the contamination does not cleanly cancel. Exposure misclassification correlates with urban density, carrier NAT, and multi-dwelling units, and those correlate with purchase behavior. That is non-random error, and non-random error does not wash out. So the honest verdict: real randomization fixes the causal question but not the attribution question, and most commercial "incrementality" products are the second design, not the first. The question that separates the real ones from the theater: randomized where, exactly, relative to your graph? Ask it. Watch the face.

Flat on the table, the way Scott laid it out, because the split is the whole story.

Actually measured, all true: the platform has real logged-in intent data. Ads were delivered to devices. Conversions happened somewhere. The graph joined some exposures to some conversions as a database operation.

Taken purely on faith: that the household behind the IP is the household in the graph (around 13% to 16%, by Truthset's own measurement). That the intent-holder is the viewer. That the viewer is the buyer. That the lift reflects media effect and not audience selection. And that the scoring is impartial when seller, medium, and scorekeeper share one bank account.

The CMO is buying a real signal, through an unvalidated pipe, graded by the pipe's owner. The word "deterministic" quietly upgrades "we made some probabilistic joins" into "we know," and a few hundred million dollars just moved on the difference between those two sentences.

Here's the whole con on one card, the thing you carry into the next pitch. Any lift number without a denominator is a magic trick, not a measurement. When the big percentage flashes, you ask four questions in a row and you watch the room: Versus what. At what cost. With what holdout. Absolute or relative. Then, for the identity crowd, Scott's fifth: randomized where relative to your graph, and what's your validated match accuracy, not your match rate?

The honest vendors have all of it loaded and ready, because they built the three-cell test that isolates whether targeting beat plain reach, and they can show their accuracy without flinching. The ones running the con will suddenly develop an urgent need to circle back offline. Take it to a smaller room. Put a real expert on it next quarter.

That flinch is the product working exactly as designed. Just not the product they sold you.

We could name Company A. We could name Company B. We're not going to yet, because the more decks I read, the longer the list gets, and the point was never one villain. The point is that if you do the math they left out, a startling number of these products are selling you a bill for making your advertising worse, printed on a slide that says the opposite. Next week, we start doing the math out loud.