Proof Is the New Flex — And It Has a Unit.

This week, the largest stage in marketing will spend five days agreeing that proof matters. Not one keynote will say what proof is per. That blank — the missing denominator — is the whole argument.

The pre-festival consensus for Cannes Lions 2026 has already hardened into a single line: the AI-novelty era is over, and "proof is the new flex." The festival even rebuilt its awards around it — new AI Craft subcategories that judge work on whether the AI produced a result prior methods couldn't, scored on craft, intent, and human contribution. Simply using an AI tool no longer wins anything. After three years of "look what the model can do," the industry has decided that the flex is evidence.

Good. That premise is right, and it is the premise this site was built on — the Lab exists because the field runs on takes when it should run on proof. So let me concede the headline in the first sentence and move past it: proof matters; everyone agrees; agreeing is worthless. The entire industry nodding at "effectiveness is back" adds nothing, because the hard question isn't whether to prove. It's what proof is measured against. And on that question, the Croisette is silent.

Proof is a ratio. The industry forgot the bottom.

Here is the thing almost no one says out loud: proof is not a number. It's a ratio. Every claim of effectiveness is some result divided by some unit — conversions per dollar, revenue per impression, lift per campaign. The result is the numerator. The unit you divide by is the denominator. And our field has a numerator obsession so total that it forgets the denominator exists.

Watch how the words behave. ROAS is revenue over spend. Incrementality is conversions over the counterfactual. CAC is cost over a new customer. Effectiveness, the word the whole festival is built on this year, is a result over… nothing specified. It floats. That's why three teams can each show their "effectiveness" is up while the business is exactly where it was — they're quoting numerators and waving at a denominator nobody agreed on. A numerator with no denominator isn't proof. It's a boast with a chart.

So name the denominator, and do it early, because everything downstream depends on it. The unit of marketing proof is the decision. Not the click, not the impression, not the campaign — the decision: the discrete choice your spend is actually buying. A decision to approve a loan applicant. A decision to enroll a member. A decision to add to a cart. Marketing's job has always been to manufacture decisions at a cost the business can bear, and once you say that plainly, the denominator that's been missing from "effectiveness" snaps into place. Cost Per Decision is simply what "proof is the new flex" sounds like once you make it countable.

What proof actually means — for one paragraph, rigorously

It's worth being exact about the word everyone's about to overuse, because the rigorous definition does real work here. In any measurement discipline, proof is a number you cannot improve by wanting it more. It survives a held-out test you didn't get to tune on. It survives a counterfactual — what would have happened anyway. It survives an adversary who's motivated to find the flaw. A case-study lift figure, generated and reported by the same system that ran the campaign, survives none of those; it's a number that gets better precisely because someone wanted it to. That's the test the AI Craft jury is gesturing at without naming, and it's the test a decision denominator passes and a self-graded "effectiveness" number fails — because the count of decisions, and what each one cost, can be observed independently of the system optimizing them. Proof you can move by trying harder was never proof.

"But we already have denominators"

The strongest objection to all of this comes from the people who know measurement best, and it deserves a real answer rather than a strawman. An effectiveness veteran — someone who's run geo-holdouts and marginal-ROAS curves for twenty years — will say: we already have denominators. Incremental conversions. Marginal return on the next dollar. Profit per acquisition. Econometrics solved "result per unit" before you were born. "Cost Per Decision" is a rebrand.

And where you have clean incrementality and honest marginal ROAS, that veteran is right — use them, they're excellent, and nothing here replaces them. But notice the quiet assumption underneath those denominators: that a knowable number of human decisions sit behind the spend, each attributable, each countable after the fact. That assumption is what just broke. In an agentic stack, the decisions don't sit still to be counted. Every bid is a decision. Every signal weighting is a decision. Every creative selection, every audience expansion, every budget shift the system makes on its own is a decision — made thousands of times a second, inside systems that also report their own effectiveness. Classical denominators measure the outcome of a few human choices. The decision denominator measures the cost and quality of the decision-making itself, including the machine's — which is the layer that just exploded in volume and the layer the optimizer is now also grading. That's not a rebrand of marginal ROAS. It's the denominator for the part of the system marginal ROAS was never built to see.

Make it concrete. A Performance Max campaign doesn't make one decision you can audit at quarter-end; it makes millions. It decides, per auction, which audience signal to weight, which of a thousand asset combinations to serve, how much to bid, where to place. Marginal ROAS can tell you the last dollar earned its keep. It cannot tell you whether the decisions producing that dollar were good ones, cheap ones, or fast ones — because those decisions never surface as countable line items. They're absorbed into the optimizer and reported back as a single outcome number. The decision denominator is what makes that invisible layer measurable again: how many decisions did this spend buy, what did each cost, and how quickly did the system act on a new signal. That's a different question than "what was the return," and in an agentic stack it's the more important one — because the return is increasingly a number the optimizer computes about its own work.

This is the same fault line I drew in You Can't Let the Optimizer Own the Evidence Layer: the moment the system that makes the decisions also certifies their results, "proof" quietly becomes self-report. The industry wants the optimizer's brain to also be the auditor's brain — and now it wants that same brain to hand out the effectiveness grades. Proof, in that arrangement, is just the word we say right before we give the grader the answer key. Qualified Future Conversions is the live example: a forecast, generated by the optimizer, booked as the result it was supposed to prove.

Why the machine makes the denominator non-optional

Step back and look at the full production line as it exists in 2026. AI now generates the creative at infinite scale. AI runs the auction. AI selects the audience, sets the bid, and — through future-tense conversion credit and the new cross-product advisor agents — increasingly reports on how well it did. When the same system makes the work, places the work, and grades the work, "we proved it was effective" can mean one of two completely different things. It can mean the platform's own scorecard went up. Or it can mean an independent count of decisions improved — more carts at a lower cost per cart, more funded accounts per dollar, the same enrollments delivered with less decision latency.

Only the second one is proof in the sense the jury claims to want, because only the second one has a denominator the optimizer can't touch. This is why "name your decision denominator" isn't a philosophy seminar — it's the operational difference between a marketing team that can verify its own results and one that can only relay the platform's. As the surfaces consolidate — the sealed auction, the platform-owned cart, the unified advisor that spends and reports in one breath — the share of your "effectiveness" that is independently observable shrinks every quarter. The denominator is the part you keep.

The decision-denominator worksheet

Here's the Lab version — proof, not takes — and it fits on an index card. Before you quote another effectiveness number, do four things, this quarter, for one campaign:

One: name the unit of decision you're actually buying. Not "conversions" in the abstract — the specific decision your P&L pays for. A member-year for Medicare Advantage. A funded account for a neobank. A cart for retail. Write it down. If your team can't agree on the unit, that disagreement is the finding — you've been optimizing a numerator nobody could name the bottom of.

Two: measure cost per decision and decision latency for that unit, from your own data. Not the platform's modeled column — your server-side, first-party count of decisions delivered and what each one cost, plus the time from signal to action. This is the denominator you own.

Three: place it next to the platform's "effectiveness" number for the same period. The gap between your independent decision count and the optimizer's self-reported lift is the single most useful number you'll carry into a budget conversation all year. It's the size of the optimizer's thumb on the scale, measured.

Four: do it before the surfaces close. Every quarter you wait, more of the funnel moves behind a layer that reports its own grade, and the independent denominator gets harder to reconstruct. You can opt into the future-tense metrics; you cannot retrofit a clean decision denominator onto a quarter you already measured without one. That's the one move here that doesn't survive being deferred.

What this looks like from the CFO chair

Strip the festival vocabulary and the transaction is simple. The vendor that sells you the media is increasingly the same system that produces the "proof" the media worked. In any other procurement category, a CFO would not accept the supplier's self-scored effectiveness as the evidence the spend was justified — they'd want a number the supplier can't move. Marketing has been the exception, and "proof is the new flex" is a chance to stop being the exception: not by chasing the festival's applause for proof, but by walking in with the denominator the applause forgot. The P&L translation chain — from a decision to CAC to enterprise value — only holds if the first link, the decision, is something you counted yourself.

Cannes will hand out Lions for proof this week, and most of the winners will deserve them. But a festival can declare that proof is the standard; it cannot tell you what proof is a ratio of. That part is yours, and it's the one piece of this whole cycle that doesn't expire on June 26: pick the decision your spend is buying, measure its cost and its latency yourself, and keep that denominator independent of the system grading the numerator. Everyone is about to start saying they have proof. The only ones who actually will are the ones who can name the bottom of the fraction.