When the Machine Won't Say Why

The automated systems that now move the bulk of digital direct-response and e-commerce budgets (Google’s Performance Max, Meta’s Advantage+, with fully agentic buyers close behind) regularly produce numbers that look fine on a dashboard. What they will not produce is a reason. Ask why performance rose, and the answer is a shrug rendered as a user interface. Ask why it fell, and the answer, with remarkable consistency, is “spend more.”

Buyers call this the black box problem. I would call it the newest chapter of an older story: our industry's long habit of accepting what it can measure in place of what it wants to know, now running at machine speed.

But let us be precise about the machine’s jurisdiction. Current black boxes are execution tools, not planning platforms. They do not allocate across linear television, CTV, audio or out-of-home, and they do not divide budget between building a brand and harvesting its demand. Today the black box owns the bottom of the funnel. Yet agentic AI moves in one direction, up. Cross-channel planning engines are already reaching for CTV, audio, and digital video budgets, and buyers who cannot get a “why” out of direct-response spend today will be offered the same shrug with brand budgets tomorrow. Digital direct response is not the whole story. It is the preview.

In fairness, the top of the funnel is better defended than the bottom ever was. The walled gardens do not share log-level data, so no single algorithm sees the whole field. CFOs do not sign eight-figure commitments on “the model preferred it.” And the current renaissance in incrementality testing and marketing-mix modeling exists precisely because marketers decided to audit the algorithms from outside. Structure, money and method all resist total opacity.

We have been here before. For sixty years we bought age and sex because age and sex were what the currency could count, though no one ever sold a case of beer to an age bracket. We bought GRPs, then clicks, then viewability, then attention, each a proxy standing in for the question that has never changed: did the advertising cause someone to buy? The difference this time is that the old proxies were at least out in the open, where we could argue with them. The new ones are inside the machine, where we cannot.

For sixty years we accepted proxies for the question. The nearer danger is a proxy for the answer. The next generation of planning tools will not refuse to explain; they will explain beautifully, wrapping allocations in generated decks of personas, channel logic and strategic rationale, composed after the math and connected to it loosely, if at all. The shrug at least admitted what the machine did not know. A fluent, confident, wrong “why” is worse, because it disarms the skepticism the industry has just begun to organize.

Last week in this space I wrote about Anthropic’s research suggesting that its AI model, Claude, organizes its behavior around internal states resembling motivations. There is a second lesson in that work. The only reason anyone knows what is happening inside that model is that Anthropic built instruments to look, a discipline the AI field calls interpretability. Note the method: instruments, not testimony. Opacity is the natural condition of large learning systems. Transparency exists only where someone insists upon it and builds the tools to get it.

The AI labs are insisting, for reasons of safety. Advertisers should be insisting, for reasons of ROI.

Consider what the black box actually costs. It is not primarily a fraud problem, though fraud is comfortable in the dark. The deeper cost is that learning stops compounding. When a campaign works and no one knows why, the knowledge evaporates on contact; nothing transfers, and every quarter starts from zero, which is a strange way to run a learning organization. Meanwhile, the systems grade their own homework, attributing conversions to themselves and optimizing toward whatever their internal math can see, which is rarely the same as incremental sales. Every medium that competes with them for budget, television included, quietly pays for that grade inflation. And a machine optimizing toward a proxy will find every inch of daylight between the proxy and the truth, at a scale no human buyer ever managed.

The remedy is not source code. Nobody needs anyone’s model weights. Buyers need what science requires of any instrument it trusts: independent validation, and explanation in terms of causes.

Step One:Grade the machine on tests it cannot grade itself. Holdout experiments, incrementality studies, marketing-mix models that weigh every channel on the same objective scale, third-party sales-based validation: any report card written outside the platform being graded. If the lift is real, it will survive independent measurement. If it does not, you have just learned the true ROI of the black box, which is also worth knowing.

Step Two:Make “why” a deliverable. The reporting we accept should explain results in human terms: which audiences, which contexts, which creative, ranked by contribution to outcome, and beneath those the causal layer. A rationale composed after the fact does not qualify; a real why survives the same outside tests as the lift. Decades of research, including the work my colleagues and I at RMT have done with Wharton Neuroscience, keep arriving at the same conclusion: advertising works when it resonates with the motivations that drive a person's choices, most of which operate beneath conscious awareness. In Wharton's EEG studies, that motivational resonance was the only statistically significant predictor of the brain synchrony associated with effective advertising. Motivations are causes; clicks, demographics, and attention are correlations wearing the uniform of causes. A system that can report which motivations it activated, in which contexts, produces learning a human being can inherit and reuse. That is what converts spend into knowledge.

Step Three: Keep the hypothesis human. Automation should execute strategy, not absorb it. The machine is superb at the how: clearing prices, allocating impressions, testing variations at inhuman speed. The why, meaning who the customer is, what she is trying to become, which of her motivations the brand genuinely serves is the advertiser's intellectual property, and renting it back from an algorithm is no way to build a brand. Give the machine a causal hypothesis and let it prove or disprove it quickly: that is a partnership. Giving it money and awaiting its verdict is something else.

None of this waits on the platforms’ goodwill. Reach and CPM guarantees went into agency contracts in the 1990s because procurement demanded them; insistence is how standards improve. Notice, though, who feels this problem most: agency and in-house buyers, the people who must operate these systems or be replaced by them, and whose calls for transparency are too easily waved off as job protection. Brand advertisers feel no such heat yet, and that is exactly why their demand would carry weight. Made now, before the machines settle into their budgets, it reads as governance, not grievance, and it writes the terms of trust in advance. Advertisers who move first will enjoy a period in which they understand their advertising while their competitors do not, which has historically been a fruitful place to stand.

In last week's column, I suggested that motivation may be an organizing principle of intelligence itself, artificial as well as human. If that is right, then “why did this work?” is not a reporting feature. It is the entire game. A system that can only say “more” is a vending machine. A system that can say why is a colleague.

Trust the ones that show their work, and pay accordingly.

Posted at MediaVillage through the Thought Leadership self-publishing platform.

Click the social buttons to share this story with colleagues and friends.
The opinions expressed here are the author's views and do not necessarily represent the views of MediaVillage.org/MyersBizNet.