Marketing that marks its own homework

Posted on: 14 October 2026

In 2017 Uber was spending around 150 million dollars a year on digital advertising to get people to download its app. Kevin Frisch, who ran performance marketing at the time, switched off 100 million of it and the number of installs stayed exactly where it was. Installs that the reports had credited to paid channels the day before simply started showing up as organic, which is to say people were downloading the app anyway while someone along the way collected the credit for making them do it.

That same year JPMorgan Chase discovered its ads were appearing on roughly 400,000 websites a month, of which only 12,000 produced anything beyond a bare impression. The bank cut the list to about 5,000 sites approved one by one, and its chief marketing officer reported that costs and visibility had barely moved. Also in 2017 Procter & Gamble took 200 million dollars out of its digital spend, moved it into other media and measured a ten per cent increase in overall reach.

The oldest case is also the cleanest. Economists at eBay suspended paid search ads on queries containing the company's own name and found that almost all of that traffic simply moved to the free results sitting just below. eBay had been paying Google to intercept people who were already looking for eBay. When the study became public the company let it be known that it had no plans to change how it spent on search advertising. That is the detail worth dwelling on.

Put these stories side by side and another one emerges that usually goes unnoticed, namely that in none of them did the discovery come from the marketing function. At Uber it all began because the founder, Travis Kalanick, was furious to see the company's ads on Breitbart, so Frisch started turning off channels for a reason that had nothing to do with effectiveness. Chase acted after the New York Times showed it one of its own ads sitting beside a conspiracy headline. Procter & Gamble moved under pressure from bot traffic and from brands ending up next to unpresentable content, while at eBay the experiment was run by economists who had nothing to gain from the advertising budget. Each time the truth arrived from outside and almost by accident, whereas the function that held every piece of data needed to see it had not seen it.

It is tempting to read incompetence or bad faith into this, but the explanation is more mundane and for that reason more serious. In almost every company the function that spends the money is the same one that certifies what the money achieved. Attribution models give every sale a parent and that parent tends to be the last paid channel before purchase, the one that costs money and justifies next year's budget. Many agencies earn more when the client spends more, and the platforms selling the space also supply the tools used to measure how well that space performed. Across the whole chain there is not a single person paid to discover that part of their own work moves nothing. Frisch himself, recounting the episode, spread the blame across everyone, including the advertiser who kept asking for more volume.

The structure has a precedent far older than digital. For more than two thousand years medicine practised bloodletting against almost every illness and did so with a clear conscience, because the physician who bled the patient was also the one who judged the outcome. When the patient recovered the credit went to the treatment, and when the patient died the conclusion was that the bleeding had come too late, which only encouraged doctors to bleed earlier and harder. In 1835 a Paris physician, Pierre Louis, did something both obvious and revolutionary, because he started counting. He compared pneumonia patients bled early with those bled later and concluded that the practice brought no benefit. The medical establishment attacked him fiercely and it took decades before bloodletting was abandoned.

Digital marketing is living through the same phase today. It has measurement tools vastly more powerful than anything a nineteenth-century doctor could imagine, yet it uses them almost entirely to count who was present at the moment of purchase rather than to ask what would have happened in its absence. The two questions sound alike, though only the second tells you whether a piece of work matters. The first merely certifies that you were there.

Until recently this ambiguity was sustainable, because the visible work of marketing had a cost and that cost served as implicit proof of value. Writing a campaign and producing its materials took people and time, so the budget seemed justified by the plain fact of having to pay whoever did it. AI systems are pushing that cost towards zero, to the point where the ad platforms themselves now offer to generate copy and images and to allocate spend across channels on their own. Once production costs almost nothing, the only thing left to charge for is proof of having caused a result. A profession that has spent twenty years getting better at claiming results than at producing them is going to find itself exposed precisely there.

For anyone working in marketing, then, the useful question is not whether their work performs, since the dashboards always answer yes to that one. It is a less comfortable question: how much less would your company sell if you were switched off tomorrow? Not how much the platform attributes to you, but how much would actually disappear. Not knowing the answer does not make you a bad professional, but it does mean you practise a discipline that has never demanded of itself the proof that nineteenth-century medicine had forced on it from outside. There is a second question about the role itself, though few people are keen to ask it: if the part of your job you can demonstrate is the part a machine will do for free, what is left of what you do?

None of this means marketing does not work. The same eBay study found that ads did influence people who barely knew the site or used it rarely, while they were useless on those who already shopped there. Marketing works, but in specific places and on specific people, so finding out which ones requires exactly the discipline the industry has avoided.

For anyone defending or signing off a budget, the most common mistake is treating attribution as a measure of effect, because it records who was present at the moment of purchase and not who caused it. From that follows the habit of rewarding growth in attributed conversions without ever having seen what happens when spend is switched off in one region or for a randomly chosen group of customers. Tests of this kind are cheap, since you give up a fraction of the budget for a few weeks, yet they answer a question no dashboard can touch.

The worst mistake, though, is waiting for the finance director or a journalist to suggest switching a channel off. In the 2017 cases the proof came from outside and those on the receiving end lost control of their own story, so it makes sense for the marketer to be the first to propose switching off part of their own budget with a proper control group, while it is still them deciding how the result will be read.