Ads for Agents, a Contradiction in Terms

Written by Brook Schaaf

Just a couple of months ago, one Jonah Goodhart announced Mobian Agent Ads, “what we believe is the world’s first ad designed and targeted specifically to an AI agent.” (This is not to be confused with using agents to buy advertising, a separate hot topic.)

Spoiler: you can’t actually advertise to AI agents, at least not in the traditional sense of paid, persuasive messaging. While it’s common for us to use anthropomorphic language, e.g., ChatGPT “weighs” or “Claude thinks,” this is a convenience. The purpose of advertising is to instill desire. As I quote in my forthcoming book, The Affiliate Hypothesis, the late, great ad man David Ogilvy wrote, “When I write an advertisement, I don’t want you to tell me that you find it ‘creative.’ I want you to find it so interesting that you buy the product.” Emotion leads to action. 

Machines have no desire or emotion; therefore, they cannot perceive or receive an advertisement beyond processing its information, which is little different than any other data input. Aren’t there still contextual clues that can influence data ingestion and agent action? Sure, but that’s still decidedly different from the human experience. 

At the end of the Parable of the Talents in the Gospel According to Matthew, Jesus says, “For to everyone who has, more will be given, and he will have abundance; but from the one who does not have, even what he has will be taken away.” I always thought this statement of Christ was kind of… dare I say… unchristian, like the Gospel of Wealth or social Darwinism. My rector suggested that it describes how the world is, not how it ought to be.

There are other explanations, but whatever its biblical meaning, this sure does describe how things are. In 1968, sociologist Robert K. Merton coined the “Matthew Effect” to describe how eminent scientists accumulated disproportionate recognition. The concept of the compounding advantage of the well-to-do (and corollary disadvantage of the struggling) spread to economics and education, where it also came to be known as the “Matthew Principle.” We see it everywhere, from wealth accumulation to students who are academically ahead in third grade maintaining that advantage throughout their education, to Malcolm Gladwell’s famous example of Canadian junior hockey players who benefit from being older and bigger than their peers at an early age, giving them more opportunities to practice and improve. 

To manufacture a Matthew Effect, today’s winning brands have to keep showing up, while today’s strivers scramble to show them up.

As a case in point, I’m replacing the smoke detectors in my house. ChatGPT recommended Kidde and First Alert, two of the most distributed, informationally abundant brands. I’d had bad experiences with both incumbents, so I looked to Consumer Reports, which steered me toward a brand called Siterwell I’d never heard of before and purchased. (Note: lacking an affiliate program made Siterwell forfeit a signal that could have reinforced future visibility.)

Siterwell would do well with many more such mentions. As conventional wisdom holds, GEO/AEO broadly depends on three things: a preponderance of relevant information, the authority of its sources, and a structure machines can readily interpret. 

Affiliate can provide all three at once: distributed corroboration across relevant shopping surfaces, authority from trusted publishers and creators, and machine-readable product information structured through feeds, links, and infrastructure such as FetchRight. A sponsored block on a general news page is an isolated commercial claim; affiliate can collectively weave a broader, transaction-oriented evidentiary tapestry.

Unsurprisingly, Goodhart got dragged in many of the comments and in a separate Digiday article, buyers expressed skepticism of the process. If this argument is correct, his effort is directionally correct and, like it or not, the term “advertising” will probably be stretched out, because it will appeal to human decision makers. Goodhart may be wrong about what it is, right about what it does and what we’ll call it, then wrong again about how it should be done.