The dawn of digital advertising came with a seductive promise: the right ad served at the right time to the right person. We were told this paradigm would finally solve John Wanamaker’s century-old conundrum, eliminating the legendary 50% of wasted ad spend.
It was a beautiful fantasy. Setting aside the dystopian, “Minority Report”-style surveillance required to pull it off, it promised peak economic efficiency for matching supply with demand.
Fast-forward 20 years, and the illusion of omni-present, resolvable identifiers has shattered. Privacy regulations, signal deprecation and platform walls have proven that a unified, deterministic identity layer across the open web is not going to happen, and it certainly won’t happen across the entire internet.
Yet, rather than mourning the death of this delusion, the ad tech industry has birthed a new one: AI-powered 1:1 ad personalization.
Armed with large language models and infinite compute power, the new engineering dream is hyper-personalization at scale: serving a unique, dynamically generated ad creative to every single consumer. But it is time to reckon with reality. We must reject the fantasy of infinite creative variations – not just because it is technically impossible but because it fundamentally misunderstands how advertising works for the vast majority of brands.
Clocks vs. clouds
As the philosopher Karl Popper noted, the world is divided into two types of systems: clocks and clouds. Clocks are neat, orderly and deterministic. Clouds are dynamic, highly complex and probabilistic.
Ad tech treats marketing like a clock science. It assumes that if you turn gear A (a specific creative asset) in front of person B (a specific data profile), you will yield result C.
But marketing is a cloud science. Advertising is a famously weak force that operates over massive populations and long horizons. While a tiny fraction of ads capture immediate, in-market demand, the vast majority of advertising’s value lies in building and refreshing memory structures across a broad population over time.
When you fragment a brand’s message into 1,000 different AI-generated variations served to 100 million people, you destroy two foundational pillars of effective marketing: shared cultural experience and scientific isolation.
The multivariate nightmare
Consider the measurement math. If you serve thousands of distinct creative permutations without a clean, deterministic signal of effectiveness, you enter a multivariate hell.
If an ad featuring a blue background outperforms one with a green background, why did it happen? Was it the color? Or was it because the blue ad happened to hit users experiencing sunny weather that afternoon? If a creative featuring one demographic group shows a lift, is it true affinity or co-variance with an unmeasured algorithmic bias in the DSP?
An automated system can easily shuffle thousands of variations and double down on whichever ones get a quick click, but it cannot measure what it cannot see. It can optimize for immediate reactions, but it remains completely blind to whether those variations are building long-term brand equity or just chasing statistical noise.
To truly isolate variables, you need controlled experiments. But true experimental design requires limiting variables, not expanding them to infinity. You can run robust tests on three or four distinct creative hypotheses in a clean environment. You cannot scientifically test 5,000.
Hyper-personalization also kills the “fame effect.” Brands are built on a collective understanding of what a brand stands for. When you fragment that message into infinite, hyper-personalized silos, you destroy the macro-cultural signal that gives a brand its authority and prestige. It dilutes a shared asset into statistical noise.
The performance trap
Hyper-personalization can work, but only in a very narrow box: strict, short-term, click-based performance campaigns with finite budgets targeting consumers who are already down-funnel and in-market. For these bottom-funnel prospects, generative AI is a useful tool to squeeze out incremental efficiency by matching immediate intent with hyper-specific creative hooks.
This is the exact terrain where “black box” solutions like Google’s Performance Max or Meta’s Advantage+ thrive. They are highly efficient engines for harvesting existing demand. But the existential danger arises when enterprise brands mistake these bottom-funnel harvesting tools for holistic marketing strategies. If you expand these black-box, hyper-personalized principles to dual-purpose campaigns – those tasked with balancing short-term sales with long-term brand equity – you are doomed to fail.
We must stop confusing what is technically possible with what is commercially realistic. Generative AI is a magnificent tool for creative ideation, production efficiency, localized testing and conversion optimization at the very bottom of the funnel. But if we deploy it to chase the ghost of 1:1 deterministic relevance across the entire media mix, we will simply repeat the mistakes of the last two decades – wasting billions of dollars chasing precision that rarely exists to convince a consumer who isn’t looking.
It’s time to change the North Star of our industry’s playbook. The goal shouldn’t be the right ad for the right person at the right time. It should be the right messages for the right population over time.
“Data-Driven Thinking” is written by members of the media community and contains fresh ideas on the digital revolution in media.
Follow Erez Levin and AdExchanger on LinkedIn.
For more articles featuring Erez Levin, click here.
