The Mirror Problem

1.The average customer does not exist

In 1952 the U.S. Air Force had a crash problem it couldn't explain. Lieutenant Gilbert Daniels measured 4,063 pilots on ten body dimensions and asked a simple question: how many of them fall within the middle 30 percent on all ten? The cockpit had been designed around exactly that average pilot. The answer was zero.[1]

The arithmetic explains why. If a fraction $q$ of people sit near the middle on any one trait, and traits are roughly independent, the share who sit near the middle on all $m$ of them is

A segment ad is a cockpit. It is written for the centroid of a crowd: women, 25 to 34, urban, interested in travel. Each added trait narrows the crowd it truly fits, and the decay is exponential, not linear.

The industry's answer has been to make the crowd smaller. Narrower segments, lookalikes, micro audiences. That helps the targeting. It does nothing for the ad, which is still one image shown to everyone inside the slice. Shrinking the crowd is a way of admitting the problem without solving it.

2.The self is the strongest signal

Psychology found the most powerful lever in persuasion almost fifty years ago and advertising mostly ignored it. In 1977, Rogers, Kuiper and Kirker showed that people remember words far better when asked whether a word describes them than when asked what it means.[2] A meta-analysis of more than a hundred studies later confirmed the self-reference effect as one of the most reliable findings in memory research.[3] The self is not one topic among many. It is the most elaborated structure in the mind, and anything attached to it is processed more deeply.

Ahn and Bailenson took that literally. They showed participants a brand being used by a virtual figure that carried the participant's own face, or someone else's. Seeing themselves with the product produced more favorable brand attitudes and stronger purchase intent.[4] Elder and Krishna found a quieter version of the same mechanism: simply orienting a product so a viewer can imagine their own hand reaching for it raises purchase intention, because the brain starts rehearsing the action.[5]

Put those together and you get a claim that sounds obvious once stated: the most persuasive image of a product is one that already has you in it. Not a model who resembles your segment. You, in the moment the product was made for.

3.Traits travel

The second lever is how someone thinks, not just how they look. In 2013, Kosinski and colleagues showed that Facebook likes alone could predict private traits for 58,000 volunteers: political party correctly in 85 percent of cases, and personality with useful accuracy.[6] Four years later, Matz and colleagues ran field experiments reaching 3.5 million people. Ads written to match a recipient's inferred personality, extroverted copy for extroverts and quieter copy for introverts, produced up to 40 percent more clicks and up to 50 percent more purchases than mismatched ones.[7]

The constraint then was production. Someone had to write each variant, so studies used two or three. In 2024 the same group showed that a language model could write psychologically matched messages on its own, and those messages were consistently more persuasive than generic ones across domains.[8] The step from a few hand-written variants to one per person is no longer a research question. It is an engineering one.

4.Specific is not the same as personal

If personalization works so well, why does the sofa follow you around? Because the industry built specificity and called it personalization. Lambrecht and Tucker tested dynamic retargeting, the ads that show the exact product you viewed, against generic brand ads. On average the specific ads performed worse. They only won once a shopper's preferences had narrowed, for example after visiting review sites.[9] Bleier and Eisenbeiss found that deeply personalized ads are strongest right after a visit and wear out faster than generic ones as time passes.[10]

The lesson is not that personalization fails. It is that repeating someone's browsing history back to them is the shallowest possible version of it. The product is the least interesting thing you know about a person. Who they are, where they'd wear it, and what it would make them feel are what move them. Retargeting describes the cart. A reflection would describe the person.

5.The line between relevant and invasive

Every conversation about personalization arrives at the same word: creepy. The research is more precise than the word. Goldfarb and Tucker studied thousands of display campaigns and found that ads matched to their context and ads that were visually obtrusive each raised purchase intent on their own. Combined, they lowered it, and the drop was largest among people who cared most about privacy.[11] Aguirre and colleagues found that collecting information overtly lowered click-through, while credible, trust-building settings recovered it.[12]

The most important result is Tucker's. When Facebook gave users more control over their privacy settings, those users became nearly twice as likely to click on personalized ads.[13] Control didn't limit personalization. It unlocked it. We model this as a response curve where relevance rises with personalization depth $d$ until it tips into intrusion, at a threshold $\kappa$ set by how much control the person feels they have:

That is the controversial claim at the center of this piece. Creepy is not a property of data. It is a property of who is in control of it. The same depth of personalization reads as service when someone chose it and as surveillance when they didn't. Consumers are not asking for less of it either: 71 percent say they expect personalization, and 76 percent are frustrated when they don't get it.[14]

6.Why now

Two curves crossed in the last few years, and either one alone would have been enough to change the industry.

The first is signal loss. For two decades, ads found people by following them: third-party cookies, device identifiers, cross-app tracking. One platform decision at a time, that machinery has been dismantled.

Safari began limiting cross-site tracking in 2017.[15] GDPR made consent a legal requirement in 2018.[16] Firefox blocked third-party trackers by default in 2019.[17] In 2021, iOS 14.5 made every app ask before tracking a person across other apps and websites.[18] The industry read this as the end of personalization. We read it as the end of personalization without permission. The passive signal is disappearing. The signal a person chooses to share is the only one left standing, and it's the richer one.

The second curve is the cost of an image. In 2023, DreamBooth showed that three to five photos of a subject are enough to render it faithfully in new scenes.[19] IP-Adapter showed that a small adapter can carry a reference image's identity into generation without retraining the model.[20] For the first time, rendering a specific person into a specific moment stopped being a photo shoot and became a function call. The economics follow directly:

Those two curves together are the opportunity. Following people is getting harder every year. Making something for a person who invited you is getting cheaper every month.

7.Lift, or it didn't happen

Any claim about advertising has to survive the hardest result in the field. Lewis and Rao analyzed 25 large field experiments and showed that the effect of a single ad is so small relative to the natural noise in purchasing that measuring it reliably takes enormous samples.[21] Gordon and colleagues compared observational attribution methods against randomized experiments at Facebook and found the observational methods often failed to recover the true effect, frequently overstating it.[22] Most of what the industry reports as performance is correlation with a dashboard.

The honest measure is incremental lift against a randomized holdout, and the sample it needs depends sharply on how large the lift is. For a baseline conversion rate $p_0$, a relative lift $r$, 5 percent significance and 80 percent power:

This cuts both ways, and we think it's the most underrated argument for going one to one. A marginally better banner needs a measurement budget larger than its media budget to prove anything. A change in kind, an ad that actually describes the person seeing it, should produce a lift large enough to show up in a holdout of a few thousand people. Ghost-ad methods have made those holdouts far cheaper to run without wasting impressions.[23] If the effect is real, it will be obvious. If it isn't obvious, it isn't real.

What to actually measure

Most advertising dashboards show impressions, click-through rate, and attributed revenue. By the argument above, all three can look healthy while the ad persuades no one. These are the numbers we'd put on the first screen:

Where this leaves us

Advertising inherited its model from print and television, where one image for everyone was the only option. The internet kept that model and bolted targeting onto the front of it. We got extraordinary at deciding who sees the ad and stayed amateurs at deciding what the ad is.

Our view is that the next era of advertising is an identity problem, not a targeting problem. It is about who a person is, what they have chosen to share, and what the moment would look like if it were made for them alone. The science has said this for fifty years. The signal environment now demands it. The cost curve finally allows it.

We are still working through what this means once every ad can be different. If you run advertising at scale and your numbers disagree with our model, we want to see them. That's the most useful email we can get: founders@clockingminds.com.