The demonstration
Uniqueness
is manufactured.
The 600 people in the table below behave absolutely identically. They pick at random, from the same basket, with no habits and no preferences. Nothing personal, by construction.
The only thing you change is how many things you record about each of them.
Every square is a person. The red disappears as you add columns — not because the people changed, but because there are too many combinations for any two to land on the same one.
The test
Three companies report the same percentage.
Each measured 2,000 customers and each found that nearly 95% are unique. But one has customers who behave absolutely identically, and another has customers who are deeply different from one another.
Guess which is which.
It was not guessable. The uniqueness percentage does not contain the information.
This is the paper's main result, in exact form: if all I tell you is what percentage of the base is unique, any amount of real individuality remains possible — from zero to any level. It is not a weak signal. It is not a signal.
The calculation
You tested personalisation and it did not work?
It usually does not mean personalisation fails. It means you did not have enough observations per person for the test to even have a chance of coming out positive.
Below the threshold shown here, a negative result proves nothing.
observations per customer, at minimum, before a simple comparison between “per-person mean” and “grand mean” can come out in favour of personalisation.
The figure is a necessary threshold, not a guarantee. Above it, personalisation starts to be worth it. Below it, you have probably cut a good programme on the strength of evidence that never existed. With a model that shrinks personal estimates toward the grand mean, the threshold disappears — but almost nobody does that on a first test.