This is the law most often used as an argument for not trying. Somebody proposes something unusual, somebody else says Jakob's Law, and the room agrees to do what everyone else does. That is a real conversation I have sat through more than once, and it is a misreading of a claim that is much more specific and much more testable than "be normal".
The actual claim is about what you are carrying when you arrive. Every interface you have ever used has been quietly installing a map — where the close is, where the back is, where the thing that finishes the job is. You did not study it and you cannot recite it, but you use it constantly, and it is the reason a product you have never opened feels immediately operable.
So the test below strips the icons off. Twelve screens, none of them a real product, each with four dashed spots where a control could be. You are told which control to find and you click where you think it is. Nothing is labelled, nothing is revealed, and there is no feedback until the end — if you had four guesses on each screen you would land about three right out of twelve. Four of the twelve put the control somewhere convention does not. Those four are the interesting ones, because a wrong answer there is not a wrong answer. It is a reading of where your map said to look.
This one only works once. By the end you will know the answers, so a second run measures your memory of this page rather than anything you brought to it.
What the four broken screens are for
Scoring the eight ordinary screens only tells you that conventions exist, which nobody doubted. The four that move the control are where the law actually gets tested, because they separate two things that look identical from the outside: not knowing where something is, and confidently knowing, and being wrong.
If you were searching, your misses on those four would scatter across the three empty spots. If you were carrying a map, they land on one specific spot — the one every other product trained you to use. That is a much stronger result than a high score, and it is the one worth taking back to work, because it makes the cost of breaking a convention predictable. You are not buying vague confusion. You are buying a particular wrong click, in a place you can name in advance.
What it costs you
Convention is not a style you can opt out of. It was installed in your visitor by a thousand products before yours, most of which you have never heard of and none of which you control. You inherited it whether you like it or not; the only decision available is whether to spend against it.
"Be like everyone else" is the wrong lesson. The map is about where things are, not what the product does. Matching it on structural controls costs you nothing distinctive — no one has ever chosen a product because its close button was somewhere surprising — and it buys you every bit of attention you were about to waste. Spend the novelty budget on the part that is actually yours.
Novelty in structural controls is the most expensive novelty there is, because the person paying is not thinking about your product while they pay. They are hunting. Whatever you wanted them to notice next is gone.
This is why the cleaner redesign gets worse numbers. Tidying an interface almost always means moving things, and every move invalidates a map that was working. The old layout was not good; it was learned, which for the first few weeks is worth more.
The map is not universal, it is a habitat. It is built from the products a particular person actually uses, which is why conventions that feel absolute in one market are merely common in another. Everything on this page assumes a left-to-right reading order, and that assumption is doing more work than it looks like.
The version of this I keep having to argue against is the crowded header. Something has to come out, search is "only used by a few percent", so it goes behind a menu — and then the few percent who used it were the people who knew what they wanted, and they stop finding it. The usage number that justified the move was measuring the convention, not the feature.