Two versions of the same page go live at the same time. One with the explanation screen, one without. A few thousand people see one, a few thousand the other, and after two weeks there’s a number: version B gets more completions.
That number costs little more than waiting time. It’s reproducible. And it ends every discussion in the room.
What it doesn’t measure is the person who needed the explanation. They didn’t drop off in version B. They got through, faster even, and agreed to a figure whose origin they don’t know. In the analysis they show up as a success. There is no column in which their misunderstanding would appear.
Put the explanation screen back, and it won’t last: it was there, in version A. And version A lost.
I have been designing financial products for more than ten years. I know which option has to be preselected for it to be chosen. And I know how an explanation has to sound for nobody to ask any more questions.
Most of these effects arise without a plan, as a by-product of the same method you use to make things easier. You remove one hurdle. Then a second: one click fewer, one sentence shorter, a restriction that only appears on request. Each of these decisions is reasonable on its own, and over a thousand decisions you end up with a surface on which nothing sticks any more.
Mostly, all that disappears is the time in which you could have checked the information.
Most people in my industry are smart, well trained and want to build products they can be proud of. Still, tests like this one push a text in one direction. Better there means fewer drop-offs, fewer support queries, better scores in the satisfaction survey. Whether anyone understood appears in none of these numbers.
For years I used the word friction wrongly myself.
In UX design, friction is almost always a problem. “Frictionless” is the highest praise a digital product can receive, and in business terms that is compelling: friction costs users, every moment of hesitation is one in which someone drops off, and this relationship can be measured cleanly.
The obvious rejoinder is that you just have to build resistance back in. Make things a bit more laborious. Insert a step that forces you to pause.
That rejoinder falls short. For perceptual hurdles the research is clearer than I would have liked: hard-to-read type doesn’t improve learning. It creates effort, not understanding.
A hint as to what works instead comes from fluency research itself. McGlone and Tofighbakhsh had shown that rhyming statements seem more credible than non-rhyming ones with the same content. A year later they told some of their participants in advance to separate form from substance. For these people the advantage of the rhymes was smaller. What worked there was no resistance in the text. It was a sentence beforehand.
Translated into products, this would be less catchy than “more friction”: a built-in checkpoint in the process and in the organisation behind it. A point at which questions are required rather than punished. At which someone has to ask whether the other person knows what is happening – and whose result is measured instead of being left to chance.
That is a process and organisational problem. Which is why no redesign solves it.
How expensive it gets when nobody in an entire organisation asks this question is shown by Zillow. In 2021 the company turned its algorithm into a price tag for houses and closed the division eight months later. My book, written in German, opens its chapter “When No One Asks” with this case.
For people with dyslexia, with cognitive impairments or who live their daily lives in a second language, smoothness isn’t a trick but access. Awkwardness quickly becomes a bouncer for them. Anyone who demands “more friction” as a general rule builds barriers for exactly the people with the least reserves.
A surface may be easy, and mostly it should be. What needs settling is at which point in the process it had better not be.
A comparison helps. Count how much resistance you meet on an ordinary morning. To get to an article, you click through consent banners, a log-in, a paywall and the newsletter pop-up after the third paragraph. A new account requires ID, video identification, a confirmation code and a wait; a larger transfer, a query as to whether the details are correct. The whole apparatus holds you up where money flows, data is collected or an identity has to be verified.
Then you read an explanation that changes your picture of something, and nobody follows up. No confirmation step, not a second’s wait. You read it, it makes sense, it’s in.
The transition from what you’ve read to what you believe is the least controlled point in your digital day – and the one where only you yourself could check.
If a surface should be anything but smooth anywhere, it is here.
With language models, something is added that makes this point smoother still. There, smoothness is trained in deliberately. What that means for every answer you get from a machine I have set out in my book under the term fluency machine.
Then there are the little explanation fields next to recommendations, meant to justify why a system recommends what it recommends. They are friendly and plausible, and they serve a function rarely spoken aloud: they end questions. After a question has been answered, you understand more. A question that has been ended, you no longer ask.
I keep building flows like these. The question of how to make them simpler remains. A second one has joined it: at which point in the process should a flow better not be easy – and who checks there whether anyone knows what is happening?
People who build systems shouldn’t leave this check to users’ self-discipline. That isn’t an extra ethical demand. It is the job.
The full essay on this topic:
