nth Labs
Start testing
Product Pricing Security Blog Help About
Sign in Start testing
Blog

You are not your user, in five numbers

You are an unusual person. You write code, you are at home in software you have never seen before, you read English natively, you are optimistic about AI, and you believe the screen when it says something worked. Almost nobody you are building for is all five. Here is the count.

nth Labs 6 min read

There is a research dataset called MatrAIx Persona: ten thousand simulated people, built by an academic team so that products can be evaluated against somebody other than the person who wrote them. It is MIT-licensed and it is sitting in this repository, because our own tester panel draws from it.

It is worth reading on its own, before any product touches it, because it quantifies something most people building software know abstractly and design as though they do not believe.

How unusual you are

Take the five things most likely to be true of you and false of the person using what you built. For each one, the share of the population that matches:

How much of the population is like the person building the product Each bar is the share matching that description. The faint bar behind it is everybody. MatrAIx Persona (MIT), 10,000 simulated records for product evaluation each row divides by the records carrying that field, printed under the bar 0%20%40%60%80%100% writes code professionally27%1,469 of 5,385 records carrying this fieldat home with new software56%2,935 of 5,228 records carrying this fieldreads English natively46%1,734 of 3,734 records carrying this fieldpositive about AI50%2,619 of 5,272 records carrying this fieldtakes what it says on trust21%951 of 4,568 records carrying this field
Counted by scripts/userfigs.cjs. Each row divides by the records carrying that attribute, not by all ten thousand, and prints the denominator under the bar. About half the records carry any given field; counting the absent ones as “no” would have inflated every number on this chart.

The one at the bottom is the one worth sitting with.

Only one in five arrives trusting you

The dataset models a disposition it calls trust level. Just over a fifth are trusting. Nearly half are verifying, which means they will believe you once they have checked. Another third are skeptical before they start.

Think about what that does to an interface built by somebody in the first group. A green “Saved!” toast is entirely convincing if you already believe the software. If you are one of the four in five, you go back and look. If the thing did not actually save, you now know two things: the data is gone, and the product lies to you. Only the second one loses you the customer.

The same logic runs through everything: a progress bar with no numbers on it, a confirmation email that does not arrive, a price that appears at the last step. Each is fine if you are trusted and expensive if you are being checked.

Digital natives are rarer than people who avoid software

We expected this one to lean the other way.

Comfort with new software Reluctant and avoidant outnumber digital natives 1.8 to one. recorded on 5,228 of 10,000 records ยท the rest do not carry this field and are not counted either way 0%10%20%30%40% Comfortable44.0%2,298Cautious adopter21.8%1,140Reluctant15.9%829Digital native12.2%637Avoidant6.2%323Aware0.0%1
Comfort with new software across the records that carry the attribute. The “Aware” row is a category of one in the upstream data; it is drawn rather than dropped, because tidying it away would report a neater set of options than the data has.

Twelve percent are digital natives. Twenty-two percent are reluctant or avoidant. There are nearly twice as many people who would rather not deal with your software as there are people who will pick it up instantly.

The middle is where most of the population sits: comfortable and cautious adopter together are two thirds. Not confused, not delighted. Willing, if the thing makes sense.

Four habits the numbers argue for

None of this is an argument for dumbing anything down. It is an argument for four specific habits, each of which follows from a number above:

  • Say what happened, not that it went well. “Saved” is a claim. “Saved at 14:32, showing the new version” is checkable, which is what four in five people are going to do anyway.
  • Never make the confident path the only path. A flow with no way back works if you already trust it. Undo, a visible history, an obvious way to reach a human: these are for the verifying majority, not for edge cases.
  • Write for someone who did not grow up doing this. Not simpler words, fewer assumed steps. The reluctant fifth are not stupid, they are unwilling to guess.
  • Remember that half your readers are not native English speakers. 46% read English natively in this data. Idiom, jokes and clever button labels are the first things to fail.

The limits of this data

It would be easy to write the sentence “80% of people do not trust software” and it would be false, so here is the boundary.

  • These are simulated people, not surveyed ones. The dataset is a research release built to give product evaluation somebody to evaluate against. It describes who that community models your users as, which is a smaller claim than a fact about the world.
  • Roughly half of the records carry any given attribute. Every percentage here divides by the records that have the field, and every chart prints that number.
  • Some attributes were deliberately removed. Ethnicity, religion, marital status and gender are dropped before the data enters this repository. They are not in the figures because they are not in the file.

The dataset’s own authors are clear that simulated users are for exploration and stress-testing rather than as a substitute for real ones, and that caveat is theirs rather than ours. What the numbers are good for is the thing above: a reminder, with a denominator attached, that the person on the other side of the screen is not you.