System Usability Scale

Usability findings used to exist only as sentences. Three of five participants couldn't find the save button. Someone tried to skip past the invite screen without choosing a role. Those observations make an excellent list of things to fix, and they can't be compared to anything.

Pagedeck hit the same wall. When Grace asked whether the product had gotten easier to use than last quarter, Noa had no number to give. The two studies had used different tasks and recruited different participants, so laying the observations side by side still couldn't say whether anything improved. Features were being added and screens were changing, and nobody could answer whether the direction was up or down.

That's the gap a single-score survey fills. What made this one stick isn't precision; it's that it never changed. The same ten questions in the same order have been in use for decades, which means your score can be set against your own last quarter and against values other organizations have published. Rewrite the questions to suit your product and that comparability disappears on the spot, which is the one modification that costs more than it saves.

Collect an answer from 1 to 5 on each of ten questions, convert each answer to a 0-to-4 value, add them up, and multiply by 2.5. Four steps.

1. Collect the answers. The questions alternate between positive and negative. Odd-numbered ones read like "I would use this system frequently," even-numbered ones like "I found the system unnecessarily complex." Answers run from 1 for strongly disagree to 5 for strongly agree.
2. Align the direction. For odd questions subtract 1 from the answer; for even questions subtract the answer from 5. Both land between 0 and 4, with higher meaning better. The negative questions are interleaved to stop respondents from ticking a single column without reading.
3. Add the ten values. The total falls between 0 and 40.
4. Multiply by 2.5. That gives a number between 0 and 100, which is one person's score. The product's score is the average across participants.

Take someone who answers 4, 4, 5, 4, 4 on the odd questions and 2, 2, 1, 2, 2 on the even ones. The odd values become 3, 3, 4, 3, 3 and the even values also become 3, 3, 4, 3, 3, for a total of 32. Multiplied by 2.5, that's 80.

You collect it right after the participant finishes a task. A common arrangement is to run a moderated test, where a facilitator sits alongside and watches, and attach the survey at the end. Because this belongs to quantitative research, though, it needs a sample. Five people is enough for observation; a score needs something like twenty before it stops moving on its own.

68 is the reference line. It's known as the average across a large pool of products, which makes it the only footing you have for saying where your own number sits relative to the outside world.

The rough bands go like this. Below 50 means people are running into walls continuously, and the answer is usually structural rather than cosmetic. Around 68 is where an unremarkable product sits. Above 80 counts as strong, and scores in the 90s are uncommon because they require most respondents to have given near-maximum answers on all ten questions.

The change matters more than the absolute value. Two points in time are only comparable when the task was the same and the participants met the same conditions. Put 71 and 78 next to each other after changing the conditions and you can't tell whether the product improved or the sample did. Small gaps deserve the same caution: with twenty participants, a difference of three or four points is inside the range you'd get from measuring the same product twice.

What the score doesn't tell you is just as definite. It doesn't locate the problem. A 62 says something is getting in the way, and what to fix comes from recordings and observation. There's also a practice of splitting the ten questions into learnability and usability subscores, but below sample sizes in the hundreds those two values move around close to randomly.

Keep it separate from net promoter score (NPS). Asking whether someone would recommend the product measures overall satisfaction; this scale asks only whether the thing is easy to operate. Products that are cheap and pleasant to use with low satisfaction exist, and so does the reverse.

"Last quarter was 71 and this one is 78. Same task, 22 participants."

The conditions were held fixed, so the two values can be compared.

"62 is below average. This number won't tell us where the problem is, so let's start with the recordings."

Treating the score as a signal and finding the cause by other means.

"Let's keep a score from eight people out of the report. That number can't support any claim."

Warning that the sample is too small for the value to mean anything.

Noa measured the invite flow the same way before and after reworking it. Both times the task was inviting three colleagues at three different permission levels, and the participants were 22 admins on paid plans at roughly 20-person teams. The score before was 64 and after was 76. What went to Grace was those two numbers plus one line on what had changed. A score on its own becomes a target to raise next quarter; a score with the change beside it leaves a record of what actually worked.

The most common mistake reads it as a percentage out of 100. A value between 0 and 100 invites that. But the number comes from converting answers onto a 0-to-40 range and multiplying the scale up, so it isn't a percentage at all. An 80 doesn't mean 80 percent of users were satisfied, and it doesn't mean eight of ten questions got good answers. Read as a percentage, 68 starts looking like a failing grade, and an ordinary product gets reported as a crisis.

The other mistake expects a low score to say what needs fixing. A number arrived, so the cause feels like it should be in there. But this survey captures only the impression a user ended up with, not which screen produced it. Walk into a meeting holding just the score and everyone names the screen they already suspected, and the thing that gets chosen may not be where anyone was actually stuck.