Back to blog
View over a man's shoulder of a laptop on a wooden desk showing a calendar app, with his right hand on a mouse, his left hand holding a phone, a separate white keyboard and a glass of water beside it
ProductNovember 17, 202610 min read

The System Usability Scale: what ten questions can and cannot tell you

The System Usability Scale turns ten agree-disagree statements into one usability score. Here is how to ask it, score it and read it, and why a 68 is average rather than a failing grade.

By SurveyLane · The team building SurveyLane

One person on your team says the new onboarding feels easier. Another says it doesn't. The System Usability Scale was made to settle that kind of argument with a number: ten short statements, a minute of a respondent's time, and decades of published scores to compare against. It also gets scored wrong, read wrong and asked at the wrong moment, often in the same project.

Where the System Usability Scale comes from

John Brooke built SUS in 1986 at Digital Equipment Corporation and called it a "quick and dirty" usability scale. Labs back then used long, expensive questionnaires, and Brooke wanted something an engineer could hand out after a test session and score in minutes. He published it in 1996 as a chapter in the book Usability Evaluation in Industry. It's free to use as long as you credit the source.

SUS lasted because thousands of teams kept using the same ten statements. That built up a large pile of reference data, which is the main reason to pick SUS over questions you write yourself.

The ten statements, word for word

Each statement gets a five-point scale from "Strongly disagree" to "Strongly agree":

  1. I think that I would like to use this system frequently.
  2. I found the system unnecessarily complex.
  3. I thought the system was easy to use.
  4. I think that I would need the support of a technical person to be able to use this system.
  5. I found the various functions in this system were well integrated.
  6. I thought there was too much inconsistency in this system.
  7. I would imagine that most people would learn to use this system very quickly.
  8. I found the system very cumbersome to use.
  9. I felt very confident using the system.
  10. I needed to learn a lot of things before I could get going with this system.

The odd statements are positive, the even ones negative. Brooke alternated them on purpose: someone who ticks down one column without reading ends up contradicting themselves.

Scoring: why you cannot just average the answers

A SUS score is never the mean of the ten answers. Half the statements point the other way, so a plain average mixes "I agree this was easy" with "I agree this was complex".

For one respondent, code each answer from 1 (strongly disagree) to 5 (strongly agree). On the odd statements, subtract 1 from the answer. On the even ones, subtract the answer from 5. Every item now contributes 0 to 4 points, higher is better. Add the ten up and multiply by 2.5. You get a score from 0 to 100.

A respondent who answers 4 on every odd statement and 2 on every even one gets 3 points per item, 30 in total, so 75. Someone who ticks "Strongly agree" on all ten without reading lands on exactly 50.

Calculate the score per respondent first, then average those scores. And don't report the ten items as separate usability measures. Brooke said plainly that the individual items mean little on their own.

A score of 68 is average, not a weak pass

The scale runs to 100, so people read it as a percentage and hear a 68 as a C-minus. Jeff Sauro of MeasuringU pooled about 500 evaluations with more than 5,000 users and found an average SUS score of 68. That's the middle.

A percentile is the better reading. In Sauro's data, a 74 beats roughly 70 percent of the products in the set. His grade curve puts an A above 80.3, the top 10 percent. A 68 is a C. Below 51 is an F, the bottom 15 percent.

So never show the bare number. Say "74, above average for software in general" or "52, in the bottom fifth", and name what you compared against.

Putting words to the numbers with the adjective scale

Aaron Bangor, Philip Kortum and James Miller tried another angle. In the Journal of Usability Studies of May 2009, they reported adding an eleventh question to nearly 1,000 SUS questionnaires: describe the product's user-friendliness in one word, from "Worst imaginable" to "Best imaginable". The words tracked the scores closely, at r = 0.822.

The mean SUS score per word: worst imaginable 12.5, awful 20.3, poor 35.7, OK 50.9, good 71.4, excellent 85.5, best imaginable 90.9.

Look at "OK". A product people call OK scores around 51, close to an F on Sauro's curve. If a stakeholder hears "users found it OK" and relaxes, show them where OK sits.

The alternating wording backfires more often than it helps

The positive-negative pattern was meant to catch careless respondents. It catches careless researchers too. In a CHI 2011 paper, "When designing usability questionnaires, does it hurt to be positive?", Jeff Sauro and James Lewis found that 11 percent of the SUS data sets they reviewed had been miscoded by the researchers, usually by forgetting to reverse the even items. And 13 percent of individual questionnaires had respondent mistakes, such as agreeing with a negative statement while clearly meaning the opposite.

They also compared standard SUS with a version where all ten statements are positive ("I found the system to be simple" rather than "unnecessarily complex"). The scores showed no meaningful difference, so the all-positive version removes a known source of error and measures the same thing.

You have two options. Use the original, build the reverse coding into your analysis once and test it on a few hand-scored responses. Or use the all-positive version and score every item as the answer minus 1. Just don't mix the two in one trend line.

Small wording changes that are allowed, and the ones that are not

Two edits are widely accepted. Swapping "system" for "website", "app" or your product name helps people who don't think of what they used as a system. Swapping "cumbersome" for "awkward" helps too: Kraig Finstad showed in 2006 that non-native English speakers often didn't know what "cumbersome" meant, and Bangor and colleagues used "awkward" in their large study.

Other changes break comparability. Seven answer points instead of five, a dropped statement, an added "not applicable" option or statements rewritten in your house style all mean the 68 benchmark stops applying. Your scale may still be useful, but don't call it SUS. Our guide to designing rating scales explains why the scale shapes the answer, and the piece on translating surveys covers the same risk across languages: use a published translation where one exists.

When to ask it, and who should answer

SUS rates an experience, so the respondent needs to have had one. In a moderated test, ask straight after the tasks and before you discuss them. Talking about what went wrong pulls scores down. A friendly facilitator pulls them up. Brooke's instructions ask for an immediate reaction to each statement, without dwelling on it.

In an in-product survey, ask someone on their first day and you mostly measure your onboarding. Ask someone who opened the app once, three months ago, and you measure their memory. Target people who have used the product enough to have a view, and use conditional logic to skip the block for anyone who hasn't done the thing you're asking about.

What SUS will not tell you

SUS tells you how usable people found something. It doesn't tell you why. A 55 can come from confusing navigation or from one broken step. The ten statements can't tell them apart.

So don't run it alone. In a test, pair it with task data: who finished, and where they got stuck. In a survey, add one open question after the scale, such as "What was the most frustrating part of using it?", and read the answers. Our guide to analysing open-ended responses shows how to code them, and where AI helps and where it starts inventing themes.

How many respondents you need

Sauro notes that SUS gives reliable results with very small samples, even two users, in the sense that the scale behaves consistently. A small sample still can't give you a precise estimate. Ten respondents produce a mean with a wide confidence interval.

Report that interval. If version A scores 70 with an interval from 62 to 78, and version B scores 74 with an interval from 66 to 82, you haven't shown that B is better. Our article on how many respondents you need explains margin of error and why subgroup comparisons need far more people.

Comparing scores over time and between products

A SUS score earns its keep when it moves, quarter on quarter or before and after a redesign. That only works if wording and timing stay the same and you ask the same kind of user. Ask power users one quarter and trial users the next, and the drop says more about your sample than your product. Our piece on repeating a survey without breaking the trend goes through the traps.

Across very different products, be careful. The 68 average pools many kinds of software, and a complex internal tool does a different job than a simple consumer app. Beating your own last score tells you more than beating 68.

Shorter alternatives when ten questions are too many

Ten statements is short for a usability test and long for an in-app pop-up. UMUX-Lite, published by James Lewis, Brian Utesch and Deborah Maher in 2013, uses two seven-point statements: whether the product's capabilities meet your requirements, and whether it's easy to use. For single tasks, many teams ask one seven-point question about how easy the task was. Neither has the benchmark depth of SUS. If you switch, start a new trend line instead of converting old scores.

Running SUS in a survey tool

Use one five-point scale with labelled endpoints for all ten statements, in the original order. On a phone, ten rows by five columns is hard to use, and our guide to designing surveys for smartphones explains why separate questions on small screens give cleaner data.

Then take the arithmetic out of people's hands. Build the scoring into your analysis once and check it against a few hand-scored responses, including one with identical answers throughout, which should come out at 50. If you analyse results with an AI assistant, for instance through SurveyLane's MCP server, give it the scoring rule explicitly and check its first result. Ask a model for "the SUS score" without the rule and it may well average the raw answers.

Frequently asked questions

Is a SUS score a percentage?

No. The scale runs from 0 to 100 only because the raw total is multiplied by 2.5. A 68 is the average across hundreds of studies, so it's a middling result. Read the score as a percentile, or against your own earlier measurements.

Can I leave out one of the ten statements?

You can, but then published benchmarks such as the 68 average no longer apply, and it isn't SUS anymore. If ten statements is too long for your context, a measure built to be short, such as UMUX-Lite, is a cleaner choice than a trimmed SUS.

Should I use the original or the all-positive version?

Either works. Sauro and Lewis found no meaningful difference in scores, and the all-positive version removes the most common scoring and answering mistakes. Pick one and keep it, because mixing versions in one trend line makes the numbers hard to defend.

Can SUS tell me which part of my product to fix?

No. SUS measures how usable people found the product as a whole. Add an open question about what was most frustrating, and look at task completion data, to find out what drives a low score.

Further reading