UX metrics: how to measure user experience, and which numbers to trust

1. What a UX metric is

A UX metric is a number that describes how people experience a product: whether they can do what they came to do, what it cost them in time, attempts and help, and what they think of it afterwards. UI metrics are the subset that describe the interface itself, such as where the first click lands or how many screens fail a contrast check.

There are a lot of them. MeasuringU's taxonomy lists more than 70 and suggests starting with a few versatile ones, among them task completion, the Single Ease Question and the UX-Lite questionnaire.

Before choosing any of them, sort them by kind, because the kind decides what a number can tell you.

2. Behavioural and attitudinal metrics

Every UX metric answers one of two questions. Behavioural metrics ask what people did: did they finish, how long did it take, where did they stop. They come from logs, recordings and observed tests. Attitudinal metrics ask what people say about it: was it easy, would they recommend it, do they trust it. They come from surveys, interviews, reviews and support tickets.

RUCF uses its own names for the same split: friction (B, behaviour) and perception (P).

The two kinds disagree, and the disagreement is large. MeasuringU compared self-reported and verified completion on 34 tasks. Participants reported completion rates almost three times higher than the verified rates, and the two correlated at r = 0.24. In 28 of the 34 tasks, more than 90% of participants said they had succeeded. A meta-analysis of 106 effect sizes outside UX found self-reported and logged digital media use correlated at r = 0.38 (Parry et al., 2021).

So an attitudinal metric cannot stand in for a behavioural one. It measures the experience people remember, which is shaped by how it ended and by who they blame when something goes wrong. That is useful to know, and it belongs in a separate column.

3. Behavioural UX metrics

Task success rate

The share of attempts that reach the goal, also called completion rate. MeasuringU's reference set of 1,189 tasks from 115 usability tests averaged 78%. A good rate depends on what failure costs: a payment or medication flow needs to be close to 100%, a low-stakes browsing task can live lower. Define success before you test, in the user's terms, and verify it from the outcome, not from the participant's answer.

Time on task

Task times are skewed. Most people finish in a similar time and a few take ten or twenty times longer, which drags the arithmetic mean upward. For samples under 25, Sauro and Lewis found the geometric mean estimates the typical time better than the median. Report the slow end as well: the 90th percentile time tells you how long the slowest tenth of your users wait, and that group is often where the real problem is.

Errors and recovery

Count errors per attempt, then count how many people get back on track without leaving or asking for help. The second number is more useful. A form that produces many errors but recovers cleanly costs less than one that produces few errors and loses everyone who hits them.

Drop-off and abandonment

The share of people who leave at each step of a funnel. Read it step by step. A total conversion rate hides which step is losing people, and the step with the biggest drop is the first place to look.

Assistance

How often people need help to finish: support contacts per thousand active users, help-centre visits during a task, internal searches before completion. Read the last 50 support tickets and count the ones that ask something the interface should have answered.

Activation, depth and retention

Whether people reach first value, use more than the entry feature and come back. These are behavioural too, measured over weeks instead of seconds. Together, time, attempts, assistance and abandonment are what RUCF calls the four currencies of friction.

4. Attitudinal UX metrics

System Usability Scale (SUS)

Ten statements answered on a five-point scale, converted to a score from 0 to 100. It was published by John Brooke in 1996 and is still one of the most widely used standard questionnaires. The average across about 500 studies is 68, which sits at the 50th percentile; a score of 75 sits at about the 73rd. SUS is not a percentage, so a 68 is average, not a failing grade.

Single Ease Question (SEQ)

One question after each task, "Overall, how difficult or easy was the task to complete?", on a seven-point scale. Across more than 400 tasks the average sits between about 5.3 and 5.6. Because it is asked per task, it is the attitudinal metric that pairs most cleanly with task success and time.

Net Promoter Score (NPS)

One question, "How likely are you to recommend us?", on a scale from 0 to 10. The score is the percentage of promoters (9 and 10) minus the percentage of detractors (0 to 6). Fred Reichheld introduced it in Harvard Business Review in 2003 as the one number a company needed to grow. A longitudinal study by Keiningham and colleagues in 2007 found it did not predict revenue growth better than other customer measures. NPS describes a relationship. It says little about where a product is hard to use.

Satisfaction, reviews and tickets

CSAT questions, app store reviews, interview notes and the wording of support tickets are all attitudinal sources. Text has to be coded before it becomes a number: count the passages about a theme and the share that are negative. If a model did the coding, check a sample before you trust it.

5. UI metrics

Interface design can be measured with the same discipline as the rest of the experience. Four UI metrics cover most of what a design review argues about:

• First-click success. The share of people whose first click lands on the element the screen is designed around. A first-click test with a static mock-up is enough.
• Contrast and legibility. The number of WCAG 2.2 AA contrast failures on the main flows. Normal text needs a contrast ratio of at least 4.5:1, large text 3:1. Free scanners such as axe, WAVE or Lighthouse count them in minutes.
• Keyboard completion. How far the primary flow can be completed with a keyboard alone.
• Naming consistency. How many different names the core concept has across interface, help pages, marketing and support replies.

Each has a perception partner: whether users describe the screens as clear or busy, whether they mention reading difficulty, whether they can say what the product is in a sentence. Accessibility conformance is different from the rest and should sit outside any composite score. A product that fails WCAG needs a remediation list, and averaging the failure into a score lets a strong product hide it (more on that here).

6. HEART and other UX measurement frameworks

Google's HEART framework (Rodden, Hutchinson and Fu, 2010) groups metrics into Happiness, Engagement, Adoption, Retention and Task success, and asks teams to move from goals to signals to metrics. It is a good way to choose what to measure. ISO 9241-11 defines usability as effectiveness, efficiency and satisfaction, which maps onto the same split: the first two are behavioural, the third attitudinal.

Both frameworks put behavioural and attitudinal metrics in one list and leave open what to do when they point in different directions. In practice the answer is usually "fix the lowest number". That is the step RUCF replaces.

7. Why you should not average UX metrics

Take two parts of a product, both scored from 0 to 100, where higher is better.

The averages are identical and the problems are opposite. In α, no survey will ever find the cost, because the survey is asking the variable that is wrong. That is silent friction, and the treatment is to remove the costly step. In β, rebuilding the flow would spend money on something that already works. That is phantom friction, and the treatment is to change what the product says about itself. A dashboard that shows only 55 cannot tell these two apart (why averages hide the diagnosis).

8. How to measure UX with RUCF

RUCF, the Real User Cost Framework, is a published method for reading behavioural and attitudinal metrics together. In five steps:

1. Name the task. Write down the job users come to do and what counts as success, in their words.
2. Pair the metrics. For each of four surfaces (Task, Commitment, Access, Belief), pick a behaviour metric and a perception metric about the same thing. Collect them separately, and never ask users to estimate their own behaviour.
3. Weight the evidence. Tag every number with its source class: observed (logs, recordings, scans) 1.00, reported (surveys, interviews, coded tickets) 0.70, inferred (analytics proxies, AI classification) 0.50, assumed (expert or team judgment) 0.30.
4. Read the matrix. Score both sides from 0 to 100 and place each surface on the say-do matrix. Above 58 on both is flow, below 42 on behaviour with high perception is silent friction, the reverse is phantom friction, below 42 on both is loud friction, and anything in between is a reading to measure again before acting.
5. Report with confidence. Write the score with its confidence, for example 3.4 @ 0.52, report the median and the tenth percentile, and multiply the confidence of a surface by 0.8 for every release that touches it.

The full arithmetic is in the scoring model. The free scorecard runs it in two minutes with your own numbers, and the RUCF agent skill runs it inside Claude, ChatGPT, Gemini or Grok on your exports and survey files.

9. 17 common UX metrics, mapped to RUCF

Where each metric belongs, and the weight it usually carries. The class depends on how the number was gathered: a task success rate read from logs is observed, the same rate estimated in a meeting is assumed.

The last row surprises people. Different experts reviewing the same product report very different problems, which is why RUCF gives expert judgment the lowest weight (the evidence behind that).

10. Where to start if you measure nothing today

Pick four pairs, one per surface, and collect each side separately:

• Task. Completion rate of the main task from your funnel, and the SEQ asked right after it.
• Commitment. Month-over-month retention, and one survey item on whether people intend to keep using the product.
• Access. An automated contrast scan and a keyboard run of the main flow, and whether users mention difficulty reading or using it.
• Belief. Drop-off at the payment or data step, and one survey item on trust.

Eight numbers are enough to place each surface on the matrix. If one side of a pair is missing, that gap is the first thing to measure, and the confidence figure will show it.

11. Common questions about UX metrics

What is a good SUS score?

68 is average, at the 50th percentile of about 500 studies. A 75 sits around the 73rd percentile. Compare against your own earlier scores and direct competitors before you compare against the global average.

What is the difference between UX metrics and UX KPIs?

A UX metric is any number that describes the experience. A KPI is a metric the business has tied to a goal and a target, for example raising checkout completion from 62% to 70% this quarter. A good UX KPI is behavioural, with an attitudinal partner reported next to it.

Is NPS a UX metric?

It is an attitudinal loyalty measure. It can show that people are unhappy. It cannot show where in the product the cost sits, and its claim to predict growth is disputed. Use it as one perception reading next to behavioural data.

How many users do you need to measure UX?

For a usability test, a handful of participants will show the big problems, but the numbers come with wide intervals: if 10 out of 10 complete a task, the true rate is somewhere between 75% and 100% at 95% confidence. For retention signals and a meaningful tenth percentile, RUCF needs behaviour data from a few hundred active users.

Can AI measure UX?

An AI assistant can classify reviews and tickets and walk through public pages. It cannot watch your users. RUCF weights AI classification of text at 0.50 and expert judgment, an AI's included, at 0.30, so an AI-only audit should come back with a low confidence figure and a list of things to measure. The RUCF agent skill works that way.

Iordanis Passas is a creative director and type designer based on Syros, and the author of RUCF.

Find your gap

---

Related: The say-do matrix · Stop reporting average UX · The four currencies of friction · RUCF agent skill