How usability actually gets measured
Usability is measured more often than it is improved, partly because the easiest measurements are the least actionable. It is worth being clear about what each common method sees, what it costs, and what it structurally cannot detect.
Asking is not watching
Satisfaction surveys are cheap, scale well and produce a number that can be tracked over time. They are also filtered through memory, politeness, and the respondent's theory of what the question is for. People routinely report that a system is fine and then, observed, take four attempts at a task they described as straightforward.
The value of asking is in the free-text answers rather than the score. People are poor witnesses to their own performance and excellent witnesses to their own irritation, and irritation reliably marks the places worth observing.
Task-based testing
The core method has not changed much: give a representative person a realistic task, watch them attempt it without help, and record what happens. It is labour-intensive and it finds more than everything else combined, because it is the only method that observes the failure rather than a trace of it.
Its main risks are a task list written by the people who built the screen, which tends to describe the path they intended, and a session conducted in conditions unlike the real environment. A test of an interruption-heavy workflow conducted in a quiet room with no interruptions has tested something else.
Time on task, error rate, completion
These three are the workhorse quantitative measures, and they should be read together. Time alone rewards a screen that lets people finish quickly with the wrong value. Error rate alone rewards a screen so slow that nobody makes mistakes. Completion rate alone hides the difference between finishing easily and finishing after three attempts.
The most useful comparisons are within a system over time, not between systems. Absolute numbers depend heavily on the task written, the people recruited and the setting, none of which travel.
Standardised questionnaires
Standardised instruments exist so that a subjective score can be compared across studies rather than invented per project. Their strength is comparability and a large body of prior use; their weakness is that a single score cannot tell you which part of the interface produced it.
They belong at the end of a session, after observation, as a summary rather than a diagnosis. Treated as the primary measure, they turn design work into score management.
Logs and what they cannot see
Instrumented use data is the only method that observes everyone, all the time, in the real setting, and it is very good at showing where people are slow, where they repeat themselves and which routes are actually used rather than documented.
It cannot see intent. A long pause is deliberation, confusion, or a conversation. A repeated action is a mistake or a habit. Logs are unmatched at finding the place to look and are almost useless at explaining what is happening there, which is why the two methods are usually run together: logs to locate, observation to explain.