Find out it does not work while it is still a file.

Everyone nodded.
Nobody completed the task.

Realistic enough that people behave normally: real copy, real data volumes, working states. Users forgive rough visuals and stumble badly on fake content.

Goals written so they do not contain the answer. "Find the export button" tests nothing; "send last quarter to your accountant" tests everything.

Per task, per participant: finished unaided, finished with a nudge, or gave up. Plus how long it took against your baseline.

Every problem with severity, how many participants hit it, and the specific screen it lives on. Sorted so the first three are obvious.

What we would change and what we would leave. Some findings are real and still not worth fixing, and we say which.

Thirty-second cuts of the moments that matter. One clip of a customer failing ends a roadmap argument faster than any amount of analysis.

The flow with the highest drop-off and the least observation. Five sessions here typically pays for the whole engagement.

Anything with a completion rate attached to revenue, where a percentage point is worth more than the test costs.

Testing the new version against the old on the same tasks, so "it feels better" becomes a number your board accepts.

Rather than debating which concept is stronger, put both in front of eight users and let the completion rates decide.

Rule builders, permission screens and setup wizards, where the designer understands the model and nobody else does.

Where a misunderstanding has consequences beyond conversion and "users should have read it" is not a defence.

Synthetic users arenot users.

Realistic data, generated in minutes

Plausible names, volumes and edge-case records instead of Lorem Ipsum. Users behave differently in front of real-looking content, and generating it used to cost a day.

Every hesitation timestamped

Transcripts aligned to screen recordings and tagged for confusion, backtracking and abandonment, so a researcher reviews moments rather than scrubbing footage.

What repeated across eight sessions

Cross-session clustering that surfaces the issue three people hit in different words. The judgement about severity stays with the researcher.

A model cannot be confused by your interface.

It has no goal, no impatience and nothing at stake. Every finding we report came from a person attempting a real task.

• No synthetic participants, ever

  Every completion rate we report comes from a human session. We will not run an AI panel even if asked, because the number would be fiction.

• Tags are reviewed before they count

  An automated confusion tag is a pointer to a timestamp. A researcher watches it and decides whether it was a real problem.

• Participant data is consented and masked

  Recordings processed under your consent terms, faces and identifiers redacted before anything is shared internally.

• Severity is a human call

  How badly a problem matters depends on who hit it and what it cost them. That is judgement, not clustering.

When testing
will not help

Usability testing measures whether someone can do a thing, not whether they would choose to. Desirability needs different research.

With real traffic, an A/B test gives you a truer answer at higher volume. Save moderated sessions for understanding why the losing variant lost.

A new illustration or a revised colour rarely changes task completion. Testing it produces a null result and a delay.

If the build is committed and the deadline immovable, testing produces a document of things you already cannot change.

Finished unaided, finished after a nudge, or abandoned. Reported per task, because an average across tasks hides the broken one.

Against your current baseline where one exists. A flow that gets slower is a finding regardless of how it looked.

How often people went wrong, and whether the interface let them get back. Recovery matters more than prevention.

What people looked at and clicked before the correct control. Usually explains the completion number better than anything they said.

How badly each problem affected the task, and how many participants hit it. These two together set the fix order.

Which recommended changes were made, and what happened to completion in the next round. Testing without a re-test is half a loop.

What gets tested,
and what changes after