Run a usability test on a live website

How website usability testing works with AI personas: task setup, viewport choice, live view, and reading the results.

Versive's website tests use an AI persona instead of a recruited participant. The persona operates a browser on your live, publicly reachable site and works toward a task you define. Each run returns an action log with screenshots, reasoning, and a session recording, typically within minutes.

The quality of those results depends on a narrow task, a representative viewport, and careful review of repeated runs. Higher-stakes findings still require confirmation with real participants.

How a website test works

A website test is one way to run an AI test in Versive, alongside testing a Figma prototype or static design images. A prototype test moves an AI persona through selected imported frames. A website test points the AI at a live URL and lets it operate a browser against the DOM, including the site's loading states, copy, and current bugs.

You can watch a run in live view. Each navigation, click, typed entry, or scroll is logged with the affected element, a screenshot, and the reasoning behind the action. Use the log to trace why the AI chose a particular link instead of inferring its path from the final result.

Configure the test

A website test needs a handful of settings before you run it.

Website URL is the page where the AI persona starts. The site has to be publicly reachable. A staging environment works if no VPN or login wall prevents the AI from reaching the page.

Task is what you want the AI user to accomplish, written as a concrete goal rather than an open invitation to browse. The docs' own example is a good model: "Find a two-bedroom apartment in Austin under $2,000 and start an application." That single sentence gives the AI a destination and gives you something you can score as done or not done.

Viewport controls the browser's screen size. Desktop options are 1920×1080, 1536×864, 1366×768 (the default), 1280×720, and 1024×768. Use 500×1080 to simulate a mobile screen. Match the viewport to your traffic because a layout that works at 1366×768 may fail at 500 pixels wide.

Test sections are the dimensions each run gets evaluated on. Versive defaults to Initial Impressions, User Action & Behavior, and Recommendations, which cover first reactions, what the AI actually did, and what it suggests changing.

Personas are the AI users who run the test. Add one or more from your persona library or write them inline, then set an execution count for each. What are synthetic users, and when should you trust them? explains why a single persona's result needs careful interpretation.

Extended limit raises the per-run cap from 50 browser steps to 100 for journeys that need more clicks, such as a full multi-page checkout. It can cost up to two credits instead of one, so enable it only when the task needs the extra steps.

Write tasks that produce a real signal

Task wording has a large effect on the usefulness of a website test.

Give the task a verifiable end state. "Add the cheapest annual plan to the cart" produces a clear success or failure. "Explore the pricing page" leaves many valid paths, so the result cannot distinguish a usability problem from reasonable browsing. If you cannot describe "done" in one sentence, revise the task before running it.

Start with a critical path such as sign-up, checkout, search, or booking before testing exploratory browsing. Prioritize flows where failure directly prevents a user from reaching an important goal.

Keep each test narrow: one flow with one clear goal. A broad "look around and tell me what you think" tour produces commentary that is difficult to attribute to a step. Several focused tests are easier to interpret and cheaper to rerun after a fix. Usability testing questions: 40 examples that work provides questions for examining each step beyond a pass-or-fail outcome.

Watch it run, then read the results

You can follow a run live or review it after completion; both options use the same recorded evidence.

Each run reports a task completion status with a confidence score, and if the AI gave up partway through, the reason it abandoned the task. Below that sits the full step-by-step replay: the click path with a screenshot at every step, plus a complete session video recording you can scrub through the way you'd review a screen recording from a real participant.

Findings are organized per test section, and above the individual runs Versive builds an aggregated summary with key takeaways and recommendations, each one linked back to the specific runs that produced it, so you can go check the exact step a finding came from rather than taking the summary's word for it. If you turned on the accessibility analysis, you also get WCAG-based checks with an impact breakdown and concrete recommendations.

Run more than once before acting. A single run can surface a real problem, but three to five executions per persona help the aggregated summary distinguish repeated findings from one-offs. Results are attributed per persona, so testing a first-time visitor and a returning customer can reveal differences between those user types. Add runs after a fix and reuse the same personas in the project to check whether the result changes. See the AI tests overview for how projects and credits work across all test types.

When to trust it, and when to bring in real users

AI website tests can quickly surface confusing flows, unclear copy, missing affordances, and heuristic violations. Their short turnaround makes repeated checks after changes practical.

Treat the results as hypotheses ranked by likelihood. An AI persona is consistent, but it is not a customer with lived context and cannot establish what a person would pay, prefer, or do outside a test. Confirm high-stakes findings and research about behavior, attitudes, or willingness to pay with real participants in a study. Moderated vs. unmoderated usability testing compares ways to structure that follow-up round.

Create an AI tests project with the live URL, define one narrow task, and run it with one or two personas. Review the session recording and findings summary before adding more tasks.