How Meerkat works

AI personas try your product in a real browser and tell you where it went wrong. Four things to know before your first run.

A run, step by step

1. You give three things. What to test, as one URL or up to three versions compared side by side. The task, written the way a customer would say it. How many personas try it, and which.
2. The task becomes steps. Written in plain words, it is planned into two to six steps when you press run. You can review and edit them under More options, or start from a template or a saved flow.
3. You can set the scene. Also under More options, a scenario gives the personas a situation to be in, such as "you open next week and need this working before then". Don't go past names where the test should end, such as "buying a phone number": a persona stops before that point, and stopping there counts as getting through, not as a problem. Both are optional.
4. Each persona drives a browser. A vision model sees the screen and decides what to do next: click, type, scroll, go back. It only knows what is on screen, so it has to find its way like a person would.
5. You can watch. The run screen shows the persona's screen with its cursor moving, a ripple where it clicks, and what it is thinking as subtitles. You can scrub back at any point, and replay the whole thing afterwards.
6. The run is reviewed. A second model checks each step against what is visible, groups what went wrong into problems rather than symptoms, and proposes a fix for each.
7. Each persona is debriefed. They tell you how it went in their own words, then answer four questions: what they liked, what they didn't, whether they would use it, and what they would change.

A run takes a few minutes. Three personas on one version is a normal first run.

The personas

A persona is a set of traits, not a character: how much patience it has, how comfortable it is with software, how much it already knows about the product, and how precisely it knows what it wants. Those traits change what it notices and when it gives up.

• Everyday user and Busy power user stand in for the middle of your audience; the first is the one to run if you only run one.
• First-timer and Vague explorer arrive knowing nothing, and find the things your team has stopped seeing.
• Phone-first regular runs on a phone-sized screen with touch.
• Careful evaluator, Domain expert, new here and Non-native speaker read closely, compare against what they know, and stumble over wording.

Every persona is available on every plan. More personas means more independent attempts at the same task, which is how a real problem separates itself from one persona's bad luck.

Reading a report

The outcome

One sentence on how it went, and what worked, at the top.

What to fix

Problems worst first, each with what happened, the personas' own words as evidence, a screenshot, and a suggested fix. Marked by how much it gets in the way, not by a severity score.

What each persona told us

Their run in their own voice, then their interview answers. These are AI opinions, labelled as such.

Step by step

Where each persona got to, where they would have left, and where a run stopped at your limit. A persona that didn't finish says why, and what to change.

The replay

Watch any persona's run again, from the start or from the moment a finding happened.

Findings you keep are tracked per product: mark one triaged, known, fixed or dismissed, assign an owner, or send it to Linear. A later run recognises a finding it has seen before, and tells you if something you marked fixed came back.

Share a report with a read-only link that needs no account, or export it as Markdown.

Credits

Live runs use credits. Everyone gets 3,000 free every month and 10,000 when they sign up, and a persona run usually uses about 1,000, more if it struggles. New run shows about what a run will use and what you have left, and every report shows what it used. Credits belong to you across every workspace you're in, and sample runs are always free. Plans with more credits are on the pricing page.

Limits

• Personas stay on your site. A link that leaves your domain is undone, except to your own sign-in or sign-up pages on a domain carrying your product's name. Sign-in through Google or Okta, and hosted checkouts on another domain, are out of reach; when a step goes on there, the report shows that step as outside the test, not as a problem.
• For flows behind a login, tick I need to log in first under the URL: a browser opens, you sign in, and the persona takes over from there. Meerkat never sees your password. One persona and one version per run.
• Personas never pay. They stop at a payment form, the same way they stop at a limit you set, and never enter real payment details.
• Personas do not solve CAPTCHAs and do not receive email or SMS codes.
• Point sign-up flows at staging. Personas create real accounts and take real actions.
• A run has a budget. A persona stops after about 25 minutes by default, or sooner once it has taken a lot of actions, and reports what it had rather than nothing.
• Reports are evidence, not proof. A persona can misread a screen. The report shows you the screenshot behind every claim so you can check it.