AI testers modelled on real Americans, and AI shopping agents, work through a task on your site while thinking aloud. You get where they got stuck, why, and what to change, in about 20 minutes.
Pay as you go from $25. Credits never expire.
Point at a URL and write the task the way a customer would say it. Add what "done" looks like. No scripts, no selectors.
Dale, Greg and Mia, or an audience from census data: seniors, nurses, small-business owners. Each tester reads, hesitates and mis-taps like that person would.
Each tester's path, out loud. Where people stumbled, ranked by how much it cost them, with a concrete fix for each. Then run it again and see what improved.
Testers are drawn from NVIDIA's Nemotron-Personas-USA, a million synthetic Americans matched to Census data. Each gets a pace (novice, average or fast) with real reading speeds, thinking pauses, typos and slips, so a 70-year-old on a phone does not behave like a developer at a desk.
Send browsing, vision and tool-using agents on real errands: find and compare, build a cart, request a quote, reorder. See where they get blocked and get an agent-readiness checklist.
The AI decides what to do; our code measures how long it would take a person, using published models of reading, pointing and typing. Calibrate it against your own Maze results if you have them.
Runs of the same mission form a series: success rate, time on task, misclicks and ease over time, with a note on what changed. Fixed and new problems are tracked across runs, and your CI can fail a build when a target is missed.
| What | Credits | Good for |
|---|---|---|
| Quick tester try | 1 | CI on every deploy, big audiences |
| Standard tester try | 3 | Most tests |
| Deep tester try | 6 | Hard tasks, expert users |
| Run report: findings, fixes, moderator review, progress | 2 | Once per run |
A typical run with three testers is 12 credits, about $5. One human tester on a recruited panel costs around $45.
No. Testers stop before any step that pays, sends, deletes or places an order, and the report scores how far they got. You can allow the final step only on a site you mark as a test environment.
Testers never solve a CAPTCHA. When a site needs a person, the run pauses and someone on your team signs in or passes the check in a live view of the same browser. The session is saved for the testers and for later runs.
Testers say who they are (AnyVisitorBot in the user agent) and some bot walls will still refuse them. If it is your site, allow the AnyVisitorBot user agent; here is how.
No. It finds the obvious problems in minutes, so the sessions you run with real people are spent on the hard ones. Each report says how far to trust it.