What they are good at
- Wording that does not say what happens next, and buttons named after the system instead of the customer’s errand.
- Paths that are too long: the persona runs out of patience or steps before it reaches the price, the contact or the checkout.
- Missing answers to what people ask before they pay: who runs the site, the total price, refunds, the reply time.
- Pages that change by device: a phone menu that hides the one link people need, or a page wider than the screen.
- The basics of search: titles, descriptions, headings and structured data of every page they opened.
What they do not do
- They do not judge visual design. They read text and controls; colours, spacing and images are outside what they get, and a picture reaches them only as its description.
- They do not send forms on a live site, type passwords or card numbers, or sign in with Google or another service. Sign-up and payment count as done, so a run cannot tell you whether the confirmation arrives.
- They do not leave your domain: a link to another site brings them back to yours.
- Their “would pay” is an opinion, not a conversion forecast.
They can be wrong
A language model sometimes misreads a word, imagines a problem or blames the site for something the test did. Each note comes with the steps and screenshots behind it, and you can mark it true or wrong. We count the share of true notes from those marks and publish it on the page about our model once enough are checked.
Where people are still needed
Use AI visitors before a release, after every change and on pages you cannot test with people every week. Keep interviews with real customers for what only a person can answer: whether they want the product at all, how it feels to use, and what they would really pay. The two work best together: the AI panel clears the obvious problems, so the hours with people go to the hard ones.
