How it works
Each simulated person goes through the site step by step in a real Chromium browser. A language model decides what to do, but what the person sees and whether the goal was reached are checked separately — we do not leave that to one model.
1. What the person sees
Before each step we take a screenshot and mark the elements that can be acted on. After the step we compare the page state: did a new page load, what appeared, what disappeared, did nothing change. This part uses no language model — it is measurement. The profile affects what the person notices: someone who skims does not get long paragraphs; someone less web-savvy does not recognise an unlabelled icon.
2. How they decide
First the goal is broken down into signals a person would recognise as success. Then at each step the model, playing the profile, chooses one action — click, type, scroll, go back — and says what it expects to happen. The person has a memory: notes on what they have found and where.
3. Who oversees the run
- Monitor — rules, not a model: it counts attempts without progress, detects going back and forth between the same states, long scrolling and pages that do not react. The profile's traits set how much patience the person has.
- Verifier — independently checks whether what is on screen really meets the goal. Saying “found it” is not enough.
- The point of giving up — when a person runs out of patience we record that point, but the study goes on to see the rest of the path. That is why results separate “got there” from “got there, but would have given up on the way”.
4. How the result is built
Across many runs we gather the site's problems and check which traits are over-represented among those affected. We compute the effect of each trait on success and mark differences too small to draw conclusions from. The report is written in the language chosen when the study starts (English or Polish).
Safety of the studied site
The agent does not submit forms — submissions are intercepted in the browser, so nothing is saved on the studied site. If a site does not let automated visitors in (e.g. bot protection), the result says so instead of pretending to study it.
Limitations
A simulation does not replace people: it does not know their emotions, history or what happens off screen. Results are probabilistic — which is why we study a population rather than one person, and show when differences are within chance.