A website can have green tests, answer correctly on every address, and still meet your customer with a button covered by something else. Nothing fails. Lens opens the page in a real browser, does what the customer would do, and keeps the picture as proof — and no change goes live before it has seen it.
The difference is not academic. A test can only see what it was told to look at — and it looks through the code. The costliest failure in software is not the one that shouts; it is the one that succeeds without having any effect. The page answers. The log is green. The button sits behind another button.
| Question | Test | Lens |
|---|---|---|
| Was the function called? | Yes | — |
| Does the page answer? | Yes | Yes |
| Can you SEE the button? | No | Yes |
| Is anything covering anything? | No | Yes |
| Is the text cut off? | No | Yes |
| Does it work in Safari? | No | Yes |
| Does it work on a phone? | No | Yes |
The whole page, the visible window, or one specific element. Sign in first if it is behind a password.
“go to the front page · click Book · type the name · expect it to say Thanks”. 25 actions, written as sentences.
Chrome, Safari or Firefox. On desktop, tablet or phone. With the right language and the right timezone.
A screenshot per step, kept durably. Not “it worked” — a picture someone can open.
A pixel diff can say something moved. It cannot say whether the result is usable. Lens measures the geometry of the finished page: is anything sitting on top of anything else, is a line of text clipped, has an element ended up off the screen, is the contrast too low to read. A high-severity finding turns the run red and stops the delivery.
The geometry is free and always on. If you want a model to look as well, it runs on Mistral in Europe — not on a US service.
A change is not finished because an AI agent says it is. It is finished when Lens has opened it in a browser and can show the picture. If the check fails, the work goes back to the agent — and round again.