When a page will not load, you get your credit back

Every film starts with a probe. A real browser opens the address, waits for the page to paint, reads its text and pictures, and hands what it found to the planner, which writes the film.

That makes the probe the cheapest place to be honest. If a page has nothing to say, the right film is no film, and the probe is the last moment where saying so costs nobody anything. Everything after it, the plan, the voice, the recording, is work done on your behalf.

A language model given nothing does not stay silent. Hand it an empty document and it will write a confident film anyway, about a page it never saw.

We wanted that to be impossible. Every claim in a PageToVid film should trace back to the site, and an empty page has no claims to trace. So the probe now checks that a page truly arrived before the planner sees a single word of it.

A small new check now runs at the probe. It refuses three cases as an unreachable site, with a sentence that names the page and says your credit is returned:

1. 1Chrome's own error page, after one retry, because some of those are a passing network fault.
2. 2A blank tab left behind when the link started a file download instead of opening a page.
3. 3A document that is still almost empty, with barely any text, pictures, video, canvas or frames, after the patient wait we give a page to paint.

The order matters. Some web apps take their time: one we checked live was blank at six seconds and perfectly fine at twelve, so the third rule only applies after the longer wait. PDFs shown in the browser's own viewer are filmed as before.

The same review looked at the other refusal the probe makes: a page that asks for a password. A sign-in wall is real, and we cannot film past it from the outside.

But many ordinary pages carry a closed login menu in their header, with a password field hidden inside it. An about page or a community site with thousands of words of content is not a wall. Two changes follow:

• only a password field a visitor can actually see counts as a sign-in;
• the probe looks for a real sign-in screen before planning, so a true login page is refused in seconds, before the plan and the voice are paid for.

Content pages with a login in the corner are now filmed like any other page. For a page that genuinely sits behind a sign-in, the Chrome extension records it from your own browser.

Unit tests alone would not have been enough here. An early version of the blank rule refused a PDF, which is exactly the kind of mistake a refusal must never make.

So every rule was run with the real probe against real pages of both kinds: pages it should refuse, and pages that have always filmed well, pagetovid.com among them. A refusal rule is only good if it says no to the right pages and yes to all the others.

If a page cannot be filmed, you hear it at once, in plain words, and your credit comes back. If it can, nothing changes: the probe reads it, the planner writes from it, and the film shows what is really there.

It is a small promise with a large effect. A product that would rather refuse than invent is one you can hand any address with confidence. You can read what PageToVid can and cannot film on the features page.