How the page is read
Our server fetches the public page once (no logins, no private addresses), extracts the title, description, headings, short paragraphs, the most used colors and the social preview image. That extract — not the whole page — goes to the model, marked as untrusted content, so text on the page cannot steer the generator.
The video is built from that material only. If the page does not state a number, a customer or a feature, the video will not invent one. Want a specific angle? Add it to the brief.