Captions step aside, and your site shows up in 3 seconds

We went back through a batch of recent films and watched them frame by frame, asking one question: what does a viewer actually experience? A few patterns stood out.

• A caption sometimes covered the very button, price or heading the voice had just named.
• Some films spent several seconds on drawn text before showing the product.
• A view with nothing to zoom into could stay still for a long stretch.
• Bright yellow and lime brands opened on a muddy olive card.

Each one is a small thing. Together they decide whether a film feels crisp or slow, so we addressed all of them in one release.

Our captions already lifted above a subject sitting low in the frame, but the lift was calculated from an estimated caption height. Now we measure the real caption area, which at full HD is closer to 244 pixels than the 150 we had assumed, and the lift works for every kind of shot: hovers, spotlights, opening views and plain recordings.

In vertical films, when the recorder cannot place a heading clear of the captions, it says so, and the caption moves above it. A lifted caption sends the free-film badge to a top corner, so the two never stack. The soft shade behind captions now starts exactly at their top edge and follows the page's brightness, and long vertical lines are re-broken before they reach the edges.

A film about a website should show the website early. Wide films now reach a real recording by their first scene and vertical films by their second, with the opening line spoken over the footage. A storyboard is refused if more than three seconds of drawn text come before the first recording, and typewriter openings finish typing within 2.4 seconds.

For still views, the camera now cuts on the narration's words between the full frame and a gentle 8% push-in, so the picture moves with the voice. On our test film, the longest still moment went from over eight seconds to three. Bright yellow and lime brands open on a dark background with their colour as the accent, which suits them far better.

Several changes make sure what you hear matches what you see. The recorder reads the main part of each page, ignoring header and footer, to recognise a loading, sign-in or error screen; such a shot is taken again or replaced, never kept. The planner never narrates cookie banners or pop-ups, and never says "here is X" unless the shot really frames X.

A scene whose filmed subject shares no word with its narration is given a subject that does. And figures on cards are always the ones the voice says, in full: a card never shows 663 while the voice says 663,000, and a unit never runs into its number.

Every film gets an automatic quality review from a vision model. Pace, though, is something you feel over time, and a model looking at single frames cannot see it. So a new step, filmMetrics, measures it directly:

1. 1how long until the first real recording;
2. 2how much of the film is real footage of your site;
3. 3how many scenes per minute;
4. 4the longest moment where nothing moves, measured on the final video at four frames per second.

Those numbers now count in the quality score. The vision review looks at every scene and at the opening seconds, and it no longer overlooks a caption covering the element the voice names. A new test suite checks these rules on real rendered pixels, not only in code.

Your next film reaches your real site sooner, keeps its captions off what matters, moves with its voice, and is judged on how it feels to watch, not only on how its frames look. You can see the result in our example films.

The lesson we keep relearning: anything a viewer feels over time, such as pace, stillness or a slow opening, has to be measured over time. A stricter ruler is also what tells us which changes really help.