Exactly how the score is made.

The frames. Every frame. Densest in the first 3 seconds, where the watch-or-scroll decision happens, then even coverage of the rest.

The sound. Volume, beat, sudden changes and silences. The things that make a thumb stop, or keep moving.

The words. A transcript of what is said and shown on screen, plus the caption you plan to post with it.

From that, Castle returns the eight factor scores with findings tied to specific moments, a second-by-second retention prediction, the timestamp where your hook peaks, and the fixes, ordered by impact. A full diagnosis takes 3 to 5 minutes.

Each one comes from decades of published research on why people watch, remember and share. Castle scores all eight on a single clip, at the second they happen.

Ask most AI tools the same question twice and you get two different answers. That makes them useless for measuring anything. Castle is built differently. The AI has one job: to watch your clip and judge each of the eight factors from what is actually in the frames, the sound and the words. How much each factor counts toward the final number is set by Castle's scoring model, not by the AI, and it does not change from one run to the next.

This is the whole point. A 41 today means the same thing as a 41 next month. Two clips can be compared and the comparison holds. And when you tighten the hook and score it again, the number moves because the video changed, not because the model did. You can see your fix working.

Score the same video twice and you will land within about 4 points. Anything smaller than that is noise, and the report tells you so instead of pretending to a precision it does not have.

The hook counts most because it decides whether anything else gets seen.

It is not a number to admire. It is a to-do list. Every score comes with the reason behind it and the change that would move it most.

Exactly where you're losing people. The second your hook peaks, and whether that is early enough. Every drop-off, with the reason next to it: a lull, a late reveal, a flat middle. Which of the eight factors is holding the video back most.

Exactly what to do about it. The one change most likely to lift the score, first. Then the next two, in order of impact. When the hook is the problem, three to five new openings written for your topic.

How it lands before it's live. A second-by-second prediction of how many viewers are still watching. A simulated panel's reaction: the average, the spread, and what they said. How the same clip plays in ten different markets.

A number you can act on. Consistent from run to run, so you can measure a fix. Comparable across clips, so you can pick the stronger cut. Explained at every step, so you're never guessing why.

Castle scores the part you control: the video itself. Post at the wrong hour to the wrong audience and even a strong clip can stall. But a weak hook loses everyone, everywhere, and that is the part Castle fixes before anyone sees it.

A panel of simulated viewers grounded in 1,000,000 real demographic records. They react to your actual frames and transcript. You get the average, the spread, a split by region, and what individual viewers said.

The panel is matched by region rather than by country. It is not a measured audience and does not claim to represent one, and it is labeled simulated everywhere it appears.

Questions this page did not answer live on the FAQ.