Questions, answered honestly.

The clip's structural psychology: hook, pacing, emotion, curiosity. Not your distribution, your audience, or luck.

Paste a YouTube link, or upload any clip: Reels, Shorts, TikTok, LinkedIn. The platform is detected from the link automatically.

A full diagnosis takes 3 to 5 minutes. Longer videos add a little processing time.

Castle scores clips up to 60 seconds with the full rubric. Paste something longer and it offers to score the first 60 seconds, where the hook lives. Videos up to an hour can be trimmed this way.

Yes, a free one. It takes one click with Google and never asks for a card. Every report you run is saved.

It is a diagnosis of how the video is built, grounded in named research, not a guarantee of the outcome. On 211 real Shorts Castle had never seen, the score's link to views relative to each channel is small but real (a correlation of 0.13).

Close, not identical. Scoring the same video again moves the overall number by about 4 points: a 5-video probe at production settings measured 3, and 211 videos each scored three times measured 3.5, which we publish rounded up. Treat any change under 4 points as noise. The product tells you this too.

Yes, in September 2026, on 211 held-out Shorts from a corpus of 491 across 170 channels, split by channel. The result is honest and modest: a small, real correlation with views relative to the channel (0.13, with a confidence range from just above zero to 0.27). How the score is made is on the How scoring works page.

Neither a secret dataset nor a checklist. Castle is a multimodal AI engine reasoning over a rubric built from published psychology research. There is no training corpus, and that is a choice: a trained black box gives you a number; this approach gives you the number and the reason, so you know what to change.

The factors come from published research by named researchers: Berger, Loewenstein, Heath and Heath, Green and Brock, Itti and Koch, Schultz. The weighting between them is ours, and we openly call it a version-one hypothesis that our pre-registered evaluations keep in check.

A panel of simulated viewers grounded in 1,000,000 real demographic records. Panelists react to your clip's actual frames and transcript. Simulated, not measured, never claimed to be representative, and labeled that way in the product.

Every score names its mechanism and the evidence it was read from. Estimates are labeled estimates, simulated is labeled simulated, and anything under 4 points of difference, the published re-run band, is noise, and the product says so.

No. Processing is transient. We keep the score and the metrics, never the clip.

Only people you share the link with. Share pages render without a login, but the link itself is unlisted.

Yes, and it stays. Every account gets 5 minutes of video analysis each month at no cost, with the same full report as every paid plan.

By the video time Castle analyzes. Each plan includes a monthly allowance of minutes or hours, and a scored video uses only the seconds actually analyzed. Refused submissions and errors do not count against you.

Billing runs through Dodo Payments and you can cancel any time from your account page. The full refund terms are on the Refund Policy page at /legal/refunds.