Updated 2026-08-20

Do AI clip scores predict views?

In our own data, no. Across 45 published posts with view counts, the correlation between our scorer's number and actual views is -0.151 — slightly negative. The single best-performing clip we have posted scored 68, our lowest band, and took 1,517 views; several clips scored 80 took between 0 and 90.

What we measured

We took every post we have published from our own accounts that has a view count attached, read the view number from the latest metrics snapshot, and paired it with the score our pipeline gave that clip before publishing.

The result is not a small positive relationship. It is a small negative one.

MeasuredSource
Published posts with view counts45scheduled_posts joined to post_metrics, latest snapshot per post, 20.08.2026
Correlation, score vs views-0.151 (Pearson)computed over those 45 pairs
Best clip we have posted1,517 views, scored 68youtube_shorts post, latest snapshot
Clips scored 800, 0, 1, 8 and 90 viewssame dataset, five posts across three platforms

What this does and does not prove

It does not prove that clip scoring is useless in general, and we are not claiming that. Forty-five posts from one small, new account is a tiny sample, views on a new account are dominated by platform distribution rather than by the clip, and the same clip posted to three platforms appears three times in the data.

What it does show is narrower and still worth knowing: at our scale, on our account, the score did not rank clips the way the platforms did. If you are choosing which clip to post and you expect the number to tell you which one will travel, our own data does not support that expectation.

Why we are not giving you a percentage lift for this: we have not measured one.

So what is the score for

It ranks how hook-like a moment is inside the window you gave it — whether a clip opens on something that stands alone. That is a structural property of the transcript, and it is genuinely useful for picking a starting point from twenty candidates.

It is not a prediction of distribution, and we have a documented example of the gap. On a cooking video, the cook's actual opening line — "Check it out, folks. If you like a chopped cheese, you're gonna love this" — scored 68, last of five. A mid-video aside about breaking up the meat scored 82 and came first.

MeasuredSource
Real opening hook, cooking runscored 68 of 5 clips (last)request 2ced1dde, clip scores 82/80/76/68/68
Mid-video filler, same runscored 82 (first)same run

FAQ

Should I just post the lowest-scored clips then?

No — that would be reading a -0.151 correlation on 45 points as a rule, which it is not. The honest conclusion is that the score is weak evidence about views either way, so use it to shortlist and then pick by what the clip is actually about.

Why publish a number that makes your own feature look worse?

Because the alternative is letting you plan around a promise we cannot support. The score does one job well — ranking hook-like structure inside a window — and we would rather be precise about which job that is.

Will this change as you get more data?

Probably. When the dataset is meaningfully larger we will update this page and say what changed and on what date. The numbers above are from 20 August 2026.