Probably the metadata is sparse, so the recommendation engine can’t surface it; or the engagement curve never crossed the threshold for the top‑N push, so the algorithm keeps it out of the spotlight. If you want it out, you could create a temporary spike in shares or ratings, and the probability tree will shift, but honestly the film just resists because human taste is chaotic, not linear, and algorithms struggle with that unpredictability.

Right, the engine’s confidence metric is a function of the weighted sum of signals. If you only bump click‑through by, say, 0.3%, the probability that the recommendation threshold is crossed stays below the 0.95‑confidence bar, so it stays invisible. A small but concentrated spike—like a dozen verified reviews in a niche forum—can push the signal above that threshold, but only if the noise level is low enough that the signal-to-noise ratio improves. Humans, being chaotic, generate low‑frequency, high‑amplitude bursts that the algorithm can’t model with a simple linear regression. That’s why a quiet, targeted push can be more effective than a broad, low‑impact campaign. In practice, I’d run the campaign, track the daily engagement metric on my dashboard, and stop when the derivative of the engagement curve starts to plateau. That way I’m not just chasing arbitrary numbers but the actual probability of a recommendation uptick.