From retention curves to channel-specific evidence.

Measure matched hooks

Creator-supplied experiment videos are analyzed together to identify the primary hook across at least two variants. AI-assisted analysis labels only that hook and a small set of controlled hook attributes. Setup, reveals, payoffs, CTAs, and other later segments are not part of the active Learnings model.

Once YouTube retention curves are available, the system interpolates retention at each hook's start and end timestamps. The Learnings model uses this retention signal; the broader Analytics report can still show watch time, engagement, and subscriber results separately.

Build like-for-like comparisons

Within each matched hook group, every usable pair of variants becomes one comparison row. An experiment with three variants, for example, produces three comparisons: 1 versus 2, 1 versus 3, and 2 versus 3. Within each comparison row, the model refers to the two cuts as the first variant and the second variant. It compares their hooks' viewer-loss rates and records which hook attributes differ. For each attribute, it records whether the attribute appears only in the first variant, only in the second variant, or matches across both.

This keeps the evidence anchored to variants from the same experiment instead of comparing unrelated uploads. Rows with no creative-attribute difference provide no signal and are discarded.

• +1 attribute appears only in the first variant
• −1 attribute appears only in the second variant
• 0 both hooks match, either both have the attribute or neither does

Fit one hook model for each channel

15sec fits a ridge-regression model for the hooks on each channel. Modeling hook attributes together helps adjust for creative choices that frequently appear at the same time, while ridge regularization shrinks unstable coefficients rather than letting sparse comparisons produce extreme estimates.

Pairwise rows from one experiment are not independent. We cap each experiment at ten hook-comparison rows, then weight its remaining rows to a combined weight of one. A five-variant experiment therefore cannot dominate several smaller experiments simply because it creates more pairs.

• At least 7 unique hook experiments are required
• An attribute must vary in at least 3 experiments
• Ridge penalty α = 2

Test whether the direction survives resampling

The model resamples whole experiments with replacement and refits the regression 1,000 times. Resampling at the experiment level preserves the dependence among pairwise rows from the same test.

Direction stability is the share of directional bootstrap coefficients that keep the majority sign. The same refits produce a percentile interval and a finite-sample, two-sided empirical p-value. These statistics describe model stability, not confidence in one individual A/B test.

Translate the coefficient into retention points

The raw coefficient is measured in retention percentage points per second. To make it interpretable, the system multiplies that coefficient by the actual comparison durations where the attribute varied, then reports the experiment-weighted median impact in retention points.

Every bootstrap coefficient is transformed through those same observed durations to form the displayed impact interval. Attributes with an effectively zero estimate, evidence from fewer than three experiments, or direction stability below 60% remain hidden.