Top-of-Screen Promotions

• Top-of-screen promotions are a digital display strategy that uses prime banner space above the fold to drive user engagement and content trial.
• RCT-based evaluations demonstrate that these promotions yield modest, statistically significant improvements in viewing time, especially for short-form and sequential content.
• Optimization frameworks, including multi-armed bandit algorithms and rolling-horizon methods, effectively allocate banner space under dynamic user behavior and supply constraints.

Top-of-screen promotions are a display strategy in digital platforms where prime above-the-fold banner space—typically the home screen’s most visible region—is allocated for promoting selected content or advertisements. This slot occupies a unique position for influencing user engagement and has significant operational and econometric complexities in both deployment and evaluation.

1. Platform Context and Design of Top-of-Screen Promotions

Prominent video streaming platforms such as ABEMA utilize a home screen “top-of-screen” banner area that alternates between promotional banners for target content and standard advertisements. ABEMA, a major Japanese streaming service with approximately 30 million weekly active users, exemplifies this structure. While the core function is content promotion, the slot may also be used for third-party advertising, with treatment assignment (promotion vs. standard ad) at the individual user level.

The target content types commonly include:

• Long-form, self-contained series (e.g., 46-minute comedy episodes)
• Short-form highlights (≈5 minutes, such as sports clips)
• Serialized narrative shows (≈40 minutes per episode, multiple episodes per series)

Randomized controlled trials (RCTs) have assessed promotional efficacy by randomizing exposure to the top-banner (D=1 for promotion, D=0 for standard ad), with approximately 10% of users allocated to treatment. The primary measured outcome is each user’s cumulative viewing time of the promoted series over the exposure window, often four weeks (Yasui et al., 16 Jan 2026).

2. Data Structure, Outcome Measures, and Viewing Patterns

User viewing time distributions in this context exhibit nonstandard features: a dominant point mass at zero (users who do not engage with the content at all), spikes corresponding to multiples of episode lengths (reflecting episode-completion drop-off), and generally heavy right tails for engaged viewers. Mean viewing time in the control group demonstrates wide variation by content type:

• 46 min comedy: 0.244 min
• 5 min sports: 0.0353 min
• Long-form reality A: 0.657 min
• Long-form reality B: 1.397 min

Let $Y_i$ denote the total viewing time (minutes) of the target series for user $i$. To further examine engagement quality, interval-based metrics such as the Probability Treatment Effect (PTE)—which estimates the shift in the probability that $Y_i$ falls within specified time intervals (e.g., one full episode)—are used to capture more granular, distributional effects than mean shifts alone (Yasui et al., 16 Jan 2026).

3. Statistical Frameworks for Impact Estimation

The evaluation of top-of-screen promotions leverages potential outcomes methodology, focusing on both marginal effect averages and full distributional changes due to the mixed discrete-continuous nature of viewing behavior.

Key statistical targets include:

• Marginal distributions under treatment/control: $F_1(y) = \Pr(Y(1) \leq y)$, $F_0(y) = \Pr(Y(0) \leq y)$
• Distributional Treatment Effect (DTE): $\Delta^{DTE}(y) = F_1(y) - F_0(y)$
• Probability Treatment Effect (PTE): $$\Delta^{PTE}(y, h) = [F_1(y + h) - F_1(y)] - [F_0(y + h) - F_0(y)]$$
• Optional: Quantile Treatment Effect (QTE): $QTE(\tau) = F_1^{-1}(\tau) - F_0^{-1}(\tau)$

Estimation proceeds via empirical CDFs, augmented by doubly robust regression adjustment. Gradient boosting with 3-fold cross-fitting is implemented for estimating conditional CDFs, and 500 bootstrap replications are applied to construct confidence bands. This framework uncovers heterogeneous effects and supports interval-based inference rather than relying solely on changes in sample means (Yasui et al., 16 Jan 2026).

4. Empirical Findings: Heterogeneity and Engagement Mechanics

Top-of-screen promotions on ABEMA drive statistically significant but modest lifts in mean viewing time for most content types, with pronounced heterogeneity based on content structure:

Case	Control Mean (min)	ATE (adj., min)	ATE/Control	Significance
1: 46 min comedy	0.244	0.0140	+5.7%	* (p<0.05)
2: 5 min sports	0.0353	0.0028	+8.0%	** (p<0.01)
3: Reality A	0.657	0.0590	+9.0%	** (p<0.01)
4: Reality B	1.397	0.0485	+3.5%	ns

Cases 1–3 demonstrate statistically significant lifts (6–9%), with the largest relative and sustained gains for short-form and serialized, strongly linked content. In contrast, longer-form narratives with weak initial hooks yield only trial viewing with no sustained engagement. Distributional and interval-based analysis confirms that top-of-screen promotion transitions users from non-viewing to trial viewing, but only strongly sequential or compact formats show increased probabilities of multi-episode or full-episode completion. Effects on long-form, self-contained content are largely restricted to initial sampling (Yasui et al., 16 Jan 2026).

5. Optimization of Top Banner Allocations under Constraints

Operationally, the delivery and selection of content or advertisements in the top-of-screen slot can be formulated as a continuous or discrete allocation problem. Frameworks from online advertising optimization applicable here (e.g., Caruso & Giuffrida (Caruso et al., 2010)) model the system over $K$ epochs and $N$ candidate creatives with forecasted supply $S_t$ and per-creative profit rates $\pi_i$. The static objective is

$$\max_{\{x_{i,t}\} \ge 0} \sum_{t=1}^K \sum_{i=1}^N \hat\pi_{i,t} x_{i,t}$$

subject to primary constraints:

• Supply cap: $\sum_i x_{i,t} \leq S_t$
• Individual campaign budgets: $$\sum_{t=1}^K \sum_{i:j(i)=j} \pi_i x_{i,t} \leq D_j$$
• Minimum delivery and no-overflow constraints for visibility or fairness

Rolling-horizon approaches re-optimize the allocation as returns, traffic, or campaign mix shift, converting $x_{i,t}$ (impressions) to real-time display probabilities for each candidate in the top-of-screen slot. In a dynamic environment, reward uncertainty mandates exploration: multi-armed bandit-based algorithms (UCB, Thompson sampling) guarantee sublinear regret $O(\sqrt{T\log T})$ relative to the best possible static strategy, and can be adapted to budget and visibility constraints (Caruso et al., 2010).

6. Managerial Insights and Implications

Evaluation on the ABEMA platform yields several operationally actionable principles:

• Top-of-screen promotions reliably increase trial viewing probabilities for all content types.
• Highest business value is derived when promotions induce repeated, sustained engagement, particularly with short-form and sequentially linked content.
• For long-form narrative, positive incremental engagement occurs only when early episodes act as strong engagement hooks; otherwise, only trial viewing is observed.
• Design of promotional strategy should prioritize short/strongly sequential content for maximal incremental effect and leverage in-content follow-on mechanisms for long-form serials.
• Simple average treatment effect estimates can obscure important heterogeneity, which is critical for targeting and sequencing campaigns (Yasui et al., 16 Jan 2026).

7. Implementation Guidelines and Adaptive System Design

Effective deployment of top-of-screen promotions integrates the following processes:

• Comprehensive event logging (impressions, clicks, registrations) and feature engineering for demand and reward modeling
• Short-horizon supply and engagement forecasting with periodic retraining
• Regular re-optimization of banner allocation via static or rolling-horizon linear programming, translating allocations into fractional display probabilities
• Mechanisms for learning (mandatory minimum delivery for new creatives) and adaptive exploration (multi-armed bandits for established choices)
• Continuous monitoring for model drift (e.g., abrupt changes in demand or conversion rates) and responsive re-solving of the allocation problem
• Real-time evaluation of cumulative performance and empirical regret versus clairvoyant optima (Caruso et al., 2010)

Combined, these elements operationalize top-of-screen promotions as a tightly integrated system that adapts to user behavior and supply fluctuations while approaching revenue or engagement-maximizing allocations under practical constraints.