Industrial recommenders give new content initial views through budgeted exploration, then use early performance to decide further delivery. On many short-video platforms, exploration is the primary way new videos reach viewers. Viewer-side tests measure consumption; the published budget objectives we review omit creator response. We analyze four experiments on a major short-video platform. An eight-month creator ablation finds production exploration raises videos posted per creator by 8.55% and creators posting at least once by 7.10% relative to a minimal floor. A budget-matched reallocation raises creator participation with no detectable short-run viewer-side change. A year-long viewer ablation finds 1.74% more video views but 2.13% less view time. A delivered view creates immediate feed value, can trigger organic take-up, and can induce creator supply. Take-up and supply replenish a shared corpus, creating two measurement limits. Viewer-side A/B tests cancel the corpus effect when both arms consume the same corpus. Giving each arm its own corpus avoids cancellation, but turnover still controls the horizon. If the corpus turns over at rate w per posting cycle, a t-cycle experiment expresses at most wt of the eventual corpus effect. More users reduce noise but do not speed turnover. Before the corpus path visibly bends, data cannot distinguish a modest fast effect from an arbitrarily large slow one, so a valid confidence interval may lack a finite upper endpoint. As predicted, the three-week co-diverted experiment cannot determine the sign of the eventual corpus effect. Within the window, it identifies the direct feed effect, and an exploratory cohort analysis detects organic lift after exploration ends. The experiments establish a positive creator response, measure the gross corpus flow visible within three weeks, and show the design and duration needed to identify total value.
Figures & tables
Figure 1. Value channels and identifying designs. Exploration has an immediate feed effect, can change creator supply, and feeds a shared corpus through both take-up and new posts. The shaded channels express slowly. Each dashed design reaches a different channel. A one-sided design shares the corpus across arms, so that channel cancels. A separate budget-matched co-diverted variant perturbs allocation at fixed nominal budget and measures the creator response (§ 5.4 ).
Target
Design
Required horizon
Rate in N
Direct effect
Viewer ablation
Immediate
N−1/2
Supply response
Creator ablation
Supply loop τ(R)
N−1/2
Allocation response
Budget reallocation
Supply loop τ(R)
N−1/2
Corpus ramp
Co-diversion
Early ramp
N−1/3
Corpus asymptote
Co-diversion
Curvature threshold
N−1/5
Table 1. Identification map for the four randomized designs. The final column gives the sample-size law. For the fast channels it sharpens the estimate at a fixed horizon. For the corpus rows it sets the required horizon.
Metric
Rel. change (%)
95% CI
Panel A: Viewer ablation (§ 5.2 )
Video views
+1.74
[+1.57,+1.90]
View time
−2.13
[−2.28,−1.98]
Views under 1.5s
+7.49
[+7.28,+7.70]
Video favorites
−2.32
[−2.72,−1.92]
Panel B: Creator ablation (§ 5.3 )
Table 2. Randomized effects at the tested operating points. Relative arm differences have 95% CIs. Positive values favor production in Panels A–B and reallocation in Panel C. The panels isolate a mixed direct trade-off, a signed supply response, and a budget-matched participation gain.
Figure 2. Observed time paths and turnover sensitivity. (a) The year-long viewer contrast appears within weeks and stays level until a common operating-point shift, matching the direct-step prediction. Intervals are approximate. (b) The eight-month posting response, indexed to its first measured period, is descriptive and follows onset with a plateau. (c) The corpus-asymptote conclusion depends on the assumed turnover. The 36-day case rules out offsetting either direct view-time cost (the ablation’s −2.13% , drawn dashed, and the probe’s own −1.02% step), whereas the 360-day case admits either sign. Panels (a)–(b) use finer analysis periods than the appendix’s series. Their shapes agree (Appendix C ).
Quantity
Estimate
Panel A: Direct step
Video views
+0.43[+0.30,+0.57]
View time
−1.02[−1.14,−0.89]
Video favorites / delivered view
−0.87[−1.21,−0.53]
Panel B: Corpus asymptote by turnover timescale
Video views, 36-day timescale
[−0.28,+0.70]
Table 3. Co-diverted three-week readout at the calibrated estimation floor. Entries are relative changes with 95% CIs. Panel B gives intervals for the set-identified corpus asymptote under two turnover timescales. Panel C gives feasible-horizon gross flow; its implied whole-cell contribution is a derived range. Posting is quoted as estimate ± margin. Panels C–D retain the inferential qualifications stated in text.
Direct step mixed; exploratory gross cohort flow +
Net corpus asymptote at three weeks
Table 4. What the four experiments establish. Signs apply at the realized horizon. Co-diverted rows condition on the realized submarkets.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3. Finite-upper-endpoint probability for a 95% profile procedure in semi-synthetic Monte Carlo at the measured noise scale. Blue is the exact single-rate family of Theorem 1 . Its C≤1 region obeys the theorem’s 21+α cap. Orange uses the measured age–yield kernel under free timescale dilation. It shares the horizontal coordinate through the kernel’s 36-day single-rate equivalent. At the probe’s 21-day window ( C≈2 ), its finite-endpoint probability is 0.21 . Markers give each curve’s median crossing.
Figure 4. The same seven cumulative estimates differenced into six non-overlapping periods (solid, with approximate bands), against the cumulative reading (faint). The dashed vertical marker is the operating-point shift both experiments turn at (§ 5.2 ). Positive = higher under production. Consecutive periods share no data, so their shape is the effect’s shape. The cumulative series is smooth by construction. Period bands are approximate upper bounds derived from the cumulative margins under independent per-period noise, scaling the cumulative margin by t2/(t2−t1) . No closed-form margin for nested increments is available from the per-horizon estimates.
Figure 5. The creator ablation’s consumption series read per period (solid, approximate bands) against its cumulative reading (faint), as in Figure 4 . The dashed marker is the same operating-point boundary both experiments turn at. Period effects on video views are positive up to it and negative after, the coincident break cited in § 5.2 .
Figure 6. In-cell organic video views and view time for videos posted during the experiment and viewed after their exploration windows, production versus floored cell, daily with 95% intervals (user-level sample). The pooled full-population view-time contrast is +37.9%[37.1,38.7] . The gray band is the placebo interval from videos posted before the experiment.
Figure 7. Theorem 2 ’s delay family, from ( 3 ) (§ A ) at ω=0.005 , Roff=0.9 . Four values of the interpolating coefficient, c/cmax∈{0,31,32,1} , light to dark. (a) The onset differs. At c=0 the corpus responds at rate ω from the first period. At c=cmax the first period expresses nothing, and the ramp extrapolates back to zero at exactly A(Roff) . (b) The same four curves over the corpus timescale, where the distinction has washed out.
Figure 8. Corollary 3 : the multiplier a feedback loop delivers and its relaxation time coincide to within one period. Both are plotted against a single axis. A target expression horizon adds the factor set by that target. In particular, t90%=ln(10)τ(λ) .
Figure 9. Proposition 4 and Corollary 2 : required horizon against traffic per arm, both relative, for the three exponents the paper derives. Markers give the traffic multiple that halves the horizon.
Figure 10. Proposition 1 ’s three channels as shares of the marginal value of a view, against loop gain R . Illustrative units ( ηe=ηo=1 , Vorg=10 , qˉ=0.1 , Bˉ=100 , q′=0.01 , δ=0.01 ). The shares depend on these. The monotonicity is structural. Stack order is bottom to top as listed.
On two-sided content platforms, symmetric two-sided isolation (assigning matched fractions of creators and viewers to isolated treatment and control submarkets) is widely used for creator-side and cold-start experiments because it removes cross-arm marketplace interference. Isolation, however, thins each viewer's candidate catalog, and intuition suggests the resulting engagement cost should fade as the platform grows: a small fraction of a vast catalog is still vast. We show that, in an order-statistics model of engagement, whether this intuition holds depends on the upper tail of match quality. Extreme-value theory yields tail-class loss laws with a sharp dichotomy: for light or bounded tails the loss vanishes as the candidate pool grows, whereas under heavy tails it converges to a size-independent constant, so expanding the candidate pool, even by orders of magnitude, does not asymptotically eliminate the cost. Evidence from two production experiments on a platform with millions of active creators is consistent with this picture: a pure A/A traffic sweep reveals a measurable, depth-graded engagement cost; a one-sided catalog ablation independently shows that per-viewer thinning contributes to the loss; and a tail index calibrated on the small exploration pool predicts an effect consistent with the one observed in the far larger full-catalog ablation. Isolation thus carries a price that experimenters should budget for, like any other cost. We give practitioners a preflight procedure that estimates it before launch, sizes traffic accordingly, and recommends a fallback design when the predicted cost exceeds a chosen tolerance.
Yuanyuan Shen, Yiren Yan, Wenjie Li +1
Snap Inc. New York, NY, USA · Snap Inc. Palo Alto, CA, USA
The rise of AI amplifies the economic phenomenon of positive spillovers: when creators contribute content that can be reused and adapted by LLMs, one creator's effort may improve the content quality of others through recombination. While such spillovers can improve content quality, they also reshape incentives, as creators may reduce effort when they can benefit from others' contributions. We introduce the Content Creation with Spillovers (CCS) model, a game between a platform and strategic creators. In this game, each creator chooses an effort level, and qualities are jointly determined by all creators' efforts. The platform contracts on qualities while seeking to maximize social welfare. We show that standard contest mechanisms can be unstable, and propose a parametrized family of Provisional Allocation mechanisms that guarantee a Pareto-dominant equilibrium. Although optimizing welfare within this family is hard to approximate in general, we develop approximation algorithms whose guarantees hold either deterministically across structured spillover classes or with high probability under random instances.
This paper introduces a real-time thumbnail optimization system deployed at a global O(B) scale on a major short-form video platform. Unlike traditional long-form content, where custom thumbnails are heavily curated by creators, a considerable fraction of short-form videos are published without human-selected artwork. To address this uncurated corpus, we present a fully automated, end-to-end framework that replaces static default frames with dynamic, data-driven selections across billions of videos. To the best of our knowledge, this is the first published work demonstrating an online Multi-Armed Bandit framework successfully deployed at an O(B) scale for uncurated short-form video discovery. Our solution pairs a multi-stage candidate generation pipeline with a low-latency serving infrastructure. By initializing the exploration framework with image-specific priors derived from a deep visual quality model, the system minimizes exploration costs and dynamically serves optimal thumbnails at serving time. Global deployment demonstrates statistically significant improvements in core user discovery and engagement metrics.