Clickbait vs. Quality: How Engagement-Based Optimization Shapes the Content Landscape in Online Platforms

TL;DR

This paper analyzes how engagement-based optimization shapes the content landscape via a game where content creators balance costly quality and cheaper gaming tricks. It proves and empirically validates a positive correlation between gaming and quality at equilibrium and shows counterintuitive downstream effects: stronger gaming-cost barriers can reduce average content quality consumed and engagement-based optimization can underperform quality-focused or random baselines in user welfare. The authors provide comprehensive equilibrium characterizations for engagement, investment-based, and random baselines across homogeneous and multi-type settings, highlighting the importance of accounting for creator incentives when evaluating platform metrics. The work thus informs platform design and policy by linking metric transparency, incentive compatibility, and welfare outcomes in recommender ecosystems.

Abstract

Online content platforms commonly use engagement-based optimization when making recommendations. This encourages content creators to invest in quality, but also rewards gaming tricks such as clickbait. To understand the total impact on the content landscape, we study a game between content creators competing on the basis of engagement metrics and analyze the equilibrium decisions about investment in quality and gaming. First, we show the content created at equilibrium exhibits a positive correlation between quality and gaming, and we empirically validate this finding on a Twitter dataset. Using the equilibrium structure of the content landscape, we then examine the downstream performance of engagement-based optimization along several axes. Perhaps counterintuitively, the average quality of content consumed by users can decrease at equilibrium as gaming tricks become more costly for content creators to employ. Moreover, engagement-based optimization can perform worse in terms of user utility than a baseline with random recommendations, and engagement-based optimization is also suboptimal in terms of realized engagement relative to quality-based optimization. Altogether, our results highlight the need to consider content creator incentives when evaluating a platform's choice of optimization metric.

Figures (4)

• Figure 1: Support of a symmetric mixed equilibrium for engagement-based optimization in \ref{['example:linear']}. The parameter settings are $\gamma = 0.1$ (left), $\alpha = 1$, $\gamma = 0$, $\mathcal{T} = \left\{t_1, t_2\right\}$ (middle), and $\alpha = 1$, $\gamma = 0$, $\mathcal{T} = \mathcal{T}_{N, \epsilon}$ (right). The support exhibits positive correlation between gaming tricks $w_{\text{cheap}}$ and investment in quality $w_{\text{costly}}$ (Proposition \ref{['prop:positivecorrelation']} and Theorem \ref{['thm:positivecorrelationhomogeneous']}). For homogeneous users (left), the slope varies with the type $t$ and the intercept varies with the baseline utility $\alpha$ (Theorem \ref{['thm:onetype']}). For heterogeneous users with $N$ well-separated types (right), the support consists of $N'$ disjoint line segments with varying slopes and intercepts, where $N' < N$ in several cases (Theorem \ref{['thm:Ntypes']}).
• Figure 2: Cumulative distribution function $H_{a, f, \mathcal{G}}$ of the number of favorites ($w_{\text{costly}} = l$) conditioned on different angriness levels ($w_{\text{cheap}} = a$) on a dataset MCPWD23 of tweets from the engagement-based feeds ($f = E$) and chronological feeds ($f = C$). The tweet genre is unrestricted (left), restricted to political tweets (middle), and restricted to not political tweets (right). The cdf for higher values of $a$ appears to stochastically dominate the cdf for lower values of $a$, suggesting a positive correlation between $w_{\text{cheap}}$ and $w_{\text{costly}}$. The stochastic dominance is more pronounced for political tweets than for non-political tweets, and it occurs for engagement-based and chronological feeds.
• Figure 3: Equilibrium performance of engagement-based optimization (EBO) in \ref{['example:linear']} with $P = 2$ creators along several performance axes (left to right). The performance is numerically estimated from 100,000 samples from the equilibrium distributions (\ref{['sec:equilibriumcharacterizations']}). The parameter settings are $\mathcal{T} = \left\{1\right\}$ (left), $\mathcal{T} = \mathcal{T}_{N, \epsilon}$, $\alpha = 1$, and $\gamma = 0$ (middle), and $\mathcal{T} = \left\{5\right\}$ (right). The equilibrium performance of investment-based optimization (IBO) and random recommendations (RR) are analytically computed from the equilibrium distributions (Theorem \ref{['thm:investmentbased']} and Theorem \ref{['thm:randomrecommendations']}) and shown as baselines. User consumption of quality can decrease with gaming costs (left; Theorems \ref{['thm:comparisonuserconsumptiongaming']}-\ref{['thm:comparisonuserconsumptioninvestment']}), realized engagement can be lower for EBO than for IBO (middle; Theorem \ref{['thm:comparisonengagement']}), and user welfare can be lower for EBO than for RR (right; Theorem \ref{['thm:comparisonuserutility']}).
• Figure 4: The support of $(V,T)$ in Definition \ref{['def:2types']} for different values of $a_{t_1} / a_{t_2}$. The red line shows the support of $V \mid T = t_1$, and the blue line shows the support of $V \mid T = t_2$. If $a_{t_1}$ and $a_{t_2}$ are sufficiently far apart (Case 1), then the supports are disjoint. When $a_{t_1}$ and $a_{t_2}$ become closer (Case 2), the supports start to overlap, and when $a_{t_1}$ and $a_{t_2}$ are sufficiently close (Case 3), the support of $V \mid T = t_2$ is contained in the support of $V \mid T = t_1$.

Theorems & Definitions (104)

• Example 1
• Example 2
• Example 3
• Theorem 1
• Proposition 2
• Theorem 3
• proof : Proof sketch of Proposition \ref{['prop:positivecorrelation']}
• Example 4: continues=example:linear
• Theorem 4
• proof : Proof sketch of Theorem \ref{['thm:comparisonuserconsumptiongaming']}