Attention Inequality on X/Twitter: Evidence from English-Language Posts

Abstract

Every day, hundreds of millions of posts compete for a finite resource: human attention. We present a descriptive analysis of how this resource is distributed among English-language posts on X (formerly Twitter), drawing on a cross-sectional sample of 8,722 tweets (February 2026, after bot filtering), a timeline panel of 17,671 tweets from 225 users, and a historical archive of 95 million tweets (2009–2018). Four main findings and one cautionary observation emerge: (1) attention inequality is extreme among text-containing posts—the impression Gini is 0.965, and after inverse-probability weighting (IPW) to correct for the stratified over-sampling of high-follower accounts, approximately 81% of total variance is between users rather than between posts (unweighted upper bound: 88%, bootstrap 95% CI: [84%, 89%]); (2) the association between followers and impressions shows increasing returns in the 100–100,000 follower range, with possible saturation above 100K— a cross-section analysis (n = 7,673) yields ˆ β2 = 0.053 (p < 0.001, R2 = 0.761) for a global quadratic approximation; this pattern persists under all filter specifications, including automation-only filtering ( ˆ β2 = 0.030, p < 0.001); (3) retweet inequality rose within the EPFL archive (verified-account Gini: 0.76 in 2012, 0.89 in 2018), a pattern consistent with both platform evolution and cohort maturation; (4) likes, quotes, and impressions are tightly coupled in equilibrium, with the coupling weakening for larger accounts—attributable to mechanical necessity (one must see a tweet to like it) and selection effects, with algorithmic amplification as one possible additional channel. We also note that follower inequality decreased within the fixed EPFL cohort, though this is likely a survivorship artifact. Pooling with the stratified timeline panel raises R2 to 0.821. These findings describe a platform where attention is concentrated far beyond what follower counts alone would predict—consistent with algorithmic mediation, though our evidence is descriptive rather than causal.

Comments