Emergent structures of attention on social media are driven by amplification and triad transitivity

Smith, A. H., Green, J., Welles, B. F., & Lazer, D. (2025). Emergent structures of attention on social media are driven by amplification and triad transitivity. PNAS Nexus, 4, gaf106. https://doi.org/10.1093/pnasnexus/pgaf106

View paper

Summary

This paper introduces the concept of the attention broker (the tertius amplificans, or “third who amplifies”) as a scaled extension of Obstfeld’s tertius iungens orientation to social media platforms with many-to-many amplification affordances. The authors argue that the well-documented tendency toward triad transitivity in directed social networks is partly produced by a local, endogenous causal mechanism: when a high-degree account amplifies another user’s content with attribution (e.g. a retweet), its followers form new following ties to the amplified account. Using a novel data-collection technique and a difference-in-differences design across two contrasting case studies—Jorts the Cat (pro-union) and J.K. Rowling (TERF advocacy)—they causally identify how amplification accelerates transitive triad closure beyond background virality. The mechanism works through exposure rather than persuasion, hastening ties among followers already predisposed to follow the amplified accounts.

Key Contributions

• Theoretical: coins “attention brokerage” / tertius amplificans, a scaled, amplification-based extension of the tertius iungens applicable wherever attributed resharing occurs (retweets, citations, TikTok duets, corporate memos).
• Methodological: documents and operationalizes a Twitter/X V1 API cursor technique (exploiting a modified Unix nanosecond timestamp) to recover time-bounded follower events, enabling precise temporal analysis of tie formation.
• Empirical: provides causal (difference-in-differences) evidence that amplification by influential accounts generates transitive triads, across two structurally and ideologically divergent cases.
• Conceptual bridge: links local micro-mechanisms to macro-level network properties, offering a causal explanation for the prevalence of transitivity.
• Open science: releases an anonymized dataset (SOMAR/ICPSR) and code documenting the cursor-based collection method.

Methods

A two-case comparative design contrasts Jorts the Cat (~200K followers, Dec 2021–Mar 2022) with J.K. Rowling (~14M followers, Jun 2018–Mar 2023). The authors collected brokers’ full timelines via the focalevents package, filtered to simple retweets (excluding quote tweets to avoid “dunking”), and hand-labeled 646 (Jorts) and 534 (Rowling) retweeted accounts along cause-alignment and interest-actor dimensions with four coders and a tiebreaker. Treatment motifs (transitive triads: follower–broker–retweeted) and control motifs (open triads: nonfollower–broker–retweeted) were constructed over 2-week pre/post windows around each retweet. Attentive population sizes were estimated using the POPAN Jolly-Seber mark-recapture model in Project MARK. Causal estimates came from a two-stage difference-in-differences event study (Gardner) with account and time fixed effects, plus Rambachan–Roth sensitivity analysis for parallel-trends violations.

Findings

• For both brokers, the day-0 (retweet day) treatment effect is positive and significant: followers follow the amplified account at much higher rates than nonfollowers.
• Effects are heterogeneous by account type: Jorts’s brokerage is strongest for union-related accounts; Rowling’s effect is significant across all types but largest for TERF interest-actor accounts.
• Smaller positive pre-retweet effects suggest incidental prior exposure also contributes; a post-spike decline in following rates suggests amplification accelerates ties that would eventually have formed, depleting the pool of latent followers.
• Rambachan–Roth sensitivity analyses show robustness: parallel-trends violations would have to be more than four times larger post-retweet than pre-retweet to overturn key results.
• Estimated attentive populations differ substantially (e.g. ~164K Jorts followers vs. ~17.9M nonfollowers; ~841K Rowling followers vs. ~2.68M nonfollowers in the sampled set).

Connections

This paper’s focus on how amplification by influential accounts reshapes the distribution of attention connects to work on influencer-driven amplification and structural roles in networks such as Rothut2026-or and Bakshy2015-rn, the latter’s exposure-based framing of what circulates on platforms being a natural counterpart to the exposure-not-persuasion mechanism argued here. Its use of case-based, mechanism-oriented network analysis also resonates with computational structural studies like Gaisbauer2025-by and Green2025-ap, where local processes are linked to emergent macro-level patterns.