AI Review Summary

The inclusion of the 2026 cross-sectional data point (the red star) in Figure 5 alongside the 2009–2018 EPFL fixed-cohort data visually and logically implies a continuous trend that the underlying data structures cannot support. The 2026 data was collected via a stopword-filtered cross-section of the entire active platform, whereas the EPFL data tracked fixed, legacy-verified cohorts longitudinally. Because the sampling frames, user populations, and platform mechanics (algorithmic vs. chronological feeds) are entirely disjoint between these two datasets, the 2026 retweet Gini cannot logically be used to validate or extend an "upward trend" from the EPFL data. I recommend removing the 2026 data points from the Figure 5 time series and revising the text to treat the historical and contemporary datasets as strictly separate analyses, avoiding causal or continuous trend language between them.

The timeline panel used to decompose variance into between-user and within-user components samples "up to 100 recent original tweets per author." As acknowledged in footnote 6, this means the temporal span covered by the data varies drastically between prolific users (a few days) and infrequent users (months or years). Consequently, the time period of observation is perfectly confounded with user posting frequency. Because engagement dynamics and platform algorithms fluctuate over time, the compressed time window for prolific users likely suppresses their within-user variance artificially, thereby inflating the overall between-user variance estimate (the reported 87% / 81%). To ensure the within-user variance is measured over a valid and comparable baseline, I recommend re-calculating the variance decomposition using a standardized time window for all sampled users (e.g., all tweets posted within a fixed 14- or 30-day period).

The primary bot filter (Criterion i) removes accounts with a mean impressions-to-followers ratio below 0.5%. Because the core regression analysis models the relationship between impressions (the dependent variable) and followers (the independent variable), filtering the dataset based on the direct ratio of these two exact variables introduces a mechanical truncation of the dependent variable's lower tail. While the authors present a sensitivity analysis omitting this filter—which notably drops the quadratic coefficient from 0.053 to 0.028—the primary models and variance decompositions still rely on the truncated data. To avoid hardcoding a performance floor into the dependent variable and artificially inflating the superstar effect, I recommend using the non-performance-based filter (e.g., automation and ghost accounts only) as the primary specification for all main text results, moving the aggressive performance-based filter to the sensitivity analysis.

The study claims to measure attention inequality across English-language posts, but the primary cross-sectional sampling relies on a disjunction of ten specific English stopwords. This method structurally excludes non-text media (images, videos without text) and short-form posts lacking these specific words. Because highly viral content on modern social media is frequently visual, this sampling strategy is perfectly confounded with content format. The findings logically support conclusions about *text-heavy* English posts, but the strength of the claim exceeds the evidence when generalized to the platform's overall attention economy. I recommend revising the manuscript's claims throughout the abstract and discussion to explicitly scope the conclusions to long-form or text-heavy posts, rather than treating the stopword query as a representative sample of all English-language platform activity.

**Residual Overdispersion in the Negative Binomial Model:** In Appendix A, the authors use a Negative Binomial GLM to check robustness against the log transformation. The text states that the model accommodates overdispersion and notes a Pearson dispersion statistic of approximately 8. However, a Pearson statistic of 8 for a *fitted* Negative Binomial model indicates severe residual overdispersion that the model's variance function has not fully captured (a well-fitting model should yield a value near 1). The authors should clarify whether the standard errors reported in Table 8 are robust/sandwich standard errors that account for this residual overdispersion. If they are not, the standard errors may be substantially underestimated, and robust standard errors should be applied to ensure the significance of the findings holds.

**Post-Treatment Variables in the "Full" Model:** In Table 5 and the associated marginal effects analysis (Table 7, Figure 4), the authors include engagement metrics (likes, retweets, quotes) as covariates to model impressions. Because engagement is a direct downstream consequence of impressions (as the authors acknowledge via the "accounting identity"), these variables act as post-treatment covariates. While the authors appropriately disclaim causal inference, interpreting the conditional association of followers *controlling for* engagement (e.g., the interaction terms and marginal effects in Table 7) is statistically problematic, as it conditions on a downstream outcome and risks collider bias. The authors should explicitly discuss the statistical limitations of conditioning on post-treatment variables when interpreting the marginal associations of follower counts in the full model.

**Reproducibility of IPW Estimates:** The authors report an IPW-corrected between-user variance share of 81%, noting that the weights are based on the ratio of cross-section strata proportions to panel strata proportions across "six log-spaced strata." However, the exact definitions of these strata (e.g., the specific follower count bin edges) and the empirical proportions used to calculate the weights are not reported in the manuscript. To ensure full reproducibility of this key population-corrected estimate, please provide the strata definitions and the corresponding weight calculations, either in the main text or a supplementary table.

**Varying Time Windows in Variance Decomposition:** The within-user variance decomposition relies on up to 100 recent tweets per user, which spans a median of 14 days but varies widely across users (from a few days to several months, as noted in footnote 6). Because engagement variance may be non-stationary (e.g., subject to temporal autocorrelation or topic shifts over time), calculating within-user variance over vastly different time scales for different users may distort the relative within-user versus between-user variance shares. The authors should consider reporting a robustness check for the variance decomposition that restricts the analysis to a standardized time window (e.g., tweets from the past 14 or 30 days) for all users to ensure the within-user variance is strictly comparable across the sample.

The study applies a rigorous composite filter to remove bot *authors* from the dataset, but it does not address the potential impact of non-human (bot or scraper) *views* on the impression counts themselves. Because impressions (screen renders) are the primary dependent variable, bot-driven inflation of views—which may disproportionately target high-profile accounts or trending topics—could skew the inequality metrics. The authors should discuss this limitation and how unmeasured bot consumption might influence the interpretation of the extreme Gini coefficient.

The temporal analysis (Fact 3) compares a 2009–2018 historical archive to a 2026 cross-sectional snapshot. The 8-year gap between 2018 and 2026 omits critical periods of platform evolution, making the assertion of a continuous or secular trend highly speculative. The authors should explicitly address the limitations of inferring trends across this missing window, particularly given the major structural and algorithmic changes the platform underwent during that specific timeframe.

The stopword query used to approximate a random sample of the English-language tweet stream ("the OR just OR that OR this OR was OR with OR but OR what OR been OR they") relies heavily on standard English syntax. This approach may inadvertently filter out highly informal speech, short meme-based text, or specific dialects (such as African American Vernacular English) that might omit these specific stopwords in short-form posts. The authors should discuss how this linguistic sampling strategy might bias the representation of certain communities and impact the generalizability of the inequality estimates.

In Section 1, the manuscript compares the 2026 impression Gini (0.965) to YouTube views from 2010 (Cha et al.) and scientific citations from 1976 (de Solla Price). Comparing contemporary, algorithmically-driven feed data to decades-old, structurally different platforms limits the utility of this benchmark. The authors should contextualize this comparison by acknowledging the temporal and structural differences, or ideally, include more contemporary benchmarks for digital attention inequality to better situate the study within the current literature.

The authors rightly note that the observed convex association between followers and impressions could be driven by a mixture of follower-driven ("Following" feed) and algorithm-driven ("For You" feed) regimes, but state that distinguishing these requires unavailable API data. To strengthen the discussion of future directions, the authors could propose specific methodological next steps—such as leveraging user-side data donation, browser extensions, or controlled sock-puppet auditing—that could empirically disentangle these two distribution channels in future research.