"Rich-Get-Richer"? Platform Attention and Earnings Inequality using Patreon Earnings Data

In the paper

“It may also lower barriers to entry for content creators, but it will not change online inequality and concentration in attention and earnings unless the underlying algorithmic mechanisms changes - even if AI-generated content has the potential to foster more egalitarian ecosystems (Zhou and Lee 2024).”

What the sources say

• Disputed

  Split into 2 atoms — 0 confirmed, 1 partially supported, 1 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Checked by a reasoning model — may be wrong.

  cites Zhou and Lee 2024 sourceExperimental — accuracy not yet measured

In the paper

“AI technologies have for some time underpinned recommendation algorithms and advertising algorithms (Covington et al. 2016; Schafer 2023; Gairola 2024).”

What the sources say

• Partly supported

  The Introduction says deep learning has had a substantial impact on YouTube recommendations, supporting the recommendation-algorithm part of the claim. The Feature Engineering section compares its ranking signals with experience in ranking ads, but does not establish that AI underpins advertising algorithms. (LLM, conf=0.86)

  cites Covington et al. 2016 sourceExperimental — accuracy not yet measured
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Schafer 2023
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Gairola 2024

In the paper

“A central challenge facing recommender systems on mature platforms is their potential to compound initial advantage (viewership), by promoting popular, existing material and focusing less on untested 'fresh' content (Covington et al. 2016; Sutton and Barto 2018).”

What the sources say

• Supported

  The “Example Age” Feature section says models trained on historical behavior can be biased toward the past, making recommendations reflect earlier watch likelihood rather than current popularity. Section 3.4 adds that training only on watches generated by recommendations would make it difficult for new content to surface and would bias the system toward exploitation. Together, these passages support the claimed potential to favor established videos over fresh content. (LLM, conf=0.86)

  cites Covington et al. 2016 source
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Sutton and Barto 2018

In the paper

“Second, platforms with stronger power laws (lower α ), such as YouTube and Instagram, exhibit weaker creator middle classes: lower mean and median earnings with less expansive interquartile ranges (Q25-Q75).”

In the paper

“Platform variation is consistent with differences in algorithmic systems: YouTube ( α = 1 . 8), Instagram ( α = 1 . 84), Twitch ( α = 1 . 93), and Facebook ( α = 1 . 94).”

In the paper

“YouTube, Instagram, Twitch, and Facebook consistently show well-defined power law distributions, ranging between 1.8 and 1.94 for the year 2024”

In the paper

“Fourth, if conversion culture alone drove the observed patterns, we might expect platforms with larger creator bases to show more diverse earnings distributions through sheer variety; yet YouTube-affiliated creators show the most concentrated earnings despite YouTube's dominance as a traffic source, suggesting that attention allocation mechanisms are a plausible driver alongside conversion norms.”

In the paper

“Third, comparing three cross-sections (2018, 2021, 2024), we observe convergence toward steeper inequality across YouTube, Twitch, Instagram, and Facebook, with declining α values implying increasing earnings concentration.”

In the paper

“Facebook is the only platform to have fewer exclusive Patreon content creators in our data in 2024 than it had in 2021 (Table 2 in Appendix).”

In the paper

“The standard deviation of earnings is highest for Twitter/X and lowest for Facebook.”

In the paper

“Patreon, however, allows creators to keep their earnings private, leading to significant missing data here, especially for larger earners, since those with more paid subscribers tend to disclose their earnings the least, we find.”

In the paper

“The data show substantial variation in earnings across platforms. Twitter/X and Patreon have the highest mean, median, maximum, and interquartile range of earnings in our sample, while Facebook and Twitch sit at the lower end.”

In the paper

“Importantly, platforms with stronger recommendation systems, such as YouTube ( α = 1 . 8) and Instagram ( α = 1 . 84), exhibit more significant earnings inequality and a weaker 'middle-class' of creators (with lower mean, median, and interquartile earnings)”

In the paper

“YouTube has the lowest value in 2024, followed closely by Instagram, with values around 1.8 and 1.88.”

In the paper

“Instagram's α dropped from around 2.1 in 2021 to around 1.84 in 2024, a period that coincided with the platform's shift toward algorithmic recommendation.”

In the paper

“The Music, comics, and writing categories all show relatively high α”

In the paper

“and the cross-validated errors are very similar, indicating limited overfitting.”

In the paper

“a few platform-year estimates, especially Patreon and Instagram in the final year, are somewhat sensitive to the estimation method used (see Appendix A).”

In the paper

“It is notable that YouTube offers one of the most generous shares of ad sales, and hosts the greatest earnings inequality among Patreon creators who monetize there (Financial Times 2023b).”

In the paper

“Temporal trends show worsening (and converging) power law earnings dynamics of Patreon creators on Instagram, YouTube, Facebook, and Twitch, where decreasing α values suggest growing dominance of top Patreon creators.”

In the paper

“With the exception of the earnings distribution for Patreon creators, the alpha values estimated by the three methods are generally consistent, confirming the robustness of the power-law pattern in the earnings distribution.”

• Also stated in: § A Robustness and Further Analysis · p. 26 · boosted

  “In nearly all other cases, the exponents estimated by the three methods are in strong agreement.”

In the paper

“As documented in our data, platforms vary substantially in their treatment of adult content as well, leading to different creator compositions and potentially different earnings dynamics.”

In the paper

“Instagram and Facebook also show a relatively high and similar proportion (0.40-0.43).”

In the paper

“Similarly, for the last year of the Instagram data, the CNS estimator deviates from the other two methods, again reflected in the faster decay of the tail relative to the predicted line.”

In the paper

“Twitter/X - and to some extent Patreon - stands out in this context, combining a high median income with a relatively equal earnings distribution up until a certain point.”

In the paper

“By contrast, if conversion is systematically rank-dependent , meaning, if larger creators convert platform attention into patrons at higher rates, because they cultivate stronger parasocial relationships, have better funnel design, or face different audience composition), then C ip may increase with A ip and can steepen or flatten the observed tail.”

In the paper

“Estimations were performed for each platform and year, and results remained robust across resampled datasets.”

In the paper

“A smaller α (closer to 2) indicates algorithms heavily favor already popular creators (lower γ ), leading to steeply skewed attention distributions.”

In the paper

“Supporting this interpretation is our finding that creators with no other social media accounts (Patreon-only) show the second least concentrated distribution (highest α ), after Twitter/X - whose historical reliance on social graphs rather than pure recommendation may explain its relatively equitable outcomes.”

In the paper

“Platform-specific power law exponents ( α ) exist and are consistent with differences in earnings inequality across platform-affiliated creators”

In the paper

“Patreon earnings data provide concrete, measurable outcomes of these dynamics given their reliance on third-party platform algorithms for visibility.”

In the paper

“An important caveat is that γ should be understood as a reduced-form parameter capturing the net effective exploration rate, not a direct measurement of any specific platform mechanism.”

In the paper

“Despite previous studies employing Patreon data (Regner 2021; El Sanyoura and A. Anderson 2022b; Hutson 2021), we provide the first systematic analysis of the earnings distribution's underlying structure and the economic processes or algorithmic mechanisms generating it.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Regner 2021
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites El Sanyoura and A. Anderson 2022b
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Hutson 2021

In the paper

“A lower power law exponent implies higher inequality, indicating greater probability of extreme tail incomes characteristic of wealth distributions (Pareto 1896; Atkinson 1970).”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Pareto 1896
• ?Not checked — no full text of the sourcecites Atkinson 1970 source

In the paper

“As creators gain visibility, they attract further attention through algorithmic amplification (Zappin et al. 2022; Hua et al. 2022), creating a self-reinforcing feedback loop that concentrates income among a small group of highly popular creators - a 'winner takes all' or viral effect (Rosen 1981; Rieder et al. 2018).”

What the sources say

• Not checked yet

  Checked before — its earlier verdict is held back until a new check runs.

  cites Zappin et al. 2022 source
• Not checked yet

  Checked before — its earlier verdict is held back until a new check runs.

  cites Hua et al. 2022 source
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Rosen 1981

In the paper

“Within this mixed model, reducing γ (relying more on popularity-driven exploitation) lowers α , intensifying inequality and concentrating attention among top creators, while increasing γ raises α and produces a more balanced distribution of attention.”

In the paper

“Our core finding is that monthly Patreon earnings across platforms fit a power law with exponent around α ≈ 2 . 0, indicating an extremely heavy-tailed distribution and a strong 'rich-get-richer' dynamic.”

In the paper

“Our data directly identify the distribution of Y ip , but do not separately identify the distributions of A ip , C ip , and R ip .”

In the paper

“We emphasize preferential attachment because it captures the global algorithmic mechanism: recommendation systems explicitly allocate attention in proportion to existing popularity.”

In the paper

“A further caveat concerns the conversion rate from platform attention to Patreon subscriptions, which may vary across platforms for reasons unrelated to algorithmic attention allocation.”

In the paper

“A larger α corresponds to more equitable attention distribution, arising from higher γ algorithms that are more exploratory and less biased toward existing popularity.”

In the paper

“Our earnings data most closely proxies for algorithmic attention mechanisms on third-party platforms where these creators are active, since Patreon monetization relies almost entirely on external platforms' algorithms to generate attention, as Patreon offers minimal native discovery capabilities.”

In the paper

“reflecting lighter tails and less earnings inequality.”

In the paper

“Our data provides a partial view of total creator earnings inequality and so should not be interpreted as representative of total earnings.”

In the paper

“In contrast, platforms with weaker power laws-Patreon and especially Twitter-support healthier middle-class creator ecosystems, which may reflect algorithmic differences in content discovery, though audience composition and platform-specific monetization norms could also contribute.”

In the paper

“In the original Barabási-Albert (BA) model there is no random component: new links attach purely in proportion to existing degree, which generates a power-law degree distribution with density exponent α = 3.”

In the paper

“Platforms that recommend the same 'viral' content to all users may generate highly unequal earnings, whereas algorithms that surface niche 'long tail' content may foster more equitable distributions (C. Anderson 2012; Brennan 2020).”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites C. Anderson 2012
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Brennan 2020

In the paper

“Patreon is perhaps the only large creator platform to disclose earnings, but this has not been analyzed systematically to test for cross-platform attention and earnings inequality (Regner 2021; Hutson 2021; El Sanyoura and A. Anderson 2022b,a).”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Regner 2021
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Hutson 2021
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites El Sanyoura and A. Anderson 2022b,a

In the paper

“suggesting relatively heavier tails and higher inequality, indicating a greater tendency for the emergence of super-hubs-nodes with degrees larger than expected in the standard preferential attachment model (Barabási and Albert 1999).”

What the sources say

• Not checked yetcites Barabási and Albert 1999 source

In the paper

“These patterns are consistent with more or less concentrated algorithmic attention allocations by each platform, but they could also reflect platform-specific conversion regimes or creator sorting into Patreon as a monetization channel.”

In the paper

“Algorithms, however, don't just create new markets and enable efficiencies (Varian 2010), they are distributional devices that decide who benefits and who does not (Hitzig et al. 2019).”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Hitzig et al. 2019

In the paper

“Applied to platform earnings, the same attention-reinforcing mechanisms shape creators' revenue streams. If visibility and engagement are governed by these dynamics, then earnings - including earnings derived from ads, sponsorships, or subscriptions - also become concentrated among a few popular creators.”

In the paper

“We show in Appendix C that the method is robust and has good predictive power, despite its simplicity.”

In the paper

“The temporal trends we document are harder to attribute to compositional change alone, since they show within-platform shifts over three-year intervals, but we acknowledge that evolving creator and audience composition could contribute to these trends as well.”

In the paper

“Creators also self-select into platforms based on content type, audience fit, and monetization goals.”

In the paper

“Facebook has recently chosen to do the opposite (TechCrunch 2025).”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites TechCrunch 2025

In the paper

“They document novel empirical patterns using a creative data source, and are consistent with the hypothesis that algorithmic attention allocation drives earnings concentration - but they do not establish this causally.”

In the paper

“It is notable that YouTube offers one of the most generous shares of ad sales, and hosts the greatest earnings inequality among Patreon creators who monetize there (Financial Times 2023b).”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Financial Times 2023b

In the paper

“we treat claims about algorithmic attention allocation as suggestive rather than definitive in the absence of direct attention measures or quasi-experimental variation in platform policies.”

In the paper

“Competition from TikTok appears to have locked the market into highly addictive, engagement-maximizing algorithms - a race to the bottom.”

In the paper

“If the most commercially ambitious creators concentrate on YouTube and Instagram while more niche creators gravitate toward Twitter/X or remain Patreon-only, the resulting earnings distributions would differ even under identical algorithms.”

In the paper

“reflecting a growing permissiveness on the platform (Table 2 in Appendix).”

In the paper

“In contrast, platforms with lower α show a 'rich-get-richer' pattern far earlier on, where a small elite captures most of the income, leaving the middle-class of creators under-rewarded.”

In the paper

“The observed earnings distributions may thus partly reflect compositional differences in who uses each platform rather than (or in addition to) algorithmic effects.”

In the paper

“It is important to restate that our findings offer a partial and suggestive view of creator earnings dynamics.”

In the paper

“Ad revenue sharing, tipping, subscriptions, and brand deals vary across platforms.”

In the paper

“the sample consists of only three cross-sections (2018, 2021, 2024)”

In the paper

“In this mixed model, choosing γ = 1 / 2 yields the same exponent α = 3, although the underlying mechanism differs from the original BA construction. See Appendix B.2 for the standard BA derivation.”

In the paper

“Formally, if individual earnings X follow a Pareto-type distribution with density p ( x ) ∝ x -α for x ≥ x min and α > 1, then the upper tail behaves as P ( X ≥ x ) ∝ x -( α -1) .”

In the paper

“But maximizing for immediate user engagement (Immorlica et al. 2024), as a quarterly economic imperative facing the platform's shareholders, can lead to less new content being created and even less longer-term user engagement - even if potentially at the expense of diminished immediate user satisfaction (Meta 2021; Yao et al. 2024).”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Immorlica et al. 2024
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Meta 2021
• Not checked yet

  Checked before — its earlier verdict is held back until a new check runs.

  cites Yao et al. 2024 source

In the paper

“More competition alone is, therefore, unlikely to correct these unequalizing algorithms, especially if they dampen rational consumer choice and the usual benefits of competition on the user side.”

In the paper

“The low marginal cost to distributing and selling content online adds to the power law, which arises from the feedback introduced by correlated decisions across a population (Easley, J. Kleinberg, et al. 2010).”

What the sources say

• ?Not checked — no full text of the sourcecites Easley, J. Kleinberg, et al. 2010 source

In the paper

“These trends are consistent with algorithmic systems increasingly favoring viral content over diverse creator engagement, though compositional shifts in the creator population could also contribute.”

In the paper

“These mechanisms are not mutually exclusive with algorithmic explanations; indeed, algorithmic design may interact with each of them.”

In the paper

“Creators on YouTube, for instance, may rely less on Patreon because YouTube's Partner Program offers substantial ad revenue, potentially creating selection effects in who monetizes via Patreon.”

In the paper

“users also show a preference for longer-form content too as evidenced by the rise of lengthy podcasts.”

In the paper

“Alternative explanations, including creator self-selection, audience composition, and platform-specific monetization cultures, remain plausible and are not ruled out by our design.”

In the paper

“Patreon and Twitter/X show moderate to lower power law proportions, suggesting that only the very top of the income distribution aligns with the power law”

In the paper

“Recent research suggests that incorporating mechanisms to surface underrepresented ('longer-tail') content (C. Anderson 2012) could enhance overall platform performance without necessarily sacrificing user satisfaction (Houtti et al. 2024; Yao et al. 2024).”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites C. Anderson 2012
• Not checked yet

  Checked before — its earlier verdict is held back until a new check runs.

  cites Houtti et al. 2024 source
• Not checked yet

  Checked before — its earlier verdict is held back until a new check runs.

  cites Yao et al. 2024 source

In the paper

“Since the probability of non-disclosure is strongly related to these observables (especially paid members), this specification approximates a setting in which earnings are missing at random conditional on the covariates.”

In the paper

“As γ decreases, emphasizing preferential attachment, the exponent α approaches 2, indicating a stronger rich-get-richer effect with a more pronounced tail in the distribution.”

In the paper

“it places creator earnings closer to concentrated capital income than labor income, consistent with compounding 'rich-get-richer' mechanisms.”

In the paper

“Platforms also differ in how prominently creators can display external links, the friction users face when clicking out, and community norms around external monetization - factors that can affect conversion rates independently of attention distribution.”

In the paper

“If we were to drop these observations, the right tail of the earnings distribution would be mechanically truncated, leading to a downward bias in both estimated inequality and the fitted power-law exponent.”

In the paper

“To the extent that this becomes baked into a platform's business model, with associated commercial incentives, it can become difficult to dislodge through market forces alone.”

In the paper

“Our single-platform restriction mitigates some selection concerns (these are creators who have chosen one external platform), but cannot eliminate them.”

In the paper

“This is why, for characterizing inequality, we focus on the estimated exponent α rather than on raw maxima, interquartile ranges, or other platform-specific summary statistics.”

In the paper

“For example, if 10,000 creators earn at least $100, we would expect around 1,000 to earn at least $1,000, about 100 to earn at least $10,000, and roughly 1 to earn at least $1,000,000.”

In the paper

“so in the notation p ( k ) ∝ k -α , the BA model gives α = 3.”

In the paper

“The preferential attachment model (Barabási and Albert 1999) is based on two mechanisms: growth, where a new node is added at each time step, and preferential attachment, where each new node connects to m existing nodes with a probability proportional to their degrees.”

What the sources say

• Not checked yetcites Barabási and Albert 1999 source

In the paper

“Instagram audiences, for example, may be less inclined toward direct patronage than Twitter/X audiences, independent of how attention is algorithmically distributed.”

In the paper

“Our findings are consistent with a central role for algorithmic design in shaping earnings inequality”

In the paper

“One interpretation is that less concentrated market structures are associated with less skewed algorithmic allocation of attention and earnings.”

In the paper

“As discussed in Sections 2.2 and 4.2, we cannot rule out alternative mechanisms, including: audience composition, creator sorting, and platform-specific conversion norms.”

In the paper

“If AI drives an explosion in content creation this risks reinforcing the dominance of existing platforms that intermediate and filter such content, as we come to rely on their algorithms for discovery even further.”

In the paper

“At the same time, the interquartile range focuses on the middle of the distribution and ignores the very top tail, so platforms with very different degrees of top-end concentration can have similar interquartile ranges.”

In the paper

“Platforms with lower α values, such as YouTube and Instagram, face significant risks, including discouraging mid-tier creators and amplifying the spread of viral or harmful content.”

In the paper

“Similarly, audience characteristics such as willingness to pay, cultural norms around patronage, demographic purchasing power vary across platforms and could independently shape Patreon earnings concentration.”

In the paper

“The key insight from solving this model is that the fraction of nodes with a given number of links k at time t follows a power law distribution. Specifically, the probability that a node has k links decays as k -α , where α = 1 + 1 1 -γ , a direct consequence of the rich-get-richer mechanism.”

In the paper

“Some econophysics models distinguish a two-class structure in income distributions, where working class income follows an exponential distribution (reflecting additive processes) and rentier or capitalist class income follows a power law (reflecting multiplicative processes). See (Ludwig and Yakovenko 2022; Drăgulescu and Yakovenko 2001).”

What the sources say

• Not checked yet

  Checked before — its earlier verdict is held back until a new check runs.

  cites Ludwig and Yakovenko 2022 source
• ?Not checked — no full text of the sourcecites Drăgulescu and Yakovenko 2001 source

In the paper

“Platform audiences may differ systematically in their propensity to financially support creators through external platforms like Patreon.”

In the paper

“Hu et al. (2023) recommends modifying the learning rates of recommendation algorithms to help balance the trade-off between exploration and exploitation, ensuring that high-quality content is recognized and then promoted quickly.”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Hu et al. 2023

In the paper

“This derivation shows that the power law in network connectivity depends on the parameter γ , which controls the balance between random and preferential linking.”

In the paper

“The relationship between platforms' algorithmic choices over what content to prioritize and their market power has been theorized in the context of advertising and data privacy (Srinivasan 2019; Strauss et al. 2024; O'Reilly et al. 2024).”

What the sources say

• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Srinivasan 2019
• ?Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  cites Strauss et al. 2024
• Not checked yet

  Checked before — its earlier verdict is held back until a new check runs.

  cites O'Reilly et al. 2024 source

In the paper

“Yet this may increase within-platform inequality if top-tier creators are the main beneficiaries of competitive poaching.”

In the paper

“An important alternative to the algorithmic explanation is that creators self-select into platforms in ways that generate the observed earnings patterns.”