PaperlayerView exact revision on arXiv

Source-linked AI summary

Middleware for Feed Recommendation in Practice: How Feed Creators Build, Maintain, and Sustain Custom Feeds on Bluesky

Tony Zhou, Leijie Wang, Amy X. Zhang

arXiv:2609.12958v1cs.HC

TL;DR

Although middleware has been proposed as an alternative to centralized feeds, little is known about the creators who build and sustain it. This study combines interviews with feed creators and tool developers with analysis of custom feeds, finding distinct creator orientations alongside constraints on data, expertise, and unpaid sustainment.

  • Problem

    Middleware has been proposed as an alternative to centralized algorithmic feeds, but feed creators’ roles and experiences remain understudied despite their importance to the ecosystem.

    Source: §1 · pp. 1, 3

  • Method

    The study interviewed 24 feed creators and two third-party developers and analyzed 88,302 custom feeds.

    Source: §1 · p. 3

  • Results

    Creators showed utility-providing and community-building orientations, while data access, technical expertise, and sustainment labor remained unevenly distributed.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Bluesky custom feeds demonstrate middleware feed recommendation at scale, but middleware ideals remain only partially realized amid unevenly distributed power and infrastructural labor.

    Source: §7 · p. 21

    The study did not solicit feed consumers’ or contributors’ perspectives, so its findings reflect creators’ experiences rather than those groups’ views.

    Source: §6 · p. 21

Abstract

from arXiv · show

Scholars have long proposed third-party middleware as an alternative to centralized algorithmic feeds: feeds built and distributed by independent feed creators. This vision saw no large-scale instantiation until Bluesky, a decentralized microblogging platform, introduced custom feeds in 2023. Although central to the middleware ecosystem, we know little about how feed creators understand their role, build feeds, and sustain them. Through interviews with n = 26 feed creators and third-party developers of feed-building tools, and analysis of n = 88,302 custom feeds, we identify two creator orientations---utility-providing and community-building. Additionally, creators struggle to maintain feeds that fully realize middleware ideals: they lack granular interaction data, receive little feedback, and lack technical expertise to act on either. Finally, creators sustain their feeds as unpaid hobbyists with little platform support and are divided on whether to monetize beyond covering costs. We conclude with design and policy implications for strengthening the middleware feed ecosystem.

1 Introduction

Middleware offers an alternative to centralized feed curation, but its governance and sustainability challenges shift to independent operators. Bluesky’s custom feeds provide a large-scale setting to study creators’ roles, practices, and challenges.

1 Introduction

  • Middleware envisions an open, composable layer of third-party services for recommendation and moderation, but shifts governance, accountability, and sustainability challenges to operators [34].Source: §1 · p. 1
  • Bluesky introduced custom feeds in 2023, allowing users to create, govern, and distribute recommendation algorithms as a large-scale instance of third-party middleware.The ecosystem also includes feed contributors, consumers, and third-party developers who build supporting tools.Source: §1 · p. 1
  • Feed creators remain understudied despite their central role and the relatively small population sustaining feeds; prior work on community moderators does not capture their distinct interactions and user expectations.Source: §1 · pp. 1, 3
  • The study asks how creators understand their role, how creators and developers build feeds, and how they maintain and sustain the ecosystem.Source: §1 · p. 3
  • Interviews and analysis of 88,302 feeds identify utility-providing and community-building creator orientations, while finding that data access, technical expertise, and sustainment costs remain unevenly distributed.The study included 24 feed creators and two third-party developers; creators and developers sustain the ecosystem through volunteer hobbyist labor.Source: §1 · p. 3

2 Related Work

Related work describes limits of centralized algorithmic curation and explores alternatives ranging from user customization to third-party middleware. Bluesky custom feeds make these ideas publicly shareable and algorithmic, while leaving feed creators’ practices poorly understood.

2.1 Algorithmic Feed Curation

  • Engagement-optimized feeds offer users limited control and can pressure creators to optimize for platform metrics, with prior research associating this dynamic with homogenized content and narrower diversity.Client-side re-ranking remains limited to content already surfaced by platform recommenders, while personal feeds require ongoing user configuration [42] [58] [10].Source: §2.1 · p. 3

2.2 Third Party Middleware for Social Media

  • Middleware could let users choose services that encode different community values, but raises privacy, security, infrastructure-cost, incentive, and competition challenges [34].Existing middleware-like approaches have only partially opened core platform services to third parties [4, 45, 57, 33, 10].Source: §2.2 · pp. 4–5

2.3 Custom Feeds and Algorithmic Choice on Bluesky

  • Bluesky custom feeds are publicly subscribable recommendation algorithms on an open protocol: users can create, discover, and adopt them, while third-party tools support creation.Creators author and maintain algorithms, unlike curators who only select or modify content; feed creation is off-platform.Source: §2.3 · p. 5
  • Prior work characterized feed popularity and keyword matching, but left creators’ identities, building and maintenance practices, and reasons for sustaining feeds insufficiently understood.Source: §2.3 · pp. 5–6

3 Method

The study combines interviews with feed creators and third-party developers with large-scale collection of public custom-feed records. Interviews were recruited from experienced participants and thematically coded; feed metadata were collected from ATProto repositories.

3 Method

  • The mixed-methods study pairs semi-structured interviews with creators and third-party feed-tool developers with a large-scale analysis of publicly accessible Bluesky custom feeds.Source: §3 · p. 6

3.1 Feed Creator Interviews

  • Researchers interviewed 24 Bluesky feed creators and two third-party developers; creators maintained custom feeds spanning varied popularity and topics, while developers built two of the platform’s most widely used third-party tools.Three creators had 10,000+ likes, seven had 1000+ likes, nine had 100+ likes, and five had fewer than 100 likes.Source: §3.1 · p. 6, §3.1.1 · p. 6

3.1.1 Recruitment.

  • Recruitment used Bluesky calls, related Discord and Reddit communities, and direct outreach to creators on a public list of top feeds; participants were screened for relevant creation or tool-development experience.Source: §3.1.1 · p. 6

3.1.2 Study procedure.

  • Interviews covered motivations, feed creation and growth, and views on custom feeds’ future; transcripts were edited for clarity and thematically coded, with weekly team meetings to align on emerging themes.Source: §3.1.2 · p. 8

3.2 Data Collection

  • For feed collection, researchers scanned accessible ATProto repositories from 146 known PDS instances and extracted feed records using the getActorFeeds endpoint.The collection yielded 97,719 unique feed DIDs before the reported dataset was narrowed to the analyzed sample.Source: §3.2 · p. 8

Appendix C.

The authors collected feed and interaction metadata, then analyzed 88,302 custom feeds from 41,955 creators. Their measures capture feed scale, participation, and the mostly small scale of individual feeds.

3.2.1 Collecting custom feed metadata.

  • Metadata collection captured feed likers, up to 500 recent posts, and interactions; feed likes proxy subscribers because pin and view signals are unavailable.Source: §3.2.1 · p. 8
  • An active consumer belongs to the subscriber base and liked a post in the 500-post window; the participation rate uses users who both liked and replied, relative to feed likers.Source: §3.2.1 · p. 8

3.2.2 Custom feed dataset and summary statistics.

  • After excluding feeds without returned metadata, the dataset contained 88,302 feeds from 41,955 creators; the median creator maintained one feed (mean 2.1).Source: §3.2.2 · p. 8
  • 49,069 feeds had subscribers; the median feed had 3 subscribers (mean 26.9; maximum 50,757), and 77.0% had 1–10.Source: §3.2.2 · pp. 8–9
  • The statistics align with prior analyses describing mostly small individual feeds alongside a handful of high-visibility feeds.The dataset also enabled ecosystem-scale analysis of creator activity, consumer participation, and feed structure alongside qualitative findings.Source: §3.2.2 · p. 9

4 Findings

The findings examine how feed creators build and maintain feeds with different orientations and levels of effort, and how they sustain them through unpaid or cost-recovery work. They also describe creators’ differing views on monetization and Bluesky’s support, alongside infrastructure costs that can exceed community support.

4.1 Characterizing Feed Creators and Feeds

  • Utility-providing creators treat feeds as content delivery, while community-building creators foster spaces where consumers contribute and interact.Utility creators focus on surfacing relevant content; community-building creators may help existing communities find each other on Bluesky.Source: §4.1 · pp. 9–10
  • Across 88,302 Bluesky custom feeds, 67.5% (59,582) had no creator-authored posts in the last 500-post window, and 78.8% had no active consumers by the study’s definition.The authors say this pattern more closely aligns with utility-providing feeds, though not always, and suggest most feed creators aim to provide utility.Source: §4.1.3 · p. 11

4.2 Building Custom Feeds

  • Most interviewed creators used third-party tools, commonly relying on keyword matching while extending feeds with vetted contributors and specialized filters.Some creators found regex difficult to learn, while tools offered different features for moderation and curation.Source: §4.2 · pp. 12–14

4.1 Characterizing Feed Creators and Feeds

  • Active-creator feeds had median average interactions per post of 11.94 (IQR = [4.00, 36.32]), versus 1.56 (IQR = [0.00, 9.22]) for inactive-creator feeds.The difference was statistically significant (p < 0.001).Source: §4.1.3 · pp. 11–12, Figure 4 · p. 13

4.2 Building Custom Feeds

  • Three creators built custom feed infrastructure when third-party tools could not support their data access, personalization, or complex feed logic needs.Examples included precomputed personalized recommendations and a feed built from scratch to support custom ranking, video, trending detection, and community membership.Source: §4.2 · p. 14

4.3 Maintaining Custom Feeds

  • Utility feeds often anticipate moderation through exclusions and labelers, while community feeds require more nuanced, reversible responses to rule violations.Creators also undertake proactive promotion and onboarding because they consider Bluesky’s feed discovery ineffective and unsupported.Source: §4.3 · pp. 15–16
  • Creators rarely receive explicit feedback, using analytics as passive signals and soliciting input through workarounds such as pinned-post comments or external channels.They report acting on feedback but retain discretion over final decisions.Source: §4.3 · p. 16

4.4 Sustaining Custom Feeds

  • Creators sustain feeds through unpaid personal investment, but they differ on whether Bluesky’s limited platform support is acceptable.Utility creators tied motivation to feed growth, while community-building creators described ongoing investment; sustaining feeds carries costs in time, motivation, and money.Source: §4.4 · p. 16
  • Creators generally accept monetization for infrastructure costs but differ on personal ad revenue: some reject profiting from others’ content, while others see modest compensation as community-supporting.Some creators solicit donations to cover premium tool costs without taking the funds personally; others argue that small payments from indie advertisers can support feed builders and their communities.Source: §4.4 · p. 17
  • C15’s high-traffic news feed demonstrates a viable model: ads make up roughly 2% of posts, generating an estimated $2,000–$3,000 over the past year, with roughly half donated to ATProto projects.C15 personally approves ads, which must carry a regulatory hashtag that users can mute to opt out.Source: §4.4 · p. 17
  • After donations from about 150 users and a $1,000 grant no longer covered server costs amid election-driven growth, D2 considered full-time work to sustain the tool.D2’s tool hosts thousands of feeds; the grant had covered a couple of months of costs, and donations that once covered most server expenses no longer sufficed.Source: §4.4 · p. 17

5 Discussion

Bluesky’s custom-feed ecosystem gives actors beyond the platform agency over feed design, but power and the costs of sustaining feeds remain unevenly distributed. The discussion traces these asymmetries and weighs ways to improve accountability without recreating centralized control.

5.1 Decentralized Middleware in Practice: Power and Cost Asymmetries

  • The ecosystem extends agency over feed design beyond Bluesky, yet power and the costs of sustaining feeds remain unevenly distributed.The authors argue that meaningful decentralization requires reducing both gaps without creating new forms of centralization.Source: §5.1 · p. 18, §5.1.3 · p. 19

5.1.1 Power asymmetries in platform and tooling control.

  • Bluesky retains privileged access to interaction data and controls interface visibility, while only a few creators can bypass third-party tooling.The paper reports that 98.9% of users’ data in 2024 remained on Bluesky’s servers; only three interviewed creators had expertise to build their own infrastructure.Source: §5.1.1 · p. 18

5.1.2 Cost asymmetries in sustaining custom feeds.

  • Feed maintenance relies largely on unpaid hobbyist labor, while creators absorb infrastructure or tool costs and remain divided over compensation and advertising.Creators also take on feedback-gathering and self-promotion with little platform support.Source: §5.1.2 · p. 18
  • The authors propose privacy-preserving engagement data, subscriber feedback, AI-assisted feed creation, and better discovery, while warning that recommendation ranking can centralize visibility decisions.AI assistance could help creators translate intent into feed logic and preview included or excluded posts; community-guided systems could also lower maintenance burdens.Source: §5.1.3 · p. 19

5.2 Characterizing the Feed Landscape Through the Lens of Powers

  • Feed paradigms distribute hosting, curation, and contribution differently; custom feeds separate these powers, trading some recommendation quality for less centralized control.Utility feeds delegate hosting to tools while creators curate; community-building feeds also let consumers contribute.Source: §5.2 · p. 20, Table 5 · p. 20

5.3 Holding Feed Creators Accountable: Transparency of Feed Logic and Its Trade-off

  • Publishing feed logic could help consumers assess whether feeds match their stated purpose, but exposing exact keywords invites contributors to game curation.The authors suggest showing examples of posts a feed surfaces and suppresses rather than revealing exactly how it works.Source: §5.3 · p. 21

6 Limitations and Future Work

The study identifies two limits on what its findings can establish: it did not directly gather consumer or contributor perspectives, and it could not classify feed algorithms at scale.

6 Limitations and Future Work

  • The study did not directly solicit feed consumers’ or contributors’ perspectives, leaving those experiences for future research.The authors focused on creators partly because they were publicly contactable and existing tools support creators.Source: §6 · p. 21
  • The quantitative analysis does not classify feed logic at scale because descriptions and recent posts often provide too little evidence about how feeds work.An attempted LLM-based inference did not yield meaningful conclusions.Source: §6 · p. 21

7 Conclusion

The conclusion presents Bluesky custom feeds as the first large-scale instantiation of middleware-centered feed recommendation. It argues that this model gives users and third parties meaningful control, while its ideals remain only partially realized amid uneven power and labor.

7 Conclusion

  • Bluesky custom feeds are presented as the first large-scale instantiation of middleware-centered feed recommendation for social media.The ecosystem gives users and third-party developers control over algorithmic curation beyond centralized recommendations.Source: §7 · p. 21
  • Feed creators have built infrastructure ranging from keyword filters to community spaces, sustained largely through hobbyist labor and third-party tools rather than direct platform support..Source: §7 · p. 21
  • The middleware ideals remain only partially realized because power and infrastructural labor are unevenly distributed, especially across third-party developers and feed creators.The conclusion also identifies tensions between personalization, transparency, and community-oriented curation.Source: §7 · pp. 21–22

A Participant Demographics

The participant-demographics material describes measures used to characterize participants and lists interview questions about motivations, feed creation, tools, and ongoing maintenance.

A Participant Demographics

  • Table 6 records participants’ years of social-media use and self-rated custom-feed expertise on a four-point novice-to-expert scale.The caption defines the scale from minimal knowledge to in-depth understanding.Source: Table 6 · p. 25

B.1.1 Motivations.

  • The interview questions covered why participants joined Bluesky, what prompted feed creation, and what kinds of feeds they made..Source: §B.1.1 · p. 26
  • Questions also asked about tool use, creation time, and whether participants continued maintaining their feeds..Source: §B.1.1 · p. 26

B.1.2 Feed Metrics and Growth.

The interview guide asks creators how they track feed metrics, assess success and growth, and decide whether to create more feeds.

B.1.2 Feed Metrics and Growth.

  • Creators are asked about subscriber counts, posting frequency, and whether a successful feed inspires them to create more.Source: §B.1.2 · p. 26
  • Creators are asked which metrics they track, including interactions, engagement, and click-throughs, and how those metrics inform growth or future planning.The guide also asks which metrics matter most to them.Source: §B.1.2 · p. 26
  • The guide asks what makes creators feel that their feed is working or successful.Source: §B.1.2 · p. 26

B.1.3 Maintenance Tasks.

The interview guide covers the work of maintaining feeds and creators’ approaches to moderating and curating content.

B.1.3 Maintenance Tasks.

  • Creators are asked about typical maintenance tasks, possible improvements, the effort involved, and whether they expect to continue over time.The questions also ask whether they wish to spend more or less time on maintenance.Source: §B.1.3 · p. 26

B.1.4 Moderation and Content Curation.

  • Creators are asked about unwanted content, their moderation role, and the satisfaction or frustration they experience while curating feeds.Source: §B.1.4 · pp. 26–27
  • The guide asks how creators identify and remove unwanted content, including use of third-party tools, Bluesky support, and updates to feed rules.Source: §B.1.4 · p. 26
  • The guide probes how creators communicate curation rules, balance transparency risks, handle subscriber reports and monitoring, and respond to rule exploitation or viewed comments.Source: §B.1.4 · pp. 26–27

C Feed Collection Data

The feed-collection section presents a sample JSON record and shows examples of feed and creator metadata.

C Feed Collection Data

D Descriptive statistics of our dataset

The section contains a table titled as descriptive statistics of the custom-feed dataset.

D Descriptive statistics of our dataset

  • Table 7 is identified as presenting descriptive statistics of the custom-feed dataset.Source: Table 7 · p. 29
Loading 2609.12958v1…