Blog · arXiv Analysis · Published: August 12, 2026 · Modified: August 12, 2026 · Last reviewed: August 12, 2026

The Identical Post Becomes the Author Test

Two accounts can publish exactly the same text and still receive different rank exposure from the same custom feed at the same observation time. The words are controlled; the surrounding histories are not.

That makes duplicate posts a useful audit instrument. It does not make them a randomized identity experiment, and it does not turn decentralized feed choice into proof of equal entry within any one feed.

The Paper

The source is Yipeng Wang and Mohit Singhal’s Whose Posts Get Ranked: Identical-Text Exposure Gaps in Bluesky Custom Feeds, arXiv:2608.13879v1 [cs.SI, cross-listed in cs.IR], submitted August 14, 2026. The title page identifies both authors with Northeastern University. The three-page paper says it is to appear at the 20th ACM Conference on Recommender Systems and gives DOI 10.1145/3773078.3841250.

Bluesky’s official documentation describes custom feeds as algorithmic feeds that developers can create, publish, and make available for users to choose. The paper asks a narrower question than whether this architecture distributes control: when different authors’ text is identical, how evenly does one feed rank the copies?

One Text, Many Positions

The researchers queried public custom feeds roughly every ten minutes through the official app.bsky.feed.getFeed endpoint. Before filtering, they observed 7.3 million distinct posts returned in Top-50 lists. Their retained collection contains 297,093 complete snapshots from February through April 2026 across 1,366 feeds. A complete snapshot has 50 distinct ranks, 50 distinct posts, and no repeated rank-post pair after deduplication.

Within that collection, 64,853 matched sets across 250 feeds contain byte-identical text from at least two authors. In 33 percent of sets, the feed returned one copy in its Top 50 and omitted another; in 6.8 percent, one reached the Top 10 while another was absent from that list. Among the 58,318 two-copy sets with at least one returned post, the mean paired discounted-gain gap was 0.102. These are differences in measured rank exposure at a poll, not measured impressions or total audience reach.

How the Match Works

Every copy in a set had been published before the same feed response, was at most seven days old, and fell within 120-minute creation-time and age spans; the paper repeats key checks at 60 and 30 minutes. The primary outcome is reciprocal rank, 1 divided by rank, with discounted cumulative gain as an alternative. A copy missing from the Top-50 response receives zero in this rank measure.

Matched-set fixed effects hold the feed, snapshot time, byte-identical text, and other set-shared conditions steady. The analysis gives each set equal weight, clusters standard errors by identical-text group, and applies Benjamini–Hochberg false-discovery corrections. This comparison removes one obvious explanation—different words—without making the authors interchangeable. Media, thread position, moderation status, arrival order, and unobserved ranking signals can still differ.

The New-Author Association

The study calls an author “new” to a feed when that feed had not returned any of the author’s posts earlier in the researchers’ logs. Under that left-censored definition, new-author status is associated with 0.061 lower reciprocal-rank exposure for the same text. The paper reports p = 0.003, within-block adjusted q = 0.023, and all-test adjusted q = 0.047. The coefficient remains negative when each of the 250 feeds is omitted in turn and after controls for arrival order.

In 6,768 matched sets pairing a new author with a previously returned author, the new author ranked better 16 percent of the time. In the 746 matchups where the new author had more followers, that author still ranked lower 74 percent of the time. These comparisons weaken follower count and simple arrival order as complete explanations. They do not identify the ranker’s rule or prove that feed history itself caused the difference.

Decentralization Does Not Equal Entry

Plural algorithms can reduce dependence on one platform-wide ranking policy. They cannot, by existence alone, certify how any chosen feed admits unfamiliar authors. Operator diversity and within-feed exposure are separate governance questions. A user may be free to select among feeds while a particular feed still rewards the history it has already produced.

That creates a possible feedback mechanism, not a demonstrated fate. Being returned supplies observed history; observed history is associated with being returned again; missing from the list withholds the very record associated with later rank. A feed operator can test that loop without claiming malicious intent or demographic discrimination. The relevant question is whether a cold start is visible, measured, and deliberately handled.

An Exposure Receipt

A useful feed audit should publish more than an average rank. Its receipt would name the feed and operator, polling interval and depth, eligible candidate pool, matching rule, rank weights, treatment of missing posts, and observation window used to define prior history. It would separate text equality from media, reply, moderation, and thread features; report cold-start comparisons with uncertainty and multiple-test correction; and keep a change log when ranking logic changes.

The receipt should also mark its visibility boundary. “Not returned in this Top-50 snapshot” is inspectable. “Nobody saw this post” is not established by that record. Appeals or operator review should target the documented allocation rule, not convert an observational residual into a judgment about an author.

What the Audit Cannot Say

The paper studies naturally occurring duplicates rather than randomly swapping author identities. Most repeated material consists of news and sports headlines, fan hashtags, and greetings, often from paired accounts; 250 feeds containing such sets are not a census of Bluesky or its users. The authors note that most rankers are black boxes and that feed history begins only when their collection begins. They tested 304 varying media and post-type features and found none significant after correction, but absence of a detected association is not proof that every relevant difference was controlled.

This review checked the arXiv abstract, metadata API, full-text HTML, version-one PDF, and source archive, plus the official Bluesky API and custom-feed documentation. The archive contains the manuscript, bibliography, class and style files, and one figure; it does not contain study code, snapshot data, per-post records, analysis outputs, or a linked artifact repository. The reported results therefore could not be independently reproduced here.

The Spiralist lesson is not that identical words deserve identical rank in every context. It is that content equality lets hidden allocation become measurable. A distributed menu of rankers broadens choice; an exposure receipt shows what the chosen ranker actually did. One is architecture. The other is evidence.

Sources

Return to Blog