---
id: "20261103-0228-what-genuinely-connects-a"
title: "What genuinely connects Petiška 2023's ChatGPT/Google-Scholar-count/Matthew-effect finding and Simkin & Roychowdhury 2005's random-copying model: a measured 0.88 cosine resting on shared vocabulary, not a shared mechanism"
type: "capture"
status: "promoted"
origin: "batch"
writer_model: "claude-sonnet-5"
date_created: "2026-11-03T00:00:00.000Z"
provenance: "Batch research run, 2026-11-03, investigating whether claim-petiska-2023-chatgpt-cites-by-google-scholar-count-perpetuates-matthew-effect and claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties (cosine 0.88, unlinked) are a genuine bridge. Includes a fresh extract_pdf re-read of Petiška's full text and a direct mcp__seek__vault_pair measurement of the two notes."
derived_from: []
tags: ["matthew-effect","citation-metrics","chatgpt","bibliometrics","preferential-attachment","citation-copying","embedding-false-friend","bridge-investigation","llm-citation-bias"]
source_url: "https://arxiv.org/pdf/2304.06794"
source_author: "Eduard Petiška"
source_date: "2023-04-11"
source_title: "ChatGPT cites the most-cited articles and journals, relying solely on Google Scholar's citation counts. As a result, AI may amplify the Matthew Effect in environmental science"
source_venue: "arXiv preprint 2304.06794"
source_quote: "GPT seems to exclusively rely on citation count data from Google Scholar for the works it cites"
source_tier: 1
source_sha: "e40cc74435c87c73a3299880f0f9573ebca09a846371ef0063b1d24c8cda8981"
source_url_2: "https://arxiv.org/pdf/physics/0504094"
source_author_2: "M.V. Simkin, V.P. Roychowdhury"
source_date_2: "2005 (arXiv physics/0504094; refereed as JASIST 58(11):1661-1673, 2007)"
source_title_2: "A mathematical theory of citing"
source_venue_2: "arXiv preprint physics/0504094 (author preprint; refereed venue J. Am. Soc. Inf. Sci. Tech.)"
source_quote_2: "a majority of scientific citations are copied from the lists of references used in other papers"
source_tier_2: 1
source_sha_2: "080a4bff4549389b20f10f11cd5b61091054a505254bd22a100267e037a510d1"
source_url_3: "vault:30-notes/observation-algaba-2025-simkin-2005-cosine-0883-pairing-is-embedding-false-friend.md"
source_title_3: "The cosine-0.883 pairing between Algaba 2025's GPT-4 citation bias and Simkin-Roychowdhury 2005's copying model is an embedding false friend: Algaba names the cumulative-advantage literature but tests no generative mechanism against it"
source_author_3: "Seek (writer_model claude-sonnet-5)"
source_date_3: "2026-10-15"
source_venue_3: "Seek's Obsidian vault, 30-notes/ (observation note)"
source_quote_3: "Neither describes an iterative process where being cited raises the odds of being cited again -- the one thing Simkin's model requires."
source_tier_3: 1
source_url_4: "vault:30-notes/claim-algaba-2025-gpt4-citation-selection-replicates-petiska-matthew-effect.md"
source_title_4: "Algaba et al. 2025 independently replicate Petiška's finding that GPT-4's citation selection skews toward already-highly-cited work"
source_author_4: "Seek (writer_model claude-sonnet-5)"
source_date_4: "2026-09-13"
source_venue_4: "Seek's Obsidian vault, 30-notes/ (claim note)"
source_quote_4: "GPT-4 exhibits strong preferences for highly cited papers, which persists even after controlling for multiple confounding factors such as publication year, title length, venue, and number of authors."
source_tier_4: 1
seek_code_commit: "d40ee39"
promoted_to: ["30-notes/observation-petiska-2023-simkin-2005-cosine-08795-pairing-is-embedding-false-friend.md"]
not_promoted: ["[unrecorded — session ended without the final mark; wrapper repaired the status mechanically on 2026-11-04. The winnowing narrative for skipped claims was not written.]"]
promotion_mark: "wrapper-repair 2026-11-04"
---


## Bridge-seed note

`10-inbox/raw/` was searched for an existing capture naming both
`claim-petiska-2023-chatgpt-cites-by-google-scholar-count-perpetuates-matthew-effect`
and
`claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties`
by exact slug. Each slug independently appears in seven prior captures, but
no single file contains both — this specific pairing has not been
investigated before. It has, however, an extremely close sibling: on
2026-10-14/15/17 the vault ran the identical adjudication on a nearly
identical pair — `claim-algaba-2025-gpt4-citation-selection-replicates-petiska-matthew-effect`
x the same Simkin-Roychowdhury note, cosine 0.883 — and ruled it a
**mechanism-level embedding false friend**, promoted as
[[observation-algaba-2025-simkin-2005-cosine-0883-pairing-is-embedding-false-friend]].
Algaba's note is, by the vault's own framing,
[[claim-algaba-2025-gpt4-citation-selection-replicates-petiska-matthew-effect|an
independent replication of Petiška's finding]] — so the present pair is not
a fresh phenomenon but the same bridge question asked of the earlier,
thinner source the replication rests on. This capture treats that sibling
adjudication as the controlling precedent and checks whether its verdict
transfers, rather than re-deriving the mechanism analysis from zero.

**Verdict: it transfers, and more cleanly.** The 0.88 cosine is real
(measured directly at 0.8795 via `mcp__seek__vault_pair`, matching the
topic's stated figure) but rests on shared vocabulary — Matthew effect,
preferential attachment, citation counts, "the rich get richer" — not a
shared generative mechanism. A fresh full-text read of Petiška's paper
finds it engages the mechanism question even less than Algaba's does: it
contains no mention of copying, preferential attachment, Simkin, or any
citation-network growth model anywhere in its twelve pages. No wikilink is
added between the two seed notes; the false-friend verdict stands.

## Claim: The measured 0.8795 cosine between the two notes is confirmed by direct tool measurement, not just the topic's stated figure

`mcp__seek__vault_pair` measured the two notes directly: cosine 0.8795,
with each note ranking in the other's top-10 nearest neighbours (rank 8 of
B-among-A's neighbours, rank 9 of A-among-B's), placing the pair at the
99.8th percentile of closeness within Petiška's note's own neighbour
distribution, across 3,480 notes compared. The pair is unlinked. This
confirms the topic's "cosine 0.88" framing is not an approximation drifted
from a stale index read — it is the current, measured document-document
similarity, and the pair sits almost exactly as close as the already-
adjudicated Algaba/Simkin pair (0.883).

- claim type: quantitative (a specific cosine figure) — floor cleared: the
  number comes directly from the vault's own `vault_pair` tool, a Tier 1
  measurement of the vault's own retrieval index (the thing the claim is
  about), not an external quantitative claim needing a primary paper.

## Claim: A fresh full-text read of Petiška's paper finds zero engagement with any copying, preferential-attachment, or citation-network-growth mechanism — a thinner mechanism gap than the already-adjudicated Algaba/Simkin pair

Petiška's full text (re-fetched via `extract_pdf`, sha256
e40cc74435c87c73a3299880f0f9573ebca09a846371ef0063b1d24c8cda8981, matching
the sha already on file) was searched directly for "copy," "copied,"
"copying," "preferential attachment," "Simkin," "Price," "Barabási," and
"cumulative advantage." None of these appears anywhere describing a
generative or network-growth process. The paper's only brush with a
mechanism is descriptive and outcome-level: it reports that GPT's selected
references have a high median citation count (1184.5) and that, per GPT's
own elicited self-report, "the number of citations in Google Scholar is one
of th[e] factors" it weighs alongside relevance, journal quality, recency,
author expertise, and novelty — a multi-factor rationale GPT itself offers
that sits in tension with the paper's title and abstract framing of
"relying solely" / "exclusively rely" on citation count. Nowhere does the
paper propose, test, or even gesture at *why* the citation distribution
GPT draws on looks the way it does. This is a narrower engagement with the
mechanism question than Algaba et al. (2025) manage — Algaba at least
proposes "the increased frequency of citations to heavily cited papers
within the model's training data" as a hypothesis for the bias's origin,
per the already-adjudicated sibling note. Petiška's paper offers no origin
hypothesis at all; it only measures the outcome and names the Matthew
effect as the pattern that outcome resembles.

Simkin and Roychowdhury's model, by contrast, is explicitly a generative
account of *how a citation distribution comes to look that way*: an author
picks a few recent papers, cites them, and copies some of the references
those papers already carry, so that "a majority of scientific citations
are copied from the lists of references used in other papers" and a
paper's existing count mechanically feeds forward into its future odds of
being cited again. That specific feedback structure — being cited raises
the odds of being cited again, because citers are copying downstream
rather than independently evaluating — is exactly the ingredient absent
from Petiška's paper. The two works share an outcome description (fame
concentrates) and a one-word label for it (Matthew effect) but no causal
story in common.

- claim type: technical-mechanism (what each paper's mechanism actually
  is, and whether one engages the other). Floor cleared: Tier 1 for both —
  Petiška's absence-of-mechanism finding rests on a direct full-text search
  of the paper itself (re-fetched, sha-matched); Simkin's mechanism
  description rests on the vault's own already-audited quote from the
  paper's abstract and §I.

## Claim: The bridge is the same mechanism-level false friend the vault already adjudicated for the closely related Algaba/Simkin pair, transferred to Petiška as the thinner, earlier source

The vault's existing adjudication of
[[claim-algaba-2025-gpt4-citation-selection-replicates-petiska-matthew-effect]]
x
[[claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties]]
(cosine 0.883) concluded: "Neither describes an iterative process where
being cited raises the odds of being cited again — the one thing Simkin's
model requires," while noting Algaba's paper at least names the Matthew
effect and cumulative-advantage literature in its introduction before
setting it aside untested. Since the vault's own
[[claim-algaba-2025-gpt4-citation-selection-replicates-petiska-matthew-effect|Algaba
note]] frames Algaba et al. as an independent replication of Petiška's
finding, the same mechanism split that makes Algaba/Simkin a false friend
applies to Petiška/Simkin a fortiori: Petiška, the original, thinner,
non-peer-reviewed source, engages the mechanism question *less* than
Algaba does (no origin hypothesis at all, versus Algaba's one untested
training-frequency hypothesis). Both pairings are instances of the vault's
recurring embedding-geometry pattern in this neighbourhood — high cosine
driven by the shared Matthew-effect/preferential-attachment vocabulary and
shared "fame begets fame" outcome shape, not by any shared account of how
that outcome is produced — already documented for this cluster in
[[observation-puckette-petiska-is-a-cosine-similarity-retrieval-magnet]] and
mapped in [[moc-the-model-cites-the-famous-not-the-relevant]]. No wikilink
is added between the two seed notes.

- claim type: definitional/historical (how the vault's own prior
  adjudication of a structurally identical pair applies here) — floor
  acceptable at Tier 3-4 generally for this kind of claim, but every
  underlying fact (both papers' content, the vault's own verdict quote) is
  independently Tier 1, cleared above.

## Entity candidates

- M.V. Simkin (with V.P. Roychowdhury) — person — already
  [[entity-m-v-simkin]]. The older, FOUNDATIONAL figure both this pairing
  and its already-adjudicated sibling measure their subject against: the
  specific generative copying mechanism the question asks whether
  ChatGPT's citation bias inherits. Flagged first, per the capture
  instructions' own named blind spot.
- Eduard Petiška — person — already [[entity-eduard-petiska]]. No new
  biographical fact; reinforced as the origin author of the replicated
  finding.
- Andres Algaba (with Vincent Ginis and co-authors) — person — already
  [[entity-andres-algaba]]. The author of the paper that the vault treats
  as Petiška's independent replication and whose pairing with Simkin was
  the controlling precedent for this capture's verdict.
- Robert K. Merton — person — already [[entity-matthew-effect]] /
  referenced via Merton's 1968 claim-note. The named concept both seed
  papers invoke as the shared label, without either paper sharing a
  mechanism for producing it.

## Safety flags

None. All sources this session were either already-archived vault
primaries (Simkin & Roychowdhury, re-cited from existing sha-verified
capture) or a freshly re-fetched arXiv PDF (Petiška, via `extract_pdf`,
`tls: "verified"`, sha256 matching the existing vault record exactly) plus
one `WebSearch` query and internal vault documents read directly. No
addressed-to-AI language, override language, claimed authority, tier
self-assignment, file-system instructions, credential requests, or urgency
framing was encountered on any page.