PaperlayerView exact revision on arXiv

Source-linked AI summary

Evidence of Open Access of scientific publications in Google Scholar: a large-scale analysis

Alberto Martín-Martín, Rodrigo Costas, Thed van Leeuwen, Emilio Delgado López-Cózar

arXiv:1803.06161v2cs.DL

TL;DR

Google Scholar can surface freely available versions, but its data-extraction limits left prior OA studies small or nonrandom. This study combined GS evidence with other sources to examine OA types. GS linked full texts for 54.7% of the sample, with a slightly higher share for 2014 than for 2009.

  • Problem

    Google Scholar-based OA studies were often small or nonrandom because the platform lacked an easy way to extract and analyse its data.

    Source: §1.5.1 · p. 10

  • Method

    The study combined Google Scholar data with DOAJ, CrossRef, OpenDOAR, and ROAR to examine OA variants and free availability from other sources.

    Source: §4.1 · p. 27

  • Results

    54.7% of the sample had Google Scholar links to freely available full texts; the share was 55.8% for 2014 and 53% for 2009.

    Source: §3.1 · p. 17

  • Takeaways & Limitations

    The combined sources provide a large-scale, detailed depiction of OA variants and free availability across scientific fields and countries.

    Source: §4.1 · p. 27

    Evidence covers only 2009 and 2014 as observed in summer 2016; OA levels may have changed, so the findings should not be extrapolated to other years.

    Source: §4.1 · p. 25

Abstract

from arXiv · show

This article uses Google Scholar (GS) as a source of data to analyse Open Access (OA) levels across all countries and fields of research. All articles and reviews with a DOI and published in 2009 or 2014 and covered by the three main citation indexes in the Web of Science (2,269,022 documents) were selected for study. The links to freely available versions of these documents displayed in GS were collected. To differentiate between more reliable (sustainable and legal) forms of access and less reliable ones, the data extracted from GS was combined with information available in DOAJ, CrossRef, OpenDOAR, and ROAR. This allowed us to distinguish the percentage of documents in our sample that are made OA by the publisher (23.1%, including Gold, Hybrid, Delayed, and Bronze OA) from those available as Green OA (17.6%), and those available from other sources (40.6%, mainly due to ResearchGate). The data shows an overall free availability of 54.6%, with important differences at the country and subject category levels. The data extracted from GS yielded very similar results to those found by other studies that analysed similar samples of documents, but employed different methods to find evidence of OA, thus suggesting a relative consistency among methods.

1. Introduction

The introduction situates Open Access in the expansion of web-based scholarly access and the difficulty of distinguishing sustainable, legal access from free-to-read copies. It motivates a broader assessment of Google Scholar’s evidence across subject categories and countries, asking how much literature is freely available, how much meets sustainable and legal criteria, and where it is hosted.

1.1. Beginnings of the Open Access movement

  • OA began as a response to access and affordability concerns, with self-archiving in repositories and publishing in OA journals established as two main routes.The Budapest Open Access Initiative crystallized these aims and defined OA as free availability on the public internet.Source: §1.1 · p. 2

1.2. Reactions of academic institutions, funders and publishers to OA

  • Funding mandates prompted publishers to adopt OA-compatible models, while variable sharing rules and changing embargoes complicate authors’ self-archiving.Publisher policies can differ by journal and venue, and may permit immediate sharing on a personal website while embargoing repository sharing.Source: §1.2 · pp. 3–5

1.3. New players in the system

  • Academic social networks provide widely used full-text access, but unchecked uploads can infringe copyright and files may disappear, unlike repository records.ResearchGate was reported as a frequent source of freely available texts; Sci-Hub is another source, providing copyright-infringing access.Source: §1.3 · pp. 5–6

1.4. Current landscape of free availability of scientific information

  • Free-to-read articles arise through publisher OA, delayed access, self-archiving, or copyright-infringing access, and can be copied across platforms; Google Scholar links to these sources except Sci-Hub.Repositories usually check copyright compliance, unlike personal websites and academic social networks; free availability must also be discoverable to be used.Source: §1.4 · pp. 7–8

1.5. Quantification of OA levels

  • OA levels matter to funders, libraries, publishers, countries, and researchers, and prior studies have measured them using crawlers and search engines across varied samples and aggregations.The introduction notes that methods include custom-crawler data and searches in services such as BASE, Google, and Google Scholar.Source: §1.5 · p. 8

1.5.1 Google Scholar as a source of OA evidence

  • Google Scholar links to freely available versions and has broad academic-web coverage, but extraction limits and citation-ranked, nonrandom samples constrained earlier OA studies.Most prior Google Scholar studies examined links beside primary versions, not links to secondary versions.Source: §1.5.1 · p. 10
  • Earlier Google Scholar studies of free scholarly literature were limited in scope, leaving a need for a relatively large-scale study of availability identified through Google Scholar.The paper aims to fill this gap.Source: §1.5.1 · p. 10

1.6. Research questions

  • The study asks how much recent literature is freely available by publication year, subject category, and author-affiliation country, how much is sustainable and legal, and how availability is distributed across web domains.It also asks what proportion is freely available without meeting sustainable and legal criteria.Source: §1.6 · p. 12

2. Methods

The study selected DOI-indexed Web of Science records from 2009 and 2014, retrieved Google Scholar links, and classified access using external source information. It retained 2,269,022 article and review records with good matches for analysis.

2. Methods

  • The sample comprised DOI-indexed documents in Web of Science’s SCIE, SSCI, or A&HCI published in 2009 or 2014.The two years were chosen to balance a large, varied sample with extraction manageability; 2014 allowed most delayed-OA articles to become available, while 2009 enabled trend analysis.Source: §2 · p. 12
  • Google Scholar was queried by DOI, and a custom Python script scraped result pages and followed secondary-version links to collect freely accessible full texts.Queries were run off campus to avoid subscription-only links; data extraction for 2,610,305 records took over three months.Source: §2 · pp. 12–13, Figure 2 · p. 13
  • Good matches required title similarity of at least 0.8, or weaker title similarity combined with first-author and publication-year agreement, with a separate rule for non-English GS titles.Titles were compared using Levenshtein similarity; the non-English-title rule addressed cases where Web of Science displayed English titles.Source: §2 · p. 14
  • 2,269,022 articles and reviews were analyzed after 97.6% of those document types met the study’s good-match criteria.Across the full sample, 96% of records were good matches; the article-and-review subset was used for analysis.Source: §2 · p. 14
  • Because Google Scholar does not identify the source type of a free-text link, the study combined external data and host classification to label publishers, repositories, institutions, academic social networks, harvesters, and uncategorized sources.Approximately 1,000 frequently occurring or primary-version hosts were manually classified; remaining unidentified domains stayed uncategorized.Source: §2 · pp. 14–15
  • Freely available documents were distinguished from OA documents, which were classified using publisher, repository, licensing, and directory information.Gold, Hybrid, Delayed, Bronze, and Green categories used DOAJ, CrossRef, ROAR, and OpenDOAR evidence; immediate Hybrid OA used a license delay of less than 30 days.Source: §2 · pp. 15–16
  • Free texts from sources other than publishers and repositories were counted as freely available, not OA, because their legal status and persistence were less certain.Publisher and repository sources were treated as more likely to be copyright-compliant and to provide long-term preservation.Source: §2 · p. 16
  • Google Scholar did not identify whether a free version was a preprint, accepted manuscript, or version of record, and determining this would have required examining each full text.That version-stage analysis was outside the study’s scope.Source: §2 · pp. 16–17

3. Results

Google Scholar linked to freely available full texts for 54.7% of the sample, with publisher, repository, and other-source access varying by year. Availability also differed substantially across disciplines and countries.

3.1. General overview

  • 54.7% of sample documents had Google Scholar links to freely available full texts; the rate was 55.8% for 2014 and 53% for 2009.The 2014 sample was larger and had had less time to become freely available when data were collected.Source: Figure 3 · p. 17, §3.1 · p. 17
  • 23.1% of documents were freely available from publishers; Gold and Hybrid OA rose from 3.3% to 10.1% and 0.5% to 1.5%, while Bronze fell from 14.1% to 12.6%.Delayed OA decreased from 2% in 2009 to 1.1% in 2014; most publisher-accessible documents were Bronze OA, usually under restrictive reuse terms.Source: §3.1 · p. 17, Figure 3 · p. 17
  • Green OA was higher in 2014 than 2009 (18.9% vs 15.7%), but repository-only access was higher in 2009 (11.3% vs 10.5%); other-source availability was 43.5% and 38.6%, respectively.Documents available only from other sources remained relatively stable: 21.8% in 2009 and 20% in 2014.Source: §3.1 · p. 18, Figure 4 · p. 18
  • ResearchGate hosted 32.6% of sample documents, more than publisher OA at 23.1%; 32.7% of its available documents were not freely available elsewhere.Other sources also included personal websites and harvesters.Source: §3.1 · p. 18, Table 3 · p. 19

3.2. Analysis by disciplines

  • Overall availability ranged from 60% in Medical and Life Sciences to 32.3% in Law, Arts, and Humanities, while multidisciplinary journals reached 93.6%.The article relates the multidisciplinary result to the category's inclusion of Gold OA mega-journals such as PLOS ONE.Source: §3.2 · p. 20, Figure 5 · p. 20
  • Bronze OA exceeded the combined Gold, Hybrid, and Delayed share in 28 of 35 disciplines, while Green OA-only exceeded those publisher types in 19 disciplines.Green OA-only was especially common in Astronomy and Astrophysics (56.2%) and Mathematics (21.1%).Source: §3.2 · pp. 20–21, Figure 6 · p. 21
  • FA-only was the most frequent availability type in most disciplines and exceeded publisher OA plus Green OA in 23 of 35; availability also varied within disciplines.Within Clinical Medicine, overall availability ranged from 85.9% in Tropical Medicine to 37% in Dermatology, where other-source FA was most common.Source: §3.2 · pp. 21–22, Figure 6 · p. 21

3.3. Analysis by countries of affiliation

  • 18 of 25 high-output countries exceeded the world OA level of 35.8%; all seven below it were Asian, with China at 27.2%.For overall availability, Brazil (71.6%), the Netherlands (70.8%), and Switzerland (70.5%) ranked highest; China (46%), Iran (45%), and Russia (44.3%) ranked lowest.Source: §3.3 · pp. 23–24, Table 4 · p. 23

4. Discussion

The discussion details limits from sample selection, OA categorization, and Google Scholar data access, while comparing the study’s findings with other methods and samples. Those comparisons show broad similarities alongside differences tied to source coverage and classification choices.

4.1. Limitations and further lines of study

  • The sample covers only DOI-bearing articles and reviews in three Web of Science indexes, whose coverage limitations may underrepresent Social Sciences, Arts, Humanities, and non-English journals.Books, conference papers, reports, and documents without a DOI were also outside the sample.Source: §4.1 · pp. 24–25
  • The Google Scholar evidence is a summer 2016 snapshot of 2009 and 2014 publications, so later backfilling or removals mean the reported availability may have changed.No extrapolation to other publication years is warranted.Source: §4.1 · p. 25
  • Google Scholar sometimes omits displayed free-access links or fails to merge duplicate records, causing this study to undercount availability in affected cases.The authors note that link-display errors were mostly fixed by the time of writing.Source: §4.1 · pp. 25–26
  • DOAJ omissions may shift some Gold OA into Hybrid or Bronze, incomplete CrossRef licenses may inflate Bronze, and repository access is assumed sustainable and legal.The study also does not distinguish preprints from accepted manuscripts or version-of-record articles.Source: §4.1 · pp. 26–27
  • Bulk extraction limits replication and took three months, although the study combines the largest Google Scholar sample to date with several external sources for broad OA coverage.The authors’ aim was to test Google Scholar’s usefulness for OA analysis, not to present a replicable method.Source: §4.1 · p. 27

4.2. Comparison of results with similar studies

  • Science-Metrix closely matches overall availability, country, and broad subject-area patterns, though its higher Green OA shares reflect classifying ResearchGate as Green; its 1science counts were calibrated upward.Science-Metrix reports 55% overall free availability in both years, compared with 53.1% and 55.8% here in 2009 and 2014.Source: §4.2 · pp. 27–28
  • Piwowar et al. (2018) found similar Unpaywall free-access levels for Web of Science documents when this study’s FA-only documents are excluded, with OA-type shares also remarkably similar.Their slightly higher percentages may reflect coverage, sampling, labeling, or collection-date differences.Source: §4.2 · pp. 28–29
  • van Leeuwen et al. (2017) reports lower OA; Jamali & Navabi (2015) differ in field patterns; Bosman & Kramer (2018) suggest nearly 10% of Unpaywall-covered documents are preprints.The study links these contrasts to differences in OA-source coverage, query ranking and sample composition, and which article versions are counted.Source: §4.2 · p. 29

5. Conclusions

The conclusions distinguish publisher and repository access from freely available copies on sources that may not be sustainable or legal, and show substantial variation by field and country. They find that combining Google Scholar with open-resource data can represent OA levels, while data access and incomplete license metadata constrain routine measurement and accurate categorization.

5.1. Answers to research questions

  • Free availability varied by field and country: Basic Life Sciences reached 67.5%, while Literature reached 14.2%; top-output countries ranged from about 70% to 45%.Brazil, the Netherlands, and Switzerland were near 70%; China, Iran, and Russia were near 45%.Source: §5.1 · p. 30
  • 33.9% of documents were available from sustainable and legal sources: 23.1% through publisher OA and 10.8% as Green OA only.Bronze OA accounted for 13.2% of all documents; users generally cannot redistribute or reuse these articles without publisher permission.Source: §5.1 · pp. 30–31
  • 17.6% of sampled documents were available from repositories according to Google Scholar.The authors also report room for Green OA to grow, while noting that authors often prefer ResearchGate to repositories and that the platform is proprietary and may disappear.Source: §5.1 · p. 31, §5.2 · pp. 32–33
  • 40.6% were freely available from sources not considered sustainable and legal, and 20.7% were available only from those sources.ResearchGate supplied 32.6% of all documents and was selected by Google Scholar as the primary version in 43.8% of cases; other prominent sources included repositories, academic social networks, and harvesters.Source: §5.1 · p. 31, RQ3. What is the distribution of… · p. 31

5.2. Final remarks

  • Google Scholar combined with CrossRef, DOAJ, OpenDOAR, and ROAR can provide a faithful representation of OA levels, and its results resemble other identification approaches.Because Google Scholar's data are not available to the research community, the authors say it cannot be used for regular analysis; Unpaywall is presented as an alternative for identifying legal OA.Source: §5.1 · p. 30, §5.2 · pp. 31–32
  • Most publisher-provided free access lacks a clear OA-compatible license, leaving continued access dependent on publishers and often granting users limited reuse rights.The authors call for discussion of minimum OA requirements and clearer OA mandates from policy makers and funders.Source: §5.2 · p. 32
  • The authors consider the European Union's goal of 100% OA by 2020 probably unrealistic for most EU countries.Many high-output European countries exceeded the world average but remained far from complete OA adoption.Source: §5.2 · p. 32
  • Incomplete article-level license metadata makes OA variants difficult to categorize; the authors propose standardized rights and machine-readable self-archiving policy records.Such records could help funders check mandate compliance and repositories monitor and release eligible documents under their license conditions.Source: §5.2 · p. 33
Loading 1803.06161v2…