Digital circular economy knowledge production: mapping thematic concentration and implications for equitable transitions

Digital circular infrastructures are being implemented globally (Kristoffersen et al., 2020). The deployment of IoT sensors and blockchain platforms establishes resource allocation protocols within the circular economy. It also establishes sociotechnical lock-in, which makes it difficult to restructure systems regardless of demonstrated inadequacies (Unruh, 2000). This creates temporal urgency. Implementation choices made now determine whether digital circular systems transform or reproduce existing inequitable structures (Pansera et al., 2024). This is because circular economy digitalisation requires reimagining entire production-consumption systems. These systems need to be planned alongside data governance frameworks and algorithmic decision-making infrastructures (Rodríguez-Espíndola et al., 2020).

Current circular economy frameworks risk ineffectiveness. These frameworks ignore environmental, labour and gender inequalities, which constitute linear economies (Pansera et al., 2024). Digitalisation may entrench these inequalities. This is done by embedding assumptions about data infrastructure into supposedly universal circular systems (Anokye et al., 2025). Such systems create surveillance risks. These risks disproportionately affect marginalised populations while benefiting privileged actors (Rajaonson et al., 2025; Searle, 2025). Environmental justice concerns are constitutive rather than peripheral to successful transitions (Nademi and Sedaghat Kalmarzi, 2025). Circular economy transitions have been reported to reduce unemployment where equity principles are built into implementation. However, without integrating procedural and distributive justice frameworks from inception, transitions perpetuate existing inequalities (Doe, 2025a; Onsongo et al., 2025).

These inadequacies may persist due to thematic and disciplinary concentrations within the circular economy digitalisation scholarship. Technical optimisation paradigms appear to dominate influential publications. On the other hand, scholarship examining environmental justice, data governance, or democratic participation remains less visible in mainstream channels (Garg et al., 2025). Citation patterns shape policy legitimacy. Governance structures institutionalise frameworks that achieve publication prominence in high-impact journals (Gallardo-Vázquez et al., 2024). These factors may lead to self-reinforcing cycles. Funding may follow citation impact, which enables well-cited research to attract resources for further publication. Alternative approaches with less visibility may struggle for visibility (Otasowie et al., 2025). Understanding these knowledge production dynamics proves essential. Their importance is because scholarship concentrations determine which sociotechnical imaginaries guide digital circular economy development and which alternatives remain underexplored.

Despite the growing literature on digital technologies (Dubey et al., 2020), artificial intelligence applications (Ahatsi and Olanrewaju, 2025) and blockchain implementations (Negi, 2024), limited studies have attempted to examine knowledge production patterns in digital economy research. Previous research has not comprehensively examined the theoretical frameworks, geographic contexts and thematic priorities that attain prominence in mainstream academic discourse. Few studies have explored how previous research shapes the sociotechnical imaginaries that materialise in digital circular infrastructures (Jasanoff and Kim, 2015).

To address these gaps, this study employs keyword co-occurrence mapping and semantic embedding techniques to map the intellectual structure of this emerging field. The analysis examines which research orientations, disciplinary approaches and geographic contexts achieve scholarly prominence and which ones do not. By examining these patterns, the study identified opportunities to broaden discourse to incorporate diverse theoretical perspectives, geographical epistemologies and socio-technical approaches. The study also identified underexplored research domains that warrant greater scholarly attention. This approach may help ensure digital circular systems serve diverse, rather than narrow, interests.

The remainder of the study is structured as follows. Section 2 presents the conceptual background. Section 3 presents the methods. Section 4 presents the results, and Section 5 presents the discussion of the results. Section 6 presents the conclusions of the study.

The circular economy has no accepted definition in literature. Kirchherr et al. (2017) identified 114 distinct definitions in use, differing in whether circularity is framed as waste management, resource efficiency or systematic economic transformation. The digital circular economy inherits this contestation and adds another layer. Corvellec et al. (2022) further argued that mainstream formulations tend toward technical optimisation while leaving distributional questions unexamined.

It refers to the application of digital technologies, including the Internet of Things, blockchain, artificial intelligence and digital product passports, to circular strategies such as tracking, sorting, remanufacturing and reverse logistics (Kristoffersen et al., 2020). These technologies are not neutral instruments. They embed assumptions about what can be measured, who owns the data and which circular activities the system can see. Informal recovery and repair economies, which handle a substantial share of global material circulation, are largely invisible to infrastructures designed around formal supply chains.

Jasanoff and Kim (2015) opined that sociotechnical imaginaries are collectively held and institutionally stabilised visions of desirable futures, made durable through their embedding in technological projects. The framework builds on earlier work on co-production, which holds that knowledge and social order emerge together rather than sequentially (Jasanoff, 2004). Imaginaries matter because they narrow the range of futures that appear feasible and because that narrowing precedes and shapes material investment. Work on path dependency shows how initial choices constrain later possibilities once systems accumulate complementary investments (Arthur, 1989; Unruh, 2000). Work on cumulative advantage in science shows how initial recognition compounds. Merton (1968) described the tendency for credit to accrue to already-recognised scientists; Price (1976) formalised the phenomenon as a stochastic process, and Barabási and Albert (1999) generalised it as preferential attachment in growing networks. Applied to a research field, these mechanisms suggest that early thematic concentrations may become self-reinforcing through their effects on citation, funding and career advancement.

Closure of this kind is not unique to circularity research. Comparable dynamics have been documented in artificial intelligence ethics, where technical fairness formulations have displaced broader questions of political economy (Greene et al., 2019), and in energy transitions research, where substitution framings have historically dominated over demand-side and sufficiency approaches (Sovacool, 2014). Establishing whether digital circular economy scholarship follows a similar trajectory requires empirical mapping rather than assumption.

Bibliometric analysis of circular economy research is well established, and several studies have mapped the field at different scopes. Türkeli et al. (2018) analysed circular economy knowledge production across the EU28 and China from 2006 to 2016, combining co-authorship, co-citation, bibliographic coupling and co-occurrence analysis at author, institutional, city, regional and national levels, supplemented by a researcher survey. Camón Luis and Celma (2020) mapped the thematic structure of 3,391 records from 2016 to 2019 and identified a shift in the field's centre of gravity from environmental science toward business administration. Negrete-Cardoso et al. (2022) examined 416 documents on circular economy and waste management, extracting five dominant research themes. Santibanez Gonzalez et al. (2023) analysed 163 articles at the intersection of circular economy and smart cities, addressing productivity, leading contributors, collaboration networks and prominent themes. Pal et al. (2025) analysed 1,223 documents linking the circular economy to the Sustainable Development Goals, mapping the intellectual and conceptual structure through co-citation and keyword analyses. Doe (2025b) mapped the intersection of circular economy research and Sustainable Development Goal 12. That study applied PRISMA screening to 7,900 documents published between 2015 and 2024, reporting concentrated publication venues and marked geographic disparity. Its design counts what the literature contains, whereas the present study asks whether thematic orientation predicts how a document is subsequently received.

Two of these findings bear directly on the present study. Camón Luis and Celma (2020) document a managerial reframing of circularity research, consistent with the technical and economic orientation reported in Section 4. Türkeli et al. (2018) identify governance and lifestyle among the themes present in the broader circular economy literature, which suggests that the thematic distribution reported here may be specific to the digital subset rather than characteristic of circularity research as a whole.

The six studies nonetheless share an analytical purpose. Each asks what literature contains: which themes recur, which authors and institutions are most productive, which countries publish the most and how these patterns shift over time. None asks whether the reception of a document depends on its thematic orientation. That question requires a different design, because prevalence and reception are separable properties. A theme may be uncommon because few scholars pursue it, because journals rarely accept it, or because readers rarely cite it once published. Counting occurrences cannot distinguish between these mechanisms, and each implies a different response.

A study using bibliometric methods to examine knowledge production must account for the limitations of those methods. Citation counts measure attention rather than quality and are shaped by field size, publication conventions and document age (Hicks et al., 2015). Database coverage is uneven: Scopus indexes English-language international journals more comprehensively than regional or non-English venues, which systematically understates scholarship from outside established publishing centres (Mongeon and Paul-Hus, 2016). Wouters (1999) argued that citation indices reflect the practices of their compilers as much as the literature they index and that treating them as transparent windows onto science obscures the truth. These limitations bear directly on the present analysis. A study documenting geographic asymmetry using a database with known geographic bias risks reproducing the pattern it reports. The findings below should therefore be read as evidence about the composition of the indexed international literature, not about the existence or quality of scholarship excluded from it.

The study employs bibliometric analysis to map the intellectual structure of digital circular economy scholarship (Donthu et al., 2021). The study integrates keyword co-occurrence mapping, lexical classification and semantic embedding techniques (Aria and Cuccurullo, 2017). The analytical approach positions bibliometric patterns as evidence revealing which research orientations dominate the field. It also revealed how geographic, disciplinary and thematic concentrations shape discourse boundaries (Cobo et al., 2011).

The study selected and used data from the Scopus database. The Scopus database is compatible with advanced bibliometric tools and provides rigorous indexing standards (Mongeon and Paul-Hus, 2016). The search strategy combined digital technology terminology with circular economy concepts in the title, abstract and keyword fields. The complete Boolean query, which was executed on September 25, 2025, is as follows:

TITLE-ABS-KEY ((“digital*” OR “industry 4.0” OR “iot” OR “internet of things” OR “ai” OR “artificial intelligence” OR “big data” OR “blockchain” OR “digital product passport” OR “machine learning” OR “deep learning” OR “smart” OR “automation”) AND (“circular econom*” OR “circularity”)) AND PUBYEAR >2016 AND PUBYEAR <2025 AND (LIMIT-TO (DOCTYPE, “ar”) OR LIMIT-TO (DOCTYPE, “re”)) AND (LIMIT-TO (SRCTYPE, “j”))

The search covered 2017–2024. The period captures when digital and circular economy discourse merged into distinct research domains. The 2017 starting point aligns with the EU's 2015 circular economy action plan implementation period and Industry 4.0 maturation (Kagermann, 2014). Preliminary analysis confirmed that publications explicitly linking these concepts appeared sporadically before 2017 but emerged systematically from this year forward. Document-type restrictions limit retrieval to peer-reviewed journal articles and reviews. It excluded conference papers, books and editorial materials.

The search strategy targets only scholarship that examines digital technology applications within circular economy contexts. It characterises dominant technical discourse rather than comprehensively mapping all sustainability or justice scholarship (Wohlin, 2014).

The study adopted the PRISMA 2020 guidelines (Page et al., 2021). The screening process was carried out across five stages. The result yields a final document of 1803 from an initial 5,391 retrievals (Figure 1). Two independent researchers assessed title–abstract–keyword relevance. They applied inclusion criteria that required substantive analytical engagement with both digital technologies and the circular economy rather than peripheral mentions. This reduced the dataset to 3,191 documents after eliminating publications mentioning terms only tangentially. Temporal verification confirmed 2017–2024 publication dates (3,166 documents). Subject area filters focused on social sciences, environmental sciences, business, engineering and interdisciplinary studies. It excludes purely material science or natural science research without socio-technical dimensions. This reduced the document set to 2,132. The final manual abstract review by independent researchers confirmed a substantive digital-circular economy intersection. Disagreements were resolved through consensus discussion. The final dataset comprised 1,516 articles and 287 reviews across 621 sources published in the English language.

Descriptive bibliometrics were adopted to characterise publication growth, citation impact and author productivity using bibliometrix v4.1.3 in R v.4.3.1 (Aria and Cuccurullo, 2017). The study used the Text Analysis for All software v.1.0.0 (Aria et al., 2024). Lexical classification and citation modelling were performed in R v.4.3.1 using the MASS package. Bradford's Law was adopted to assess whether core journals, moderate contributors and peripheral sources followed expected 1:n:n² patterns (Nicolaisen and Hjørland, 2007). The study adopted degree centrality to identify highly connected nodes and betweenness centrality to detect bridging positions (Freeman, 1978). The study also adopted the Louvain algorithm clustering with resolution = 1.0 and a minimum co-occurrence threshold of five to analyse keyword co-occurrence networks (Blondel et al., 2008). Thematic mapping was adopted to position research areas using Callon density and centrality indices (Callon et al., 1991). This distinguishes motor themes from basic, emerging and niche themes.

Lexical classification assigned each document a thematic orientation using three term lists specified before analysis. Matching was performed on the combined title, abstract, author keyword and index keyword fields. Text was lowercased, and publisher and copyright strings were removed before matching was carried out. Terms were matched as word stems so that optimis matches optimise, optimise and optimisation.

The technical and optimisation list comprises optimis*, efficien*, performance, algorithm, predict*, machine learning, deep learning, neural network, internet of things, IoT, blockchain, digital twin, supply chain, logistic*, forecast*, simulation, Industry 4.0, automat*, sensor, big data, traceabilit*, productivit*, lean and six sigma.

The social and governance list comprises justice, equit*, fairness, inclusi*, marginalis*, vulnerab*, povert*, informal sector, informal worker, waste picker, labour, worker, trade union, gender, women, indigenous, communit*, participat*, democratic, deliberat*, governance, power relations, distributive, procedural, human rights, well-being, livelihood, decent work, just transition, sovereignt* and colonial. The distributive list is a narrower subset applied as a stricter test. It comprises justice, equit*, fairness, marginalis*, povert*, informal sector, waste picker, women, gender, indigenous, just transition, decent work, human rights and colonial.

A document matching one or more terms from a list was recorded as engaging that vocabulary. Documents were then assigned to four mutually exclusive categories on this basis. Of the 1,803 documents, 1,146 match only the technical list and 92 match only the social list. A further 295 documents match both lists and the remaining 270 match neither. Eight terms in the technical list also appear in the Boolean search query and are therefore guaranteed some presence in the corpus. These are machine learning, deep learning, internet of things, IoT, blockchain, Industry 4.0, automat* and big data. A restricted version of the list excluding these eight was computed alongside the full version. On the restricted list, the technical share falls from 79.9% to 67.6% of documents. Both shares are reported throughout, and the restricted figure is treated as the preferred estimate. The social and distributive lists share no term with the query; so, their shares are unaffected by this restriction. Classification is lexical rather than substantive in character. Documents engaging distributive questions without using these particular terms are misclassified by the procedure. Reported proportions are therefore lower bounds rather than point estimates of engagement.

The study adopted multiple correspondence analysis, which transformed high-dimensional keyword matrices into interpretable dimensions. Dimensions were retained where the eigenvalue exceeded 1/Q = 0.02000, following Benzécri (1979), with inertia reported (Greenacre, 1993) after adjustment (Nenadic and Greenacre, 2007). Sentence transformers (all-MiniLM-L6-v2 model) were adopted to generate 384-dimensional semantic embeddings from abstracts (Reimers and Gurevych, 2019). Pairwise cosine similarities between term embeddings were computed, and term pairs above the similarity threshold were retained to form the network reported in Section 4.2. Reinert clustering with chi-squared tests (p < 0.001) was reported alongside the classification, noting that at this sample size, the test reaches significance under most partitions.

Publications increased from 13 (2017) to 560 (2024), representing a 71.18% compound annual growth rate that signals rapid epistemic consolidation (Table 1). This growth rate is well above rates typically reported for established scientific fields (Bornmann and Mutz, 2015). The rapid expansion indicates a limited temporal window for paradigm diversification before citation patterns consolidate into resistant structures. The average document age of 2.59 years confirms the field remains in a formative period when alternative approaches could still achieve mainstream visibility. The 1,803 documents distributed across 621 sources suggest potential fragmentation. However, subsequent network analysis revealed systematic concentration. The average of 44.13 citations per document indicates substantial scholarly attention despite the emerging nature of the field. The 40.38% international co-authorship rate appears high. However, the geographic analysis below further examines power asymmetries. Temporal dynamics reveal accelerating rather than stabilising growth. Publications increased 43-fold from 2017 to 2024, with particularly sharp acceleration post-2020. This acceleration coincides with the COVID-19 pandemic catalysing digitalisation discourse. This suggests external events shape research attention rapidly. Such responsiveness to external pressures raises questions about whether scholarship critically examines or uncritically promotes technological solutions.

Keyword co-occurrence was computed at document level across the 60 most frequent keywords, yielding 1,552 co-occurring pairs (Table 2). Circular economy co-occurs most often with sustainability (314 documents), Industry 4.0 (305) and sustainable development (297). Cosine similarity is reported alongside raw counts because association strength, which normalises by the product of term frequencies, penalises high-frequency terms and would rank the strongest pairs among the weakest.

The composition of the frequent-keyword set is informative. No keyword relating to justice, equity, gender or informal work appears among the 60 most frequent terms. At document level, circular economy co-occurs with justice in two documents and with equity in one, against 314 for sustainability.

The thematic map (Figure 2) positions research areas by Callon centrality and density (Figure 2). Circular economy and digital technology terms occupy the motor-theme quadrant, combining high centrality with high internal density. Sustainability and management terms occupy positions of lower centrality, and review and literature terms are centrally located, reflecting the substantial share of synthesis work in the corpus.

No term relating to social justice, governance or distributional questions appear on the map. This is consistent with the keyword frequency distribution reported in Table 2, where no such term reaches the 60 most frequent keywords, and with the thematic classification in Table 3, where no cluster is organised around distributive vocabulary.

The conceptual structure map in Figure 3 positions keyword terms using multiple correspondence analysis. The first dimension accounts for 55.06% of inertia and the second for 14.44%.

The first dimension separates two distinct vocabularies at opposite poles of the horizontal axis. Business, Industry 4.0, circular economy, supply chain management, blockchain, business models and industrial economics cluster at one pole. Economic aspect, human and article occupy the opposite pole, at some distance from the remainder of the map. Between them lie terms describing technologies and processes: artificial intelligence, machine learning, internet of things, recycling, waste management and life cycle analysis.

The second dimension separates strategic and policy vocabulary in the upper region of the map. The upper region contains green economy, sustainable development goals, environmental economics and digitisation terms. Applied and material vocabulary sits below, including electronic waste, product design, environmental technology and energy efficiency.

Citation impact differs across the six clusters (Kruskal–Wallis H(5) = 67.53, p < 0.001). Cluster composition proved only moderately stable under variation in the length of the text segments. Reclassifying at 30-word and 50-word segments yields adjusted Rand indices of 0.393 and 0.487 against the 40-word solution. These values indicate moderate agreement, so individual document assignments shift with the segmentation chosen. Interpretable technical, managerial and policy groupings recurred in all three runs, though their membership varied. Cluster membership should therefore be read as indicative rather than as a fixed property of documents (Table 3). The effect is modest, with eta-squared of 0.035 indicating that cluster membership explains 3.5% of citation variance. Literature review, firm capability and product lifecycle research are cited above the corpus, at rank-biserial correlations between −0.098 and −0.115. Environmental impact research does not differ detectably from the remainder of the corpus (p = 0.890). Two clusters are cited below at p < 0.001: policy and institutional research (r = 0.159) and technical methods (r = 0.232). These two low-cited clusters differ in kind and should not be read together as one finding. The technical methods cluster engages circular economy concepts only loosely, so its position reflects search breadth rather than selective reception. The policy cluster engages circularity directly, which makes its lower citation position the more consequential.

Figure 4 shows the semantic similarity network of the most strongly connected corpus terms. Nine communities emerge, each corresponding to a recognisable vocabulary within the digital circular economy literature. Supply chain, blockchain and Industry 4.0 terms form one community, and life cycle assessment, building and materials terms form another. Review and synthesis terms occupy a third, and energy, recycling and plastics terms a fourth. The term social sits in a small separate community with enterprise and collaboration terms. It connects to no other community in the network at the similarity threshold applied. No term relating to justice, equity, governance, labour or gender appears anywhere in the network.

Bradford's law identifies a concentrated core of sources within the corpus (Table 4). Eleven Zone 1 journals produced 618 articles or 34.3% of publications across 621 sources. Sustainability leads on volume with 222 articles, while Journal of Cleaner Production records the highest citation impact. Its h-index of 60 and 12,199 total citations come from 131 articles, fewer than Sustainability publishes. Volume and impact therefore vary independently, and publication counts alone are a poor proxy for influence.

Source-level m-index values are consistently higher for those journals that entered the field earliest. Journal of Cleaner Production reaches 7.5 and Sustainability 6.375, both having published since 2018. Journals that entered the field in 2021 record m-index values of 3.8 despite comparable h-index values. This pattern should be read with caution, because the m-index divides h-index by years since first publication. Citation accumulation is not linear in early years, so the measure is biased downward for recent entrants. The difference is therefore consistent with the properties of the measure as well as with cumulative advantage.

Bradford Zone 1 contains 11 journals, of which five published at least one document engaging distributive vocabulary and six published none. Journals associated with critical sustainability scholarship are almost absent from the corpus as a whole. Ecological Economics contributes one document and Local Environment one, while Geoforum and Environmental Justice contribute none.

Author-level metrics diverge from one another in ways that complicate any simple productivity-to-impact reading (Table 5). Bag and Tseng each published 17 documents but accumulated 2,920 and 840 citations, respectively. Kumar records the highest h-index in the table at 15, alongside one of the lowest citation totals at 980. Jabbour accumulated 2,306 citations from 12 documents, the highest citation rate per article in the table. Total citations, h-index and productivity therefore rank these authors differently, and no single measure orders them consistently. Fractionalised counts describe the average size of co-authorship teams rather than the position of authors. Bag's 4.45 fractionalised articles across 17 documents imply a mean team of approximately 3.8 authors. Kurniawan's 1.90 fractionalised articles across 11 documents imply a mean team of roughly 5.8 authors.

The most-cited documents indicate which framings have attracted the most sustained scholarly attention (Table 6). The most-cited article proposes a research agenda linking Industry 4.0 and the circular economy. The next four address Industry 4.0 enablement, blockchain for supply chains, machine learning in construction and big data adoption. Nine of the ten most-cited documents use exclusively technical vocabulary, and one combines technical and social terms. No document among the ten engages with justice, equity or distributional vocabulary at all.

Individual author comparisons cannot separate thematic orientation from the effects of venue, timing and collaboration. Citation counts across the corpus are therefore modelled directly rather than illustrated through paired cases. Each document was classified by lexical orientation using the lexicons described in Section 3.3.1. Documents matching only the social lexicon were compared against a reference category comprising documents matching only the technical lexicon together with those matching both lexicons. Counts were modelled by negative binomial regression with controls for publication year, document type, team size and journal citation level.

Table 7 reports three negative binomial models of total citations across the whole corpus. Thematic orientation enters as a three-level factor with technically oriented documents as the reference category. Model 1 controls for publication year, document type and the size of the author team. Model 2 adds the mean citation level of the publishing journal, computed on a leave-one-out basis so that a document's own citations are excluded. Model 3 replaces that scalar with journal fixed effects, restricted to the 28 journals that published documents of both orientations.

Socially oriented documents receive an estimated 58% fewer citations than technically oriented documents in Model 1 (IRR = 0.415). The estimate attenuates to 50% in Model 2 once journal citation level is taken into account (IRR = 0.500). In Model 3, it falls further to 28% and no longer reaches conventional significance (IRR = 0.719, p = 0.070). Documents matching both lexicons do not differ detectably from technically oriented documents in any of the three specifications.

The three models bound how much of the difference is attributable to where work appears. Conditioning on journal citation level reduces the estimated coefficient by about a fifth on the log scale. Holding journal identity constant reduces it by about three-fifths, since the scalar measure captures only one property of a venue. Journals of similar average citation level differ in readership, disciplinary scope and indexing coverage. Conditioning on that scalar is therefore not equivalent to holding the journal itself constant. The share of the difference attributable to venue therefore lies between a fifth and three-fifths, depending on how venue is measured. The within-journal residual is not distinguishable from zero at conventional levels, and Model 3 rests on 430 documents.

Three features of this model specification bear directly on how the estimates should be interpreted. Journal citation level is plausibly a mediator of the relationship between orientation and citation rather than a confounder of it. Conditioning on it therefore estimates a controlled direct effect and does not identify a total causal effect. Conditioning on a mediator can also induce bias where unobserved factors determine both journal placement and citation.

The count distribution contains 83 documents with no citations, or 5.4% of the estimation sample. A zero-inflated negative binomial specification of Model 2 does not improve on the standard model, with Akaike values of 13,631 and 13,630 respectively. The standard negative binomial specification is therefore retained throughout the analysis on that basis. Two further checks bear on how the journal citation measure was constructed and imputed. Excluding the 346 documents in single-document journals, whose journal mean is imputed as the corpus mean, moves the Model 2 estimate from 0.500 to 0.550. The direction is as expected, since mean imputation attenuates the journal coefficient and correspondingly inflates the residual orientation effect. The estimate is nonetheless stable in sign, magnitude and significance across that restriction of the sample. A journal mean that includes the focal document's own citations is mechanically correlated with the outcome. The leave-one-out construction removes that source of correlation and is used throughout the analysis reported here.

Figure 5 maps country scientific production across the corpus using affiliation entries as the unit. A document contributes once for every author affiliation recorded in a given country on this basis. India records the most affiliation entries at 482, followed by China with 476 and the United Kingdom with 432. The United States records 229, Brazil 165 and Australia 142 on the same measure. Counts assign Hong Kong and Macao to China and report Taiwan separately, which the bibliometric software does not do consistently. Papers with many authors or institutions inflate these counts relative to the number of underlying documents.

Table 8 reports the same corpus at document level, counting each document once per country represented. The ranking changes under this unit, and the United Kingdom leads with 272 documents. China follows closely with 238 and India with 230, while Italy records 201 documents. The United States records 131 documents, or 7.3% of the corpus, in fifth position. This unit is used for all citation analysis, since citations accrue to documents rather than to affiliations.

Per-country citation means partly reflect collaboration patterns rather than national research influence on its own. Internationally co-authored documents average 61.4 citations across the corpus, against 33.5 for single-country documents. Countries recording the highest per-document impact are also those with the highest international co-authorship rates. France, Malaysia and Morocco each exceed 85% international co-authorship, and all record above-average citation impact.

African participation in this corpus is limited in volume but not in citation impact. Ninety-three documents, or 5.16% of the corpus, include at least one African-affiliated author. These are distributed across 20 countries, with South Africa contributing 32 and Morocco 29. South Africa averages 128.50 citations per document, the highest of any country in the corpus. This figure rests on 32 documents, of which 26 are internationally co-authored and should be read accordingly. The University of Johannesburg is the second most productive institution overall, with 27 documents.

African scholarship is therefore under-represented in volume while being well cited when it appears. This pattern is not readily described as a form of exclusion operating through citation practices. It is more consistent with a restricted pipeline of entry into the indexed international literature.

Collaboration network analysis (Figure 6) shows a co-authorship structure organised into a number of identifiable communities. Degree within the restricted network averages 4.42 co-authors, with a median of four and a range from one to 20. Twelve authors, or 5.0% of the network, have a degree of ten or above. The most connected are Garza-Reyes with 20 co-authors, Luthra with 19 and Kumar Mangla and Kumar with 18 each. High degree indicates position within a collaboration structure rather than standing within the field. It does not establish influence over research agendas, funding allocation or editorial standards, none of which this analysis observes.

Restricting the co-authorship network to the 260 authors contributing three or more documents yields a graph of 240 nodes and 530 ties, fragmenting into 30 components. The largest component contains 146 authors and 406 ties, or 60.8% of the restricted network. Louvain community detection identifies 14 communities at a modularity of 0.791.

Ties are concentrated within communities, though not exclusively: 77 of 406 ties, or 19.0%, connect authors in different communities. Several authors occupy positions linking multiple communities. Kumar Mangla (betweenness 0.227) connects four communities, Bag (0.212) five, Luthra (0.192) five and Jabbour (0.189) three. These positions indicate connectivity across the collaboration structure. They do not establish influence over research agendas or evaluation standards, which this analysis does not observe.

Community structure in a collaboration network of this size and sparsity is expected, and the modularity value should not be read as evidence of unusual fragmentation without comparison against a degree-preserving null model. What the structure supports is a descriptive claim: collaboration is organised into identifiable groups, with a minority of authors connecting them.

Three analyses that use different procedures return findings pointing in the same overall direction. Lexical classification shows that 67.6% of documents use technical or optimisation vocabulary on the restricted list. The share is 79.9% on the full list, against 21.5% using social or governance vocabulary. Only 3.5% of documents engage justice, equity, gender or informality at all. No term relating to these appears among the 60 most frequent keywords in the corpus.

Thematic classification identifies six clusters, none organised around distributive vocabulary. The policy and institutional cluster is cited below the rest of the corpus, though the effect is small. Citation modelling estimates that socially oriented documents receive 58% fewer citations after controls. This estimated difference attenuates to 50% once the citation level of the publishing journal is included. Most of the difference therefore persists among documents appearing in journals of comparable citation level.

Lexical classification uses no citation data, and the thematic clustering uses neither lexicons nor citations. The regression uses no clustering. That all three place socially oriented scholarship below the corpus mean supports the inference more than anyone would alone. Publication volume grew 43-fold between 2017 and 2024, from 13 documents to 560. Across the same period, the share engaging distributive vocabulary showed no trend at all. The estimated slope is 0.14% points annually, which is not statistically distinguishable from zero.

Six of the 11 Bradford Zone 1 journals published no justice-engaging document across the eight years. Geographic participation is nonetheless broad, with India third by volume and Brazilian work cited above the corpus mean. Representation and thematic diversity therefore appear to be distinct properties of this field. These findings describe a distribution and its persistence over eight years.

The findings are situated within work on sociotechnical imaginaries (Jasanoff and Kim, 2015). That literature examines how collective visions of desirable futures become embedded in technological systems. Imaginaries matter because they narrow the range of futures that appear feasible before material investment occurs. This study examines an earlier stage in that process, namely the composition of the scholarship from which such visions are drawn.

Four separate analyses in this study converge on a single pattern of thematic concentration. Technical and optimisation vocabulary appears in 67.6% of documents on the restricted list and 79.9% on the full list. Social and governance vocabulary appears in 21.5% of documents in the same corpus. Documents engaging justice, equity, gender or informality account for 3.5% of the corpus. Thematic classification produces six clusters of moderate stability, none of which is organised around distributive vocabulary. No term relating to these questions appears among the 60 most frequent keywords in the corpus.

The persistence of this composition over the study period is more notable than its existence. Publication volume grew 43-fold between 2017 and 2024, from 13 documents to 560 annually. Across the same eight years the share engaging distributive vocabulary showed no measurable trend. Expansion on this scale reproduced the thematic composition of the field rather than broadening it.

This bears on how path dependency is understood in emerging research fields (Arthur, 1989; Unruh, 2000). Lock-in is usually described as a process in which early choices progressively constrain later ones. The pattern observed here is different, since the composition did not narrow over time but remained constant.

A field's thematic composition may reflect several distinct processes that prevalence counts cannot separate. Scholars may rarely pursue certain questions, journals may rarely accept work addressing them, or readers may rarely cite such work once it appears. Distinguishing among these three possibilities requires modelling reception directly rather than simply counting occurrence.

The models in Section 4.3 speak to the second and third of these possibilities. Socially oriented documents receive 58% fewer citations than technically oriented documents after controls for year, type and team size. Accounting for the citation level of the publishing journal reduces this estimate to 50%. Holding journal identity constant reduces it further, to a difference that is not distinguishable from zero at conventional levels.

Where a document appears therefore accounts for a substantial part of the difference in how often it is cited. The estimate is sensitive to how venue is represented, ranging from a fifth to three-fifths of the log-scale difference. A residual difference within journals may remain, but this design cannot establish it at the available sample size. The evidence is consistent with selection at the point of publication carrying most of the effect. It does not exclude a further contribution from how published work is subsequently read and cited.

Two further descriptive observations drawn from the corpus sit alongside these two model estimates. Socially oriented documents appear in journals averaging 22.1 citations per article, against 48.6 for technically oriented documents. Six of the 11 Bradford Zone 1 journals published no document engaging distributive vocabulary across the eight-year period.

These estimates identify an association only and do not establish the direction of causation. Journals with higher citation levels may attract technically oriented submissions for reasons unrelated to editorial preference. Authors may also direct such work to other venues in anticipation of its likely reception.

Critical scholarship on knowledge production often describes Global South contributions as numerically marginal and citationally subordinated. Neither characterisation holds in the present corpus. India ranks third among producing countries with 230 documents, ahead of the United States in fifth place. Brazilian-affiliated work averages 67.08 citations per document, against a corpus mean of only 44.13.

African participation in the corpus is limited in volume rather than in scholarly reception. Ninety-three documents include at least one African-affiliated author, amounting to 5.16% of the corpus. The University of Johannesburg is the second most productive institution overall, with 27 documents. These per-country figures partly reflect collaboration patterns, since internationally co-authored work is cited more across the corpus.

What distinguishes this scholarship is its thematic orientation rather than its geographic origin alone. The most productive authors in the corpus work predominantly on supply chain optimisation and Industry 4.0 adoption, irrespective of location. Southern institutional presence is therefore substantial, while work addressing informal circularity remains largely absent from the corpus. This suggests that participation in the field operates on terms substantially set by technical framings. Scholars in Southern institutions achieve both visibility and citation impact, though predominantly through research aligned with the dominant orientation.

The distribution documented here has consequences for what evidence is available to those implementing digital circular systems. Technical optimisation research generates evidence about system efficiency and comparatively little about distributional outcomes. Questions concerning labour displacement, value distribution and data governance remain correspondingly underexplored in the indexed literature.

Work on just transitions treats procedural and distributive justice as constitutive of successful sustainability transitions rather than peripheral to them (McCauley, 2025). Convergence between just transition and circular economy research nonetheless remains limited despite evident potential (Schultz et al., 2025). Empirical work indicates that circular transitions reduce unemployment when implemented with equity principles built into design (Doe, 2025a; Onsongo et al., 2025). Such findings depend on research traditions that this corpus appears to contain only sparsely.

Alternative framings exist within the wider literature but occupy peripheral positions in this corpus. De Coninck et al. (2025) propose a social circular economy in which personal well-being advances alongside environmental improvement. That framing explicitly integrates the dimensions which optimisation approaches treat as externalities or trade-offs. Whether such framings remain peripheral because of limited scholarly interest or limited publication opportunity cannot be determined from publication records alone.

Data governance illustrates the practical cost of this thematic distribution particularly clearly for practitioners. Digital circular systems require extensive tracking, which generates data of value beyond the sustainability purposes that justify its collection. Questions of data ownership, algorithmic accountability and surveillance risk follow directly from these architectures. A literature concentrated on optimisation provides limited conceptual resource for addressing any of them. These observations describe what the indexed literature currently contains and what it does not. They do not establish that the absent research would produce different implementation outcomes, which this design cannot test.

Preferential attachment and cumulative advantage are established results in network science and the sociology of science (Merton, 1968; Price, 1976; Barabási and Albert, 1999). Path dependency and lock-in are similarly well theorised in existing work (Arthur, 1989; Unruh, 2000). This study applies these frameworks to a new case and does not extend them.

The contribution to sociotechnical imaginaries is more specific and can be stated in one claim. Jasanoff and Kim (2015) locate imaginaries in their performance by states and publics and in their materialisation in infrastructure. The analysis presented here identifies an earlier stage prior to both of those sites. The range of imaginaries available for public performance or institutional adoption is constrained by what academic publishing makes visible. Section 4.3 locates a substantial part of that constraint at the point of publication itself. Journal placement accounts for between a fifth and three-fifths of the estimated citation difference. The narrowing is therefore traceable in large part to which venues accept which kinds of work.

This carries a further consequence for how imaginaries research treats its own evidence base. Such studies typically draw on policy documents, public discourse and expert testimony, treating academic literature as a record of what is known.

A second contribution concerns the relationship between geographic representation and thematic diversity of research. Geographic participation in this corpus is broad, and Brazilian and South African work is cited above the corpus mean. Scholarship addressing the informal circularity characteristic of many such contexts nonetheless remains largely absent. Representation and epistemic pluralism are therefore distinct properties, and the first does not establish the second. Accounts relying on authorship counts alone will misdescribe fields where participation selects for particular framings.

The findings indicate where intervention is likely to have effect, and the answer depends on how venue is measured. Journal placement accounts for between a fifth and three fifths of the estimated citation difference between socially and technically oriented work. Within journals that publish both kinds of work, the remaining difference is not distinguishable from zero.

This locates a larger share of the difference after publication rather than at acceptance. Measures addressing how published work is read and cited therefore reach a greater part of the effect. Reading lists, citation practice guidance and review-article coverage are relevant here in a way that a purely editorial account would not suggest.

Editorial practice nonetheless accounts for a real and measurable share of the overall difference. Six of the 11 journals constituting Bradford Zone 1 published no distributive work across eight years. Business Strategy and the Environment reached 8.9%, or four of its 45 documents. Variation of that size among similar outlets is consistent with an editorial rather than a subject-matter constraint. This analysis does not directly observe either submission or editorial decisions at any point.

A countable indicator for individual journals follows on directly from this particular set of findings. The share of published documents engaging distributive vocabulary stands at 3.5% across the corpus and 3.9% within Zone 1. Any journal can compute this figure annually from its own record without additional data collection.

For funders, the relevant finding concerns the composition of output rather than its sheer volume. Publication volume rose 43-fold between 2017 and 2024 while the distributive share showed no measurable trend. Instruments that expand output without specifying subject matter can be expected to reproduce the present distribution.

For researchers, the implication concerns subsequent reception at least as much as venue choice. Distributive work accumulates fewer citations than comparable technical work even within similarly cited journals. The disadvantage is therefore not removed by placing such work in a stronger outlet.

For those using this literature as evidence, its volume should not be read as coverage. A practitioner asking who bears the costs of digital circular systems will discover little in the indexed record. Commissioning primary work is likely to be necessary, since the gap has persisted through eight years of rapid expansion.

The study contributes to SDG 12 by characterising the evidence base on which the circular economy is based. Target 12.6 asks firms to adopt sustainable practices, and target 12.8 asks that people hold relevant information for sustainable development. Both depend on a research literature covering the consequences of circular systems as well as their technical performance. This analysis shows that 3.5% of the digital circular economy literature engages distributive questions. Implementation guidance drawn from this base will therefore be well supplied on efficiency and thin on distribution. The study also offers a countable indicator, since the distributive share can be computed annually from any journal's output.

The contribution to SDG 10 concerns the way inclusion in knowledge production is measured. Target 10.2 concerns social and economic inclusion, which research policy commonly operationalises through geographic participation in authorship. This study finds broad geographic participation alongside a narrow thematic range within the same corpus. India ranks third by volume, and Brazilian work is cited above the corpus mean, while distributive questions remain at 3.5%. Authorship diversity and the diversity of questions are therefore separable, and the first does not evidence the second. Monitoring frameworks that rely on authorship counts alone will record inclusion that the research content does not reflect.

Reliance on Scopus introduces linguistic and geographic bias which the present analysis inherits directly. Scholarship in non-English or regional venues is underrepresented, which plausibly understates output from outside established publishing centres. The citation models estimate association only and do not identify the direction of causation. Journal placement is not randomly assigned with respect to thematic orientation, and authors may direct work to venues in anticipation of acceptance. The residual difference after venue adjustment is sensitive to how venue is represented in the model. It is large and clearly significant against a scalar journal measure and small and not significant against journal fixed effects. Screening reliability was not recorded in a form that permits a reliability coefficient to be computed. The dual-screening procedure is described in Section 3.2, but no quantitative agreement statistic can be reported.

Bibliometric methods document structural patterns but cannot recover the individual decisions that produce them. Whether authors avoid distributive framings, meet resistance in review, or publish them outside this corpus cannot be distinguished from publication records. Interviews with authors and editors, analysis of review outcomes and examination of funding calls would address this directly. Such work would test the interpretation advanced here against evidence of a different kind. Knowledge produced outside academic publication channels is absent from the present corpus by construction. Practitioner networks, civil society organisations and communities operating informal circular systems generate knowledge that indexed journals do not capture. Its absence here is evidence about the composition of that literature rather than about the knowledge itself. Participatory and ethnographic research approaches would reach such knowledge, whereas bibliometric methods evidently cannot.