Review Reports - Mapping Sustainable Digital Commerce: A PRISMA-Guided Semantic Network Review Triangulated with an Anchored Search-Interest Landscape (2022–2026)

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Overall, this article is impressive. The author presents their ideas logically, making the manuscript easy to understand and follow. However, I have a few suggestions to strengthen the paper:

Citations are needed in several sections, particularly the introduction and literature review. Some statements appear to be the author's own ideas but are actually facts that require sourcing.
In the first paragraph, when discussing the growth of environmental and ethical consumption, it would be beneficial to include supporting statistical data.
When mentioning "existing reviews," please provide specific examples of those articles.
Briefly define the PRISMA framework upon its first mention in the text.
The research questions should be formatted in standard text rather than bold font.
Additional citations may be needed in the discussion section to support the claims being made.

Note: I wrote this feedback myself and only used AI to help correct the grammar.

Author Response

Response to Reviewers

Manuscript: Mapping Sustainable Digital Commerce: A PRISMA-Guided Semantic Network Review Triangulated with an Anchored Search-Interest Landscape (2022–2026)

Journal of Theoretical and Applied Electronic Commerce Research | Manuscript ID: jtaer-4596813

Dear Editor and Reviewers,

We are grateful for the assessment that the argument is logical and easy to follow, and for six specific suggestions, all of which we have adopted.

1.1 Citations are needed in several sections, particularly the introduction and literature review

"Some statements appear to be the author's own ideas but are actually facts that require sourcing."

Agreed. We had written several empirical claims as assertions. The introduction now sources the growth of sustainability-oriented consumption (ref 49), the expansion of the research literature (refs 31, 35), and the normalisation property of search platforms (ref 42, and refs 43–46 in Section 2.5). Section 2.5 now carries eight new citations covering both the applied and the methodological literature. Nineteen references have been added in total.

1.2 Supporting statistical data for the growth of environmental and ethical consumption

Added to the opening paragraph:

The commercial footprint of this shift is measurable: across roughly 600,000 product lines in 32 consumer categories, items carrying environmental, social or governance claims grew about 28% cumulatively between 2017 and mid-2022 against about 20% for items without them, and accounted for 56% of all category growth over the period [49].

We note in passing that this market statistic sits in productive tension with our own search-side finding, and the Discussion now treats the discrepancy as something to be explained rather than assumed away.

1.3 Give specific examples when mentioning "existing reviews"

Agreed, and this overlaps with Reviewer 2’s first point. The introduction now names them, and a new paragraph in Section 2.1 maps all seven syntheses we were able to locate, with their boundaries stated: Rita and Ramos [31], Cano et al. [32], Mucowska [33], Aziz et al. [34], Zhang [35], Bălan [36] and Shafi et al. [37]. Six of the seven rely on a single database, four are bounded to one function or product category, and none contains a measure of public behaviour. Our claim of distinctiveness is now stated against that specific set rather than in general terms.

1.4 Briefly define the PRISMA framework at first mention

Added in Section 1:

PRISMA, the Preferred Reporting Items for Systematic Reviews and Meta-Analyses, is the reporting standard that requires every record identified, screened, excluded and included to be counted and justified in a published flow diagram [28].

1.5 Research questions should be in standard text rather than bold

Done. RQ1–RQ6 in Section 3 are now set in standard text.

1.6 Additional citations in the discussion

Section 7 now cites the prior reviews where it claims novelty, the search-trend methodological literature where it claims a correction, and van Eck and Waltman [47] and Mann [48] for the two new analyses. We have also removed two claims in Section 7 that no citation could support, as Reviewer 2 asked in their ninth point.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript argues that previous literature reviews have not addressed public search behavior and that search-trend studies have generally failed to make cross-query comparisons comparable. However, these claims are presented rather broadly without sufficiently mapping the existing review and search-trend literature. The authors should identify and discuss relevant prior systematic/bibliometric reviews of sustainable digital commerce, as well as previous applications of Google Trends/Naver DataLab in consumer or sustainability research. This would allow readers to better assess whether the proposed triangulation and anchored normalization represent a genuine methodological contribution.
The review relies on 700 relevance-ranked Google Scholar records retrieved via Publish or Perish, followed by post hoc screening using the three-axis criterion. Because recall cannot be quantified and the authors themselves acknowledge instability in repeated retrieval, the annual increase from 38 studies in 2022 to 108 in 2025 should not necessarily be interpreted as evidence of actual growth in the research field. I recommend either validating the corpus against Scopus/Web of Science or substantially moderating claims concerning the expansion of the literature.
Axis C requires perception or behavior-related terminology. Consequently, studies focused on regulation, governance, certification, platform accountability, technological infrastructure, or environmental impacts may be excluded if they do not explicitly use perception/behavior terms. Therefore, the finding that governance represents only 3.6% should be interpreted as the proportion of governance-related studies satisfying the study's specific eligibility criterion, rather than necessarily the proportion of governance research in sustainable digital commerce as a whole. The authors should clarify this scope more explicitly.
The thematic classification is based on deterministic keyword/regular-expression rules applied to titles and abstracts. Although this improves reproducibility, construct validity remains uncertain because no human coding or inter-coder reliability assessment was performed. This is particularly important for categories such as “platform and governance,” “greenwashing and trust,” and “green products and purchase behavior,” where terminology can overlap. A manually coded validation sample, with agreement statistics or at least an error analysis, would substantially strengthen confidence in the thematic findings.
The conclusion that subdomains are “mutually invisible” is interesting, but the comparison of mean raw co-occurrence frequencies (1.40 versus 12.6) may partly reflect differences in keyword frequency. Similarly, the k-core results are dependent on the selected co-occurrence threshold. I recommend supplementing the raw co-occurrence analysis with a frequency-normalized association measure such as Jaccard similarity, cosine similarity, association strength, or PMI, together with an appropriate null/randomization model. This would establish whether the observed separation remains after controlling for the frequency of individual terms.
The manuscript treats secondhand trading as the benchmark for mainstream commerce while simultaneously classifying secondhand/circular exchange as a sustainability-oriented category. This creates a conceptual concern because the denominator is not necessarily a neutral measure of mainstream commerce attention. In addition, a single anchor may substantially influence the reported magnitude of 0.7–4.1%. I recommend testing alternative anchors representing mainstream commerce and reporting whether the main conclusions remain stable across anchors.
The reported verification shows zero deviation and a rescaling factor of exactly 1.000 across requests. However, this appears to show that the probe series is reproduced under the authors' collection configuration rather than that the anchored ratio is accurate relative to an independently obtained common-query comparison. The authors should clarify what the validation establishes and, if technically feasible, compare the anchor-bridged results with results obtained from a single joint query containing the same terms. The methodological contribution would be considerably stronger with such a sensitivity check.
Several search terms were replaced following an initial validation exercise, and the ESG group was subsequently excluded based on its magnitude and temporal behavior. Although the authors provide a rationale, these decisions may introduce post-hoc selection effects, particularly because the excluded ESG group was substantially different from the remaining terms. The authors should report sensitivity analyses, including the original terms and/or ESG group, and explain whether the principal conclusions change under alternative reasonable keyword specifications.
The analysis contains only four complete annual observations for the principal change analysis, and no formal trend or uncertainty analysis is presented. More importantly, search interest represents information-seeking behavior, not directly sustainable attitudes or purchasing behavior. Therefore, statements such as “whatever suppresses sustainable purchasing is already suppressing sustainable searching” appear causal and are not established by the present design. Similarly, the claim that the study identifies a “pre-search component” of the attitude–behavior gap should be softened unless the authors provide stronger empirical evidence connecting search behavior with attitudes and subsequent purchasing.
There are a number of issues requiring careful cross-checking. For example, the visual presentation in Figure 2 appears inconsistent with some values in Table 2; the manuscript states both that the principal brokers remain unchanged across thresholds and that changes occur in the core structure; and the abstract reports a 39–63% contraction, whereas the discussion describes the decline as “half to two-thirds.” In addition, the conclusion states that governance vocabulary “appears nowhere in the network,” although the platformis actually present as a network node (k-core = 10); the more accurate statement is that governance-specific vocabulary does not reach the top-33 keyword selection. These inconsistencies should be resolved, and the conclusions should be aligned.
The manuscript presents an interesting and potentially valuable combination of systematic literature mapping, semantic network analysis, and anchored search-interest data. The topic is timely, and the authors make a useful effort to document methodological limitations. However, the strongest conclusions; particularly the **3.6% governance “blind spot,” the “mutually invisible” subdomains, and the 0.7–4.1% search-interest magnitude—currently depend heavily on corpus construction, keyword selection, network thresholds, and the choice of anchor. Addressing these methodological and interpretive issues would substantially improve the manuscript and make its contribution more convincing.

Author Response

Response to Reviewers

Manuscript: Mapping Sustainable Digital Commerce: A PRISMA-Guided Semantic Network Review Triangulated with an Anchored Search-Interest Landscape (2022–2026)

Journal of Theoretical and Applied Electronic Commerce Research | Manuscript ID: jtaer-4596813

Dear Editor and Reviewers,

We thank both reviewers for reviews that were detailed, fair and, in several places, correct in ways that changed our conclusions. Reviewer 2 in particular identified two problems we had not seen: that our central network finding was substantially a keyword-frequency artefact, and that our headline search-magnitude figure was conditional on an anchor that is not a neutral benchmark. We have re-analysed both, and we report the results whether or not they favour our original claims. One of them does not, and the manuscript now says so.

All changes are marked in red in the revised manuscript. Section and table numbers below refer to the revised version. New analyses add Tables B4, B5, C4 and D3, and Figure 2 has been regenerated. Nineteen references have been added (refs 31–49).

Summary of the principal changes

The semantic-network finding has been re-estimated against frequency-normalised association measures and a randomisation model. Most of the reported ninefold gap is a frequency effect. We now report what survives, which is narrower but better identified (Section 5.4, Table C4).
The 0.7–4.1% magnitude is now reported as anchor-conditional, with a full alternative-anchor analysis and the excluded ESG group restored for inspection (Section 6.1, Table B4).
Formal trend tests with uncertainty intervals replace the comparison of annual point estimates (Section 6.2, Table B5). The central differential survives; the causal language does not and has been removed.
The thematic coding has been validated against two blind coders. They agreed almost perfectly with each other (κ = .91) and with the rule only 57.6% of the time. The errors are systematic, their cause is identifiable, and two of our reported shares change as a result (Sections 4.2, 5.2, 8.3; Tables D3 and D4).
Growth-of-the-field claims have been withdrawn and replaced with corpus-level statements (Section 5.1).
Figure 2 was inconsistent with Table 2. It was plotted from a superseded coding run and has been regenerated from the Table 2 values. We are grateful this was caught.
Prior reviews and prior search-trend applications are now mapped and cited (Sections 1, 2.1, 2.5).

Reviewer 2

This review identified problems in the two findings we had considered strongest. We have followed the recommendations, and two of our conclusions have changed as a result.

2.1 Prior review and search-trend literature is not sufficiently mapped

Accepted. Our claim of a gap was asserted rather than demonstrated. Section 2.1 now carries a new paragraph mapping the seven available syntheses of sustainable digital commerce with their database coverage, corpus size and scope limits. Section 2.5 now maps the search-trend literature in two parts: applications, including the nowcasting template [38], conservation attention [39], environmental awareness as a construct [40], food-sustainability interest [41], and Naver DataLab shopping-category volumes in the Korean market [42]; and the methodological literature on limitations [43–46].

That mapping also obliged us to narrow our methodological claim. Benchmark terms are already discussed in the infodemiology literature [43], so anchoring is not novel in itself. Section 2.5 now states what is actually offered:

The anchoring logic is not novel in itself — benchmark terms are discussed in the infodemiology literature [43] — and what is offered is its systematic application, empirical verification, and transparent reporting in a commerce setting, together with the observation that the resulting magnitudes are themselves conditional on which term serves as the anchor.

2.2 Recall cannot be quantified, so the rise from 38 to 108 studies should not be read as growth

Accepted without reservation. We could not validate against Scopus or Web of Science within this revision, so we have taken the second option offered and withdrawn the claim. Section 5.1 now reads:

—but this series describes the corpus, not the field. Because identification proceeded through a relevance-ranked query whose recall cannot be quantified, and because re-executing the same query returned materially different counts (Section 8.3), the rising counts are consistent with growth in the literature but do not establish it: they are equally consistent with improving index coverage of recent years, which is a known property of the source. The within-corpus comparisons that follow — thematic shares by year — are unaffected, because they are proportions within each annual slice rather than counts across slices, and these are the quantities on which the argument rests.

The corresponding claims in Sections 7.1 and 8.1 have been amended, and validation against an indexed database remains the first item in Section 8.4.

2.3 Axis C excludes governance, regulation and infrastructure studies, so 3.6% is not the governance share of the field

Correct, and we have quantified the effect rather than merely conceding it. Re-applying the three axes to the 700 retrieved records isolates the bias: of the records carrying both a commerce and a sustainability term, 119 fail axis C, and governance vocabulary appears in 12 of them (10.1%) against 6.2% of records satisfying all three axes. Governance research is over-represented among the exclusions by a factor of about 1.6. Section 4.2 reports this, and concludes:

Every theme share reported below is accordingly a share of consumer-facing sustainable digital commerce research, not of the field at large, and the governance figure in particular is a lower bound.

Section 5.2, the abstract and Section 8.1 have been amended to match. The manual validation reported under point 2.4 bears on the same question from the other direction and puts the corrected governance share at 6.2% (95% CI 1.5 to 12.7), above the 3.6% the rule returns and in the direction this axis-C analysis predicts.

2.4 Construct validity of the thematic coding; a manually coded validation sample with agreement statistics or an error analysis

We have carried out the validation. It did not come out in our favour, and we report it in full.

A stratified random sample of 100 records was drawn — every record assigned to the four smallest substantive themes, where boundary error matters most, plus a random draw from the remainder. Two authors coded each record independently from title and abstract, blind to the rule assignment. The coders agreed with one another on 92 of 100 records: Cohen’s κ = 0.907, Krippendorff’s α = 0.907. The eight disagreements were left unadjudicated and excluded rather than resolved by us, so everything below rests on records two blind coders reached independently.

Against that consensus the rule assignment is correct in 53 of 92 cases — 57.6%. The errors are not random. Precision is perfect for livestream and social commerce and for technology and artificial intelligence, 75% for logistics and packaging, and poor for the three categories highest in the priority order: platform and governance 50%, circular economy 40%, greenwashing and trust 21%. Recall runs the other way — 100% for greenwashing and circular economy against 43% for logistics and 54% for livestream and for technology.

The cause is identifiable, and it is our priority rule rather than the patterns. Assignment runs from the most lexically specific category to the most general on the assumption that specific vocabulary is diagnostic; for three categories it is not. Trust, certification and resale appear in studies whose object is a channel or a logistics process, and because those categories sit above livestream, logistics and technology in the order, they capture records belonging elsewhere. Five records we assigned to greenwashing were coded by both coders as livestream studies; four we assigned to circular economy were coded as logistics. The categories the order harms are precisely the ones it ranks lowest.

Because the sample over-samples the small themes by design, the manual codes support a re-estimate of the thematic distribution with each stratum weighted back to its population (Table D4; 10,000 stratified bootstrap replications). Two reported figures change:

Platform and governance: 3.6% by rule, 6.2% corrected (95% CI 1.5 to 12.7).
Greenwashing and trust: 5.3% by rule, 1.1% corrected (95% CI 0.0 to 2.6).
The ordering between them reverses, with probability .97 across the bootstrap distribution. Greenwashing, not governance, is the thinnest substantive theme in the corpus. We have corrected that claim wherever it appeared.

What does not change is the finding the argument rests on. The two credibility-related themes together account for 7.3% of the corpus (95% CI 2.6 to 14.0) against 26.3% for green products and purchase behaviour — a ratio of 3.6 to one whose interval, 1.5 to 11.7, excludes parity. The field studies whether consumers respond to sustainability attributes far more than whether those attributes can be believed, and the validated coding states this more strongly than the rule-based coding did.

Tables 1 and 2 and Figure 2 are retained as the rule-based coding of all 303 records, which is deterministic and fully reproducible, and Table D4 is reported as a check on them rather than a replacement: it rests on 92 records with a sampling weight of six in the larger stratum, so its intervals for the four large themes are wide. Section 8.3 states the limitation and names the coding design that would resolve it — requiring the matched term to occupy a syntactic position indicating the study’s object, or multi-label assignment with explicit tie-breaking. The completed coding sheets, the confusion matrix and the scoring script are in the supplementary materials.

The earlier error analysis stands alongside this: 29.7% of records match two or more theme patterns under multi-label coding, which the validation now shows to be a realised rather than a potential source of error.

2.5 The 1.40 versus 12.6 comparison may reflect keyword frequency; recommend a frequency-normalised measure and a null model

This was the most consequential comment in either review, and the reviewer was right. We computed the Jaccard index, cosine coefficient, association strength [47] and pointwise mutual information for all pairs, and tested against a randomisation model holding each keyword’s document frequency fixed at its observed value over 10,000 draws. Table C4 reports the full results. The outcome does not support our original claim:

The randomisation model alone produces a ratio of 7.48 (95% interval 6.41 to 8.67) against the observed 9.06. Most of the ninefold gap is a frequency effect.
Association strength, which equals one under independence, is 0.82 for subdomain pairs against 0.96 for subdomain–core pairs — a ratio of 1.17, not nine.
Subdomain terms attach to the core at 12.64 observed against 12.79 expected (p = .32). The claim that they are specially bound to a shared abstract vocabulary is not supported and has been withdrawn.
Two results survive. Subdomain pairs co-occur less often than chance: 1.40 observed against 1.72 expected, p = .003. And 35 of 91 pairs never co-occur where 19.7 are expected, which no draw in ten thousand reproduced (p < .001).

Section 5.4 has been rewritten accordingly. The finding is now stated as avoidance rather than shared attachment:

The accurate formulation is therefore not that subdomains are bound to a shared abstract vocabulary — they are bound to it no more than chance requires — but that they avoid one another: the subdomain literatures are disconnected from each other to a degree that their individual prominence in the corpus does not explain.

Sections 7.2 and 8.1 have been amended to match. We think the revised claim is narrower and better identified than the one it replaces, and we would not have reached it without this comment.

2.6 Secondhand trading is both the mainstream benchmark and a sustainability category; a single anchor may drive the 0.7–4.1% magnitude

Accepted. The conceptual objection is sound and the sensitivity test confirms it matters. Table B4 now reports every group against two denominators. Against live commerce, a channel expression with no sustainability content, sustainability expressions range from 23% to 130% rather than 0.7% to 4.1%. The rank order of groups and the rank order of declines are identical under both; the absolute magnitude is not. Section 6.1 now states this explicitly and reports the magnitude as anchor-conditional, as do the abstract and Section 8.1.

2.7 The verification shows the probe reproduces under the authors’ own configuration, not that the anchored ratio is accurate

A fair criticism of how we described the test. Taken alone, the zero-deviation result is close to circular. There is, however, a non-circular result in the same data that we had not drawn out. The three requests differed in composition — two carried five keyword groups and one carried three — and the anchor and probe series returned by the platform were numerically identical across all three to full precision, as were their ratios. The platform’s normalisation of these series is therefore invariant to query composition, which is precisely the property the bridge requires. Section 4.4 now distinguishes the two and states the residual limit:

What the design cannot establish is agreement with an independently executed joint query containing terms that were never co-queried here; that check requires re-collection and is identified as a priority in Section 8.4.

We were unable to re-collect within the revision window, as the platform API quota and the renormalisation of indices at query time mean a fresh collection would not be comparable with the reported series. We have therefore stated the limitation rather than approximated the test.

2.8 Keyword substitutions and the ESG exclusion may introduce post-hoc selection effects

Agreed, and the ESG exclusion turns out to be consequential for one reported quantity. Table B4 now includes the ESG group throughout. Including it raises the upper end of the magnitude range from 4.1% to 13.1% of the anchor, and Section 6.1 says so. It does not affect any comparative or temporal conclusion: ESG declines by 46.6% anchor-relative with a significant monotonic trend (τ = −0.51, p < .001), behaving like the aspirational expressions rather than like greenwashing. The three pre-collection substitutions are recorded in Appendix B and in the configuration file supplied with the supplementary materials; they were made before any outcome was observed, on a non-zero-months criterion fixed in advance, and the replaced expressions returned too few non-zero months to be analysable at all.

2.9 Only four annual observations, no formal trend or uncertainty analysis; and search interest is not attitude or purchasing

Both parts accepted. On the first, we have replaced the comparison of annual point estimates with tests on the 48 monthly anchor-relative observations. Mann–Kendall tests [48] confirm a significant monotonic decline for every expression except greenwashing, which shows none (τ = 0.07, p = .49). Bootstrap intervals from 10,000 replications exclude zero for every expression except greenwashing, whose interval runs from −28.8% to +34.7%. Table B5 reports both. The central differential therefore holds on formal testing and no longer rests on comparing point estimates.

On the second, the sentence the reviewer quotes has been deleted. Section 7.1 now reads:

Search interest records information seeking, not attitudes and not purchases, and this design observes no individual moving through the sequence; what it establishes is that the aggregate search signal is small and contracting while surveyed attitudes are not, which is a discrepancy in need of explanation rather than a demonstration that attenuation occurs at a particular stage … the alternative that consumers who intend to buy sustainably simply do not express that intention as a sustainability query cannot be excluded from these data.

The "pre-search component" claim in Section 7.2 is now offered as a design to be tested rather than an estimate delivered, with the linked-data requirement stated.

2.10 Internal inconsistencies requiring cross-checking

All four are confirmed and corrected. We are grateful for the care this took.

(a) Figure 2 versus Table 2. The reviewer is right, and the error was ours: Figure 2 was plotted from a superseded coding run and never regenerated. Table 2 is correct and reproduces exactly from the coded corpus. Figure 2 has been regenerated from the Table 2 values and now also shows the greenwashing series. The figure sits directly beneath the table so that correspondence can be checked at a glance.

(b) Broker stability. Section 5.4 claimed the two principal brokers were unchanged at every threshold; the note to Table C1 said they were unchanged from the sixtieth percentile upward. The table note was correct and the text has been corrected to match. We have also added the related qualification the reviewer implies: core membership is threshold-dependent, containing 16, 12 and 13 keywords at the sixtieth, seventieth and eightieth percentiles respectively.

(c) 39–63% versus "half to two-thirds". The range is correct; the paraphrase was not. Section 7.2 now reads "two-fifths and two-thirds" and cites the exact range.

(d) Governance "appears nowhere in the network". The reviewer is right that platform is present as a node at k-core 10, and the proposed formulation is the accurate one. Sections 7.2 and 8.1 now state that governance-specific vocabulary does not reach the top-33 keyword selection, and note that platform is present but indexes the marketplace as a commercial form rather than as an object of institutional analysis, occurring in 56 records while anchoring only 11 governance-coded studies.

2.11 The strongest conclusions depend on corpus construction, keyword selection, network thresholds and anchor choice

This is the right summary of the problem, and the revision is organised around it. Each of the three headline claims has been tested against the dependency the reviewer identifies, and each has been revised to the level the test supports.

The 3.6% governance figure has been tested from both directions and re-estimated at 6.2% (95% CI 1.5 to 12.7); the companion greenwashing figure falls from 5.3% to 1.1% and the ordering between the two reverses. What survives is the comparison the argument needs: the credibility-related themes together are 7.3% against 26.3% for green products.
The "mutually invisible" subdomain claim did not survive frequency normalisation and has been replaced by a narrower one that does.
The 0.7–4.1% magnitude is now reported as conditional on the anchor, with the alternative-anchor and ESG-inclusive figures given alongside.

Section 8.3 carries three further limitations arising from the new analyses: the randomisation null does not reproduce topic and venue clustering; the anchor sensitivity tests two denominators on one platform in one market; and composition-invariance was verified within a single collection configuration. We hope the revised manuscript claims only what it can support.

We are grateful to both reviewers. Three of our reported findings changed under the checks they asked for, and we have reported each change rather than working around it. The paper is more cautious than it was and, we think, more useful for it.