Abstract
Consumer-facing technology claims often rely on credibility cues that can move acceptance ahead of evidence. This protocol defines a targeted narrative synthesis and an auditable analytic workflow for evaluating trust signals in product claims, with emphasis on claims about digital, AI-enabled, and review-mediated products rather than technical performance itself. The problem is conceptually similar to prior work on direct-to-consumer marketing of unproven stem cell interventions, where persuasive “tokens of legitimacy” can make weak claims appear evidence-based, and to influencer marketing research showing that trust, authenticity, content quality, and word-of-mouth shape consumer perception. The protocol proposes a two-level framework: first, a claim-level coding scheme that distinguishes evidence claims, authority cues, social proof, and verifiability markers; second, a computation-assisted audit that combines explainable sentiment analysis, review-based recommendation signals, and human–machine comparison to estimate how trust cues alter perceived credibility. The central operational idea is that credibility should be assessed as a difference between what a claim asserts and what its supporting evidence permits. We define a simple rule for audit purposes: Trust Signal Excess = asserted certainty score minus evidence support score. The article specifies variables, a decision sequence, comparator conditions, validation design, failure cases, and limits of generalization. The contribution is a bounded, testable protocol for studying when trust signals legitimate informative claims and when they outpace evidence in consumer technology marketing.
1. Introduction and research question
Technology product markets increasingly depend on short claims that present a system, app, device, or service as safe, effective, intelligent, or widely trusted. The claims are rarely neutral descriptions. They often contain trust signals: references to experts, user counts, ratings, awards, media mentions, scientific language, platform badges, or selective testimonials. The research question guiding this protocol is: how can we evaluate whether such signals increase consumer acceptance of a technology product claim beyond the support actually provided by the claim’s evidence? The question is narrower than technical performance evaluation. It concerns consumer-facing credibility assessment, especially where promotional language may precede or exceed substantiation. [1] [2]
The topic matters because previous work on direct-to-consumer stem cell marketing shows that persuasive legitimacy markers can be used to make unsupported interventions appear credible, trustworthy, and scientifically grounded. Related consumer studies on influencer marketing report that trust, content quality, authenticity, and word-of-mouth shape perception and purchase intention. These strands motivate a protocol focused not on whether a product works, but on how claim language, platform presentation, and social cues affect belief formation. [1] [2]
2. Scope of evidence and narrative synthesis strategy
This article uses a targeted narrative synthesis rather than a systematic review. The evidence scope is limited to three overlapping bodies of work. First, direct application evidence on persuasive marketing claims and credibility cues, including studies of problematic direct-to-consumer claims. Second, consumer response research on trust, authenticity, and purchase-related perception in digitally mediated settings. Third, methodological work that is transferable at the level of procedure rather than substantive domain: explainable sentiment analysis for interpretable text classification, sentiment analysis of product reviews as a source of structured textual signals, and human–machine comparison designs for contrasting expert, algorithmic, and hybrid judgments. The purpose is not to estimate pooled effects or to claim comprehensive coverage, but to define a methodically explicit research program. [1] [2] [3] [4] [5]
The synthesis logic is functional: each source is used only where the cited passage supports a transferable step in the protocol. Digital archiving work supports preserving the exact wording and context of online claims. Explainable sentiment analysis supports the requirement that automated classifications should expose reasons rather than output labels alone. Sentiment analysis of review corpora and sentiment-aware recommendation studies support the narrower point that review text can be processed as structured textual data, although the present manuscript does not treat those studies as direct validation of credibility auditing. Human–machine comparison studies support using expert-only, machine-only, and hybrid conditions as comparator designs, while leaving this application to be tested prospectively. [1] [3] [6] [4] [5]
3. Conceptual basis: what counts as a trust signal
A trust signal is any feature of a product claim that is designed, or plausibly functions, to increase perceived credibility without itself constituting evidence for the underlying performance claim. In this protocol, trust signals are separated into four classes. Evidence-signaling cues include references to studies, trials, regulatory approval, or data summaries. Authority cues include expert titles, institutional names, or scientific terminology. Social proof cues include testimonials, ratings, follower counts, usage counts, and “most trusted” language. Verifiability cues include links, citations, timestamps, archivable pages, and reproducible method descriptions. These classes are analytically distinct because a claim can contain several trust signals while providing little support for the stated outcome. [1] [2]
The distinction matters because the presence of a cue does not establish evidential adequacy. The stem cell marketing literature describes “tokens of legitimacy” used to make advertised interventions seem credible and backed by current scientific evidence, even when robust support is lacking. Influencer marketing research likewise indicates that trust and content quality shape perception, meaning that a persuasive cue can change consumer response even when the informational basis is uncertain. This protocol therefore treats trust signals as measurable text features rather than as proof of reliability. [1] [2]
4. Proposed operational definitions and variables
The protocol defines three primary variables at the claim level. Let c be a claim and let E(c) be the evidence presented with that claim. The audit uses explicit integer-coded ordinal rubrics. Their subtraction defines a proposed triage rule, not a measured interval-scale quantity or proof that category distances are equal. Asserted Certainty Score A(c) is coded on a 0-4 scale: 0 no performance or safety assertion; 1 possibility or tentative benefit; 2 moderate assertion with qualification; 3 strong assertion with limited qualification; 4 categorical or superiority claim stated as fact. Evidence Support Score S(c) is coded on a parallel 0-4 scale: 0 no supporting evidence; 1 vague reference without accessible detail; 2 some accessible detail but weak direct support; 3 direct and checkable support with stated scope or limits; 4 direct, checkable, and claim-matched support with material constraints disclosed. Trust Signal Load T(c) is a normalized cue index defined as T(c)=(w_e n_e+w_a n_a+w_s n_s+w_v n_v)/m(c), where n_e, n_a, n_s, and n_v are counts of evidence-signaling, authority, social-proof, and verifiability cues, m(c) is the number of claim units in the audited page section with a minimum value of 1, and default weights are w_e=1, w_a=1, w_s=1, w_v=1. For reproducibility, each cue token is counted once at its smallest visible unit of presentation: a sentence-level cue is assigned to the sentence in which it appears; a cue in a shared heading, badge row, sidebar, or testimonial block that applies to several claim units is counted once in section-level totals and not repeated for each claim unit; duplicated tokens with the same wording and function within the same audited section are counted once; and near-duplicate cues are merged only when a coder can identify no substantive difference in wording, source, or informational content. The primary audit quantity is then X(c)=A(c)-S(c), which ranges from -4 to 4 under this fixed scoring rule. The default decision rule is to flag a claim when X(c)>=2. Sensitivity analysis should then re-estimate classifications under two preregistered alternatives, X(c)>=1 and X(c)>=3, without changing the coding rubric. [1] [3]
A second rule is used when the audited page includes testimonial or review text. Testimonial text is any endorsement selected or reproduced by the seller on the product page; customer review text is any review displayed through a platform review interface or linked review feed. The protocol first labels each item by source type, removes exact duplicates, and treats near-duplicates with the same rating and near-identical wording as one item. For each claim c, Review Sentiment Direction R(c) is the majority sentiment label across linked customer reviews within the archived page context, with ties recorded deterministically as mixed. Review Evidence Density D(c) is the mean item-level specificity score on a 0-2 scale, where 0 denotes vague praise or dislike without checkable detail, 1 denotes one concrete but limited product-specific detail, and 2 denotes multiple concrete and checkable details about use, function, conditions, or failure. If no customer reviews are present, R(c) and D(c) are coded missing and do not enter the primary flagging rule; seller-selected testimonials are instead counted only as social-proof cues within T(c). When R(c)=mixed, the value is carried into the audit report as a descriptive review-context indicator only; it does not increase or decrease X(c), does not change the binary flagging threshold, and is not converted into a numeric score. This deterministic handling rule separates the primary evidence-gap decision from heterogeneous consumer sentiment while preserving review context for later interpretation. The methodological aim is to distinguish sentiment intensity from informative density, because favorable language can function as a trust signal even when it adds little evidence. [6] [4] [3]
5. Proposed analytic workflow
The analytic workflow proceeds in seven steps. Step 1 archives the target claim in its original web context, including screenshots, page source, timestamp, and linked testimonials. This step follows the logic of digital archiving used in research on direct-to-consumer stem cell marketing, where preserving exact wording is necessary because web pages can be changed or removed. Step 2 segments the page into claim units, each unit being a sentence or proposition that asserts an effect, benefit, comparison, or endorsement. Step 3 annotates each unit for A(c), S(c), and T(c) using a codebook with examples and exclusion rules. Step 4 extracts lexical and rhetorical markers, including certainty language, scientific labels, citation forms, superlatives, and social-proof tokens. [1]
Step 5 applies explainable sentiment analysis to customer reviews and seller-selected testimonial passages attached to the product claim. The aim is not to infer product quality directly, but to identify whether the page’s accompanying discourse is emotionally favorable, ambiguous, or cautionary, and whether the model can expose the textual basis of its label. Explainable sentiment analysis is relevant here because it demonstrates how classification can be paired with explicit reasoning rather than opaque output alone. Step 6 compares claim-level trust signal load with the review-derived profile to determine whether the page relies on favorable affect or social endorsement while providing limited claim-matched support. In this comparison, mixed review sentiment is reported descriptively and never used to override the primary flagging rule based on X(c). Step 7 generates an audit report that flags claims with high trust signal excess, strong social proof but weak support, or inconsistent combinations of evidence and assertion. [3] [6]
6. Comparator design and human–machine comparison
The comparator is central to the protocol. There are three conditions. In the human-only condition, trained coders assign A(c), S(c), T(c), and any review variables using the codebook. In the machine-only condition, the system proposes the same variables from prespecified text features and outputs an explanation trace for each proposed score. In the hybrid condition, the machine produces an initial audit, after which a human reviewer may accept or revise each component score and must record the textual basis for any revision. This design follows the comparative logic of expert, machine, and hybrid assessment used in human-machine research, while applying it here to credibility auditing as a prospective extension rather than as an already validated transfer. [5] [3]
The comparator is not intended to rank humans against machines on a single absolute scale. Instead, it tests whether the machine improves consistency, whether human reviewers correct systematic overcalling of trust cues, and whether the hybrid workflow identifies ambiguous claims better than either component alone. This approach also differs from recommender-system methods because the object is not to predict preference. Review-based recommendation research is transferable only in the sense that it shows how textual reviews can be combined with structured signals. In this protocol, recommendation-style ranking is replaced by credibility auditing. [4]
7. Validation design and outcome measures
Validation is prospective and multi-stage. The primary outcome is flag agreement: the degree to which human-only, machine-only, and hybrid assessments converge on the binary decision produced by the default rule X(c)>=2. Secondary outcomes are component-score agreement for A(c) and S(c), explanation adequacy, and archive-format stability. Explanation adequacy asks whether a reviewer can trace each assigned score to identifiable text evidence rather than to general impression. Archive-format stability asks whether the same claim receives materially similar scores when assessed in its original web context, in a screenshot archive, and in stripped text form. This matters because marketing pages often change over time, and the protocol must remain usable after modification or deletion. [1] [3]
Validation should also examine false positive and false negative failure cases. A false positive occurs when the rule flags a claim despite direct, accessible, and claim-matched support that the coding procedure failed to recognize. A false negative occurs when a polished but thin claim is not flagged because persuasive wording or favorable testimonials obscure low support. Domain portability should be tested by preregistered stratified sampling across at least three consumer technology categories, such as software subscriptions, consumer electronics, and AI-enabled services, with equal target counts per category selected before coding begins. If category-specific conventions require adaptation, only the examples in the codebook may be revised; the 0-4 scoring rubric and the default threshold X(c)>=2 should remain fixed for the main analysis. The article does not claim these tests have been completed; it specifies them as required validation steps. [1] [2] [3]
8. Boundary conditions, conflicting evidence, and limits
The protocol is designed for claims in which trust signals are visible in the marketing text itself. It is less suitable for hidden recommendation systems, offline sales scripts, or claims embedded in video where text extraction is incomplete. It is also limited when evidence is distributed across long documents, because a local claim may appear unsupported while a later appendix provides qualification. In such cases, the unit of analysis must expand from sentence to page or campaign. The protocol also avoids assuming that all trust signals are manipulative. Some cues are legitimate and necessary, especially when they identify methods, limitations, or regulatory status. [1] [5]
There is also a boundary condition in consumer response research. Influencer studies indicate that trust, authenticity, and content quality can positively shape perception, but they do not establish that every trust cue reduces judgment quality. In some settings, trust signals may be informative because they help consumers navigate complex offerings. The protocol therefore distinguishes persuasive trust signals from informational trust signals by asking whether the cue adds verifiable substance or merely intensifies assurance. This distinction keeps the analysis open to mixed evidence rather than presuming that every strong claim is deceptive. [2]
9. Discussion: relation to prior methods
The proposed method draws selectively from prior work but changes the object of analysis. Digital archiving research contributes the preservation principle: claims must be stored in their original form before analysis. Explainable sentiment analysis contributes the interpretability principle: classification should expose reasons, not only labels. Review-based recommender research contributes the idea that consumer text contains structured signals that can be modeled alongside metadata. Human–machine comparison contributes the evaluation principle: a method should be tested against expert judgment and a hybrid workflow, not only against itself. Combined, these foundations support a credibility audit for technology product claims that is both text-sensitive and operationally explicit. [1] [3] [4] [5]
What this protocol does not do is infer product truth from persuasive style. That would collapse credibility into rhetoric. Instead, it treats trust signals as a measurable part of marketing discourse and evaluates whether they are proportionate to the evidence presented. This bounded contribution is relevant to consumer protection, platform governance, and advertising research because it offers a way to mark claims that are linguistically convincing but evidentially thin. [1] [2]
Atomic evidence matching and bounded trust inference
Split each independent predicate by product, time and claimed effect before scoring A and S. Shared evidence attaches only when it explicitly supports that predicate. A compound parent receives a flag if any atomic child has X>=2; supported neighbors cannot offset an unsupported atom. T is descriptive cue load, not a calibrated probability of credibility. For review context, freeze the product identifier, the 90-day pre-capture window and a maximum N per feed; select the latest eligible reviews in timestamp and immutable-ID order. Seller-selected testimonials and independent feeds remain separate. Exclude unrelated products and undisclosed affiliate substitutions; absent or inaccessible reviews stay missing, never positive by default. Report percent agreement, Cohen kappa and coding coverage for two independent coders, while acknowledging prevalence sensitivity. Thresholds are proposed rubric settings, not established diagnostic cutoffs. Every evidence-support decision links the atom to a captured artifact and location. Archive stability means the same claim can be traced across preserved formats, not that the claim is true. Versioned category examples retain the same A/S anchors. No score establishes fraud, legal noncompliance or verified technological performance.
10. Conclusion
This manuscript establishes a conceptual framework for evaluating trust signals in technology product claims and proposes a precise audit protocol for future study. Conceptually, it distinguishes claim strength, evidential support, and trust-signal load, and it links these variables to a human–machine comparison design that can be implemented in text-based credibility audits. It also adapts explainable sentiment analysis and review-based modeling to a new task: not recommendation or market prediction, but credibility assessment of consumer-facing claims. [3] [4] [5]
What remains untested is the empirical performance of the protocol: inter-coder reliability, machine–human agreement, robustness across product categories, and sensitivity to claims that mix valid evidence with persuasive excess. Those questions require prospective evaluation. Until then, the value of the protocol is methodological rather than evidentiary: it provides an explicit way to study when technology marketing signals inform consumers and when they may outpace the support they appear to represent. [1] [2]
Contribution and availability
The manuscript proposes a bounded, auditable protocol for credibility assessment of consumer-facing technology product claims. Its core contribution is an operational distinction between trust signals and evidence support, implemented through claim archiving, explainable sentiment analysis, review-informed signal extraction, and human–machine comparison. The protocol is designed to identify when persuasive marketing language exceeds the evidential basis presented to consumers, without treating every trust cue as deceptive or every strong claim as false.
Funding: Roman Vlasov’s own funds. Editorial responsibility and competing interests: Roman Vlasov is the founder and Editor-in-Chief of Smart World Journal and the creator of the BILDRY construction-planning platform. Other competing interests have not been declared.
This conceptual study reports no newly collected empirical data. Its proposed procedures require the prospective validation described above.
References
- Turner, Leigh (2026). Digital Archiving of Direct-to-Consumer Marketing Claims Made by Businesses Selling Unproven Stem Cell Interventions. Journal of Medical Internet Research. 28, e95156-e95156. doi:10.2196/95156.
- Babu, Md. Asaduzzaman; Urmi, Hafsa Akter; Tofayel, Abzal Hosen; Sabuj, Md. Mehedul Islam; Alam, Md Tamjid Ul (2025). The impact of influencer marketing in the tourism industry: A digital marketing perspective. PLOS One. 20(12), e0338423. doi:10.1371/journal.pone.0338423.
- Zhao, Li; Zhou, Jinwei; Cao, Jinde; Zhu, Weina (2025). EMSA: Explainable multilingual sentiment analysis models providing sentiment analysis across multiple languages. PLOS One. 20(11), e0333508. doi:10.1371/journal.pone.0333508.
- Kollu, Siva Rama Prasad; Garapati, Yugandhar (2026). DeepSentRec: a deep learning-based sentiment-aware product recommendation system. Scientific Reports. 16(1), 10580. doi:10.1038/s41598-026-45953-9.
- Kmen, Christopher; Navratil, Gerhard; Kattenbeck, Markus; Giannopoulos, Ioannis (2026). Evaluating human–machine collaboration through a comparative analysis of experts, machine learning, and hybrid approaches in real estate valuation. Scientific Reports. 16(1), 4044. doi:10.1038/s41598-025-34099-9.
- Ayman, Umme; Akash, Md. Tanvir Ahmed; Akhter, Taslima; Mridul, Saiham Zaman; Sutradhar, Yadab (2026). BanglaEcomReviewCorpus: A dataset for e-commerce product review sentiment analysis. Data in Brief. 66, 112663. doi:10.1016/j.dib.2026.112663.