Abstract
Marketing chatbots increasingly mediate customer inquiry, product recommendation, and promotional contact, yet their interaction design often leaves users unable to identify what is disclosed, what is inferred, and what promotional influence they can decline. This protocol specifies a transparency framework for marketing chatbots that treats disclosure, archiving, and user-controlled contact as separate but linked obligations. The proposed model builds on chatbot architecture and interaction logging requirements from financial customer-service systems, where query parsing, knowledge-graph completion, and layered classification support structured logging of dialogue events. It also draws on experimental disclosure research showing that transparency interventions do not automatically alter downstream behaviour, field evidence that opt-out controls can reduce exposure and harm, archival methods for preserving marketing claims, and policy work emphasising public support, local context, and evidence adequacy when regulating availability. The protocol defines a minimum disclosure set, a machine-readable archive record, and a decision sequence for classifying each chatbot turn as service, promotional, or mixed content. It proposes evaluation criteria focused on user comprehension, traceability of claims, exposure reduction after opt-out, and the proportion of prompts with auditable provenance. A prospective validation design combines simulated dialogue tests, expert review of disclosure completeness, and field deployment with randomised comparator conditions. The manuscript establishes a bounded method for making marketing chatbots more legible and contestable; it does not claim empirical efficacy, which remains to be tested.
1. Introduction
Marketing chatbots are now used not only to answer service questions but also to recommend products, route customers to offers, and maintain repeated contact across digital channels. Their utility depends on speed and consistency, yet these same features can make the origin of a recommendation difficult to inspect. When a chatbot blends service information with promotional language, users may not know whether they are receiving assistance, advertising, or a hybrid of both. A transparency protocol is therefore needed for the narrower problem of marketing chatbots: how a system should disclose its role, record the basis of its outputs, and respect user choices about promotional contact. This manuscript proposes such a protocol as a research-ready framework rather than a report of completed evaluation.
The protocol is intentionally narrow. It does not attempt to solve general chatbot quality, sentiment, or task success. Its concern is transparency in three domains: disclosure of the chatbot’s function and sponsorship, archiving of claims and dialogue events, and user control over promotional exposure. This focus follows prior work in adjacent domains where transparency problems were operationalised as information disclosure, documentary preservation, or opt-out design. Experimental work on price disclosure suggests that transparency can improve perceived information without necessarily changing downstream behaviour. Field-experimental work on direct marketing suggests that reducing exposure through opt-out can change behaviour and harm. Archiving work shows that digital preservation is essential when claims are mutable or contested. Policy research on availability further indicates that public support and evidence sufficiency matter when controlling promotional environments. These strands motivate a protocol that is technical, auditable, and behaviourally testable. [1] [2] [3] [4]
2. Evidence scope and synthesis approach
The evidence base for this manuscript is a targeted narrative synthesis, not a systematic review. The purpose is to assemble method-relevant findings from a small set of directly applicable studies and to translate them into operational requirements for a chatbot transparency protocol. The direct-application anchor is a study of an AI-powered customer-service chatbot in a marketing setting that used knowledge-graph and neural-network components and relied on large-scale logs of chatbot and customer-service interactions. That study is useful here because it makes explicit that chatbot systems can be designed around structured query parsing, relational inference, and detailed interaction logging. The other sources are transferable rather than topic-identical: one tests how price disclosures affect perceived information and shopping intentions; one argues for systematic digital archiving of online marketing claims; one uses a field experiment to test opt-out from direct marketing; one discusses policy priorities around availability, public support, and evidence; and one shows how digital marketing research benefits from theory-context-method alignment. [5] [1] [3] [2] [4] [6]
The synthesis is bounded by three assumptions. First, transparency is treated as a property of the interaction system, not as a general moral attribute of the organisation. Second, disclosure is meaningful only if users can distinguish service content from promotional content and can later inspect the basis of what the chatbot said. Third, user control over contact must reduce future exposure in a traceable manner; a control that exists only in policy language is insufficient. These assumptions constrain the protocol and make it testable.
3. Conceptual basis and prior methods
The proposed protocol draws on prior methods in four ways. First, the customer-service chatbot study demonstrates a logging-oriented architecture: user queries are transformed into a structured representation, relational information is inferred, and outputs are classified through layered processing. Although the study is framed as performance-oriented rather than transparency-oriented, it indicates that marketing chatbots can be instrumented at the level of each dialogue turn and each response component. That logging capacity is a prerequisite for any transparency standard that requires provenance, replay, or post hoc review. [5]
Second, the price-disclosure experiment provides a methodological template for separating perceived information from behavioural change. Its central lesson is that a disclosure intervention may improve what users think they know while leaving treatment-seeking or comparison-shopping intentions unchanged. For a chatbot protocol, this means that disclosure quality should not be judged only by self-reported clarity. It must also be evaluated against subsequent user actions, such as opting out, ignoring an offer, requesting a source, or escalating a complaint. [1]
Third, the digital archiving paper shows why evidence preservation must be built into the system. Marketing claims can change, disappear, or be reworded after scrutiny begins. A transparency protocol therefore needs a time-stamped repository of the chatbot’s claim text, retrieval context, and the version of the underlying prompt or policy that generated the answer. Archiving is not a postscript; it is part of the transparency mechanism itself. [3]
Fourth, the gambling direct-marketing experiment shows how exposure can be reduced by a real opt-out intervention and how that reduction can be evaluated in a field setting. The study’s logic is transferable: if promotional contact is a meaningful exposure, then opt-out is an intervention on exposure, not merely a preference setting. For marketing chatbots, the protocol should therefore track whether users can suppress promotional turns, whether that suppression persists, and whether exposure density falls after opt-out. [2]
Finally, policy work on availability indicates that transparency rules gain traction when they are aligned with public support, local context, and evidence sufficiency. The present protocol therefore avoids universal claims about all marketing chatbots and instead specifies a bounded regulatory design suitable for contexts where promotional contact is frequent and user consent is not always obvious. [4]
A recent systematic review in digital marketing research shows that field context, theory, characteristics, and method should be specified together. That lesson matters here because transparency interventions can be overgeneralised if their intended context is not defined. The protocol is therefore limited to chatbots used in marketing, customer acquisition, cross-selling, retention, or promotional servicing, and it excludes bots used only for internal operations or non-commercial support. [6]
4. Proposed transparency protocol
The protocol defines transparency as a jointly satisfied condition with three components: disclosure, archiving, and user control. Let T be the transparency status of a chatbot interaction. T = 1 only if all of the following are true: (i) before the first substantive response, the user is informed that the system is a chatbot and whether it may generate promotional content; (ii) each promotional assertion is logged with claim text, source basis, timestamp, and model or prompt version; and (iii) the user can decline further promotional contact through an opt-out path whose burden does not exceed the burden of opting in. For audit purposes, burden is measured by a prespecified tuple B = (S, A, R), where S is the number of user action steps after the disclosure screen, A is the number of additional authentication events beyond those already required for the session, and R is the elapsed completion time under a standard usability script. The default compliance rule is B_opt-out <= B_opt-in on all three elements, with R assessed separately as the median within each fixed channel, authentication-state and identifier-class script cell, never pooled across unlike cells. If any component fails, T = 0 for that interaction. This is an operational rule that can be tested across channels.
The protocol requires a disclosure block before the first substantive response. The block contains four fields: system identity, commercial role, claim provenance, and user control. System identity states that the user is interacting with an automated system. Commercial role states whether the bot is serving only support, only promotion, or both. Claim provenance states whether the upcoming response is based on a policy document, product catalogue, conversation history, or inferred preference. User control states how to stop promotional messages, request a human agent, or request a non-promotional mode. The wording may vary by channel, but the fields may not be omitted.
The archiving module stores each conversational event as a machine-readable record. Minimum fields are: interaction identifier, user-initiated or system-initiated status, channel, timestamp, disclosure text shown, user response to disclosure, response text, promotional flag, source document hash, policy version, and opt-out status. For claims that involve prices, availability, or product features, the system also stores the underlying evidence artifact, such as the product page snapshot or policy excerpt that justified the statement. This design is intended to prevent later disputes about what was said and on what basis. [3]
The opt-out rule applies at the level of promotional exposure. Once a user opts out, the chatbot must switch to a non-promotional mode within the same session and across subsequent sessions linked to the same persistent identifier unless the user later re-authorises promotional contact. The default identifier hierarchy is: authenticated account ID when present; otherwise a verified channel address or number used for contact; otherwise a first-party device or browser token stored solely for preference persistence. If no persistent identifier is available, the system may enforce opt-out only for the current session, but it must disclose that limitation at the moment of opt-out and must not claim cross-session persistence. Non-promotional mode may still answer service questions, but it must not introduce discounts, urgency framing, cross-sells, or personalised offers. If the system cannot technically suppress all promotional content for the applicable identifier class, it must block the promotional function and route the user to a non-promotional service path. For validation, persistence should be tested separately for each identifier class so that anonymous sessions, shared devices, and cleared browser storage are not misclassified as protocol failures when cross-session linkage is unavailable. [2]
The protocol also requires separation between service and promotion when outputs are mixed. A response is classified as mixed if it contains both a service answer and a promotional appeal. Mixed responses are permitted only if the promotional segment is clearly demarcated and separately suppressible. For example, a chatbot may explain an account issue and then ask whether the user wants information about a related product; it may not bury that invitation inside the service answer. This separation is designed to make the promotional purpose inspectable by users and auditors.
5. Variables, comparators, and operational measures
The primary experiment randomizes users, with all linked sessions assigned consistently, to three nested arms: role disclosure alone; the same disclosure plus user-accessible archival traceability; and both features plus the proposed enhanced opt-out path. Every arm retains legally and operationally required user controls. The primary outcome is comprehension under original user assignment. Arm two versus one estimates the incremental traceability package conditional on disclosure; arm three versus two estimates enhanced control conditional on both preceding features. These are not standalone factorial effects of each component. Use the frozen user-cluster design and report all assigned users, including delivery failures. A generic bot-label comparator may be studied separately, without changing the registered primary family.
Primary outcomes are defined as follows. Disclosure comprehension is the proportion of users who can correctly identify the chatbot’s role, the existence of promotional content, and the route to opt out after a single interaction. Claim traceability is the proportion of promotional assertions that can be matched to an archived source artifact within the system’s repository. Exposure reduction is the change in the frequency of promotional turns per user after opt-out relative to the pre-opt-out baseline, estimated only within the same identifier class used to enforce persistence. Behavioural response is measured as the proportion of users who request a human agent, seek more information, or refuse a suggested offer after the disclosure appears. To evaluate the burden-equivalence rule, the audit must also record opt-in and opt-out completion burden as the tuple B = (S, A, R), where S is post-disclosure action steps, A is additional authentication events, and R is elapsed completion time under the preregistered usability script. These outcomes are distinct: comprehension measures understanding, traceability measures documentary fidelity, exposure reduction measures the system’s contact behaviour, and burden equivalence measures whether user control is practically symmetrical with acceptance. [1] [3] [2]
A useful secondary measure is disclosure salience, defined as the share of sessions in which the user actively acknowledges the transparency notice before receiving promotional content. This can be observed through click-through, typed acknowledgement, or continued interaction after the notice. However, salience is not equivalent to understanding. The protocol therefore treats it as a process indicator only.
The protocol should also track failure cases. A failure occurs if promotional content is generated without an archived basis, if a user who opted out later receives promotional messages, if the system presents a promotional recommendation as a neutral service answer, or if the disclosure text is buried in a manner that is unlikely to be seen before the offer. These failure conditions are important because they indicate where technical compliance and user experience diverge.
6. Validation design
Validation should proceed in three stages. Stage 1 is a controlled simulation test using scripted conversations that vary by intent, promotional content, user identity state, and user response. Identity state should include at least authenticated account, verified contact channel without login, anonymous first-party browser token, and no persistent identifier. The goal is to verify whether the decision sequence correctly classifies turns, attaches evidence records, enforces the burden-equivalence rule, and suppresses promotional follow-up after opt-out for each identifier class. Stage 2 is expert review by reviewers trained in consumer protection, marketing law, and dialogue systems. Their task is to assess whether the disclosure block is complete, whether mixed responses are clearly demarcated, whether archived records are sufficient to reconstruct the interaction, and whether opt-out limitations are accurately disclosed when cross-session persistence is unavailable. Stage 3 is a field deployment with random assignment to comparator conditions.
In the field stage, the main test is whether the protocol changes actual interaction behaviour relative to the comparator. A feasible design is a parallel-group evaluation in which new users are assigned either to the standard chatbot or to the transparency protocol. If the deployment permits user-level opt-out, the protocol should additionally examine within-user change before and after opt-out. The primary hypothesis is not that transparency will increase sales or reduce sales, but that it will improve explainability and reduce unsolicited promotional exposure without breaking service continuity. [1] [2]
Because transparency can be attractive while being behaviourally inert, the evaluation should avoid relying on self-report alone. A user may say the disclosure is clear but still be unable to recall the promotional role or the opt-out route. Likewise, a system may allow opt-out yet continue to produce promotional prompts through hidden variables or downstream templates. The design therefore combines user comprehension tests, audit logs, and exposure counts. This separation follows the logic of prior transparency experiments, where perceived sufficiency of information was distinct from behaviour, and the logic of field experiments, where actual exposure change was the causal target. [1] [2]
Generalisation should be bounded. The protocol is most appropriate where a chatbot is used in customer acquisition, cross-selling, or retention, and where promotional content is generated or curated by the system. It is less directly applicable to transactional service bots that never present offers and to high-stakes advice systems with separate regulatory duties. For that reason, the protocol should be treated as a modular standard for marketing contexts rather than a universal chatbot rule. [5] [6]
7. Boundary conditions and limitations
The main limitation is that transparency does not guarantee user welfare. Users may understand that a message is promotional and still accept it, ignore it, or prefer it. Disclosure can improve accountability without altering behaviour, and that outcome should not be treated as failure. The protocol is designed to make the system legible and contestable, not to predict consumer choice. [1]
A second limitation concerns archiving scope. Perfect preservation is unrealistic because legal constraints, privacy requirements, and platform changes may limit what can be stored. The protocol therefore distinguishes between full content archiving, hashed evidence anchoring, and metadata-only logging. A deployment may satisfy the protocol if it preserves enough information to reconstruct the claim and its basis, even when some user content must be redacted. [3]
A third limitation is context dependence. Public tolerance for promotional contact, the legal status of direct marketing, and the expectations attached to customer-service chatbots vary across sectors and jurisdictions. The policy literature suggests that availability and marketing controls are easier to justify when the public can see the harm and the evidence is specific. This implies that the protocol should be calibrated to local regulatory conditions and should not be assumed appropriate wherever chatbots are used. [4]
A fourth limitation is technical drift. If the dialogue model is updated, the archive must preserve version identity; otherwise the system may become impossible to audit across time. This is particularly important when the architecture includes retrieval components, knowledge graphs, or post-processing layers, because a change in any layer may alter the meaning of the same surface response. [5] [3]
Matched burden cells and service separation
Compare opt-in and opt-out burdens only within the same channel, authentication state and identifier class, using the same script and timing procedure. Apply the componentwise gate in every eligible cell and report cell coverage. If no comparable opt-in path exists, relative burden is unavailable; use a separately registered absolute opt-out limit and do not claim equivalence. A promotional segment must have both a visible and programmatic label and be suppressible while leaving the service answer unchanged. A snapshot-diff test checks this property on paired scripted responses. Cross-identifier contamination is recorded for sensitivity analysis rather than removed from the intention-to-treat population. Pharma, gambling, stem-cell marketing and tobacco studies motivate disclosure, archiving and user-control questions in their own settings; they do not demonstrate that this chatbot protocol improves comprehension or reduces harm. Those outcomes require the prospective experiment above.
8. Conclusion
This manuscript establishes a bounded conceptual contribution: a transparency protocol for marketing chatbots that separates disclosure, archival traceability, and user-controlled promotional contact into operational requirements. The protocol is grounded in structured chatbot logging practices, disclosure research, digital archiving methods, opt-out field evidence, and policy reasoning about evidence sufficiency and public support. It provides a concrete decision sequence, a mathematical compliance rule, comparator conditions, and prospective measures for evaluation. [5] [1] [3] [2] [4] [6]
What remains untested is whether this protocol, when implemented in live marketing systems, improves user comprehension, reduces unwanted exposure, and sustains service quality under routine commercial use. Those questions require a field evaluation with preserved logs and clearly defined comparator conditions. Until then, the protocol should be read as a technical standard for making marketing chatbots more inspectable and more governable, not as evidence that it already achieves those aims.
Contribution and availability
This manuscript proposes a bounded transparency protocol for marketing chatbots that unifies disclosure, archival traceability, and user-controlled promotional contact into a single auditable framework. The contribution is methodological: it translates prior work on chatbot logging, disclosure experiments, digital archiving, opt-out exposure reduction, and policy support into an explicit decision rule, operational variables, comparator conditions, and a prospective validation design tailored to marketing chatbots.
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
- Hollin, Ilene L.; Ball, Jennifer Gerard (2022). Does price disclosure in pharmaceutical advertising result in price transparency? Evidence from a randomized experiment. Exploratory Research in Clinical and Social Pharmacy. 8, 100180. doi:10.1016/j.rcsop.2022.100180.
- Rockloff, Matthew; Browne, Matthew; Hing, Nerilee; Russell, Alex M. T.; Rawat, Vijay; Newall, Philip (2026). Direct gambling marketing, direct harm: A randomised experiment. Addiction. 121(7), 1907-1919. doi:10.1111/add.70369.
- 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.
- Dimova, Elena D.; Shortt, Niamh K.; Mitchell, Richard J.; Lekkas, Peter; Pearce, Jamie R.; Clemens, Tom L.; Emslie, Carol (2025). “Availability is the poor cousin of marketing and pricing”: qualitative study of stakeholders’ views on policy priorities around tobacco and alcohol availability. Drugs: Education, Prevention and Policy. 32(1), 51-62. doi:10.1080/09687637.2023.2282355.
- He, Ji; Luo, Yi; Wang, Ting (2025). iDigiChat: intelligent digital marketing service chatbot for providing efficient customer services using artificial intelligence. Scientific Reports. 15(1), 33074. doi:10.1038/s41598-025-14722-5.
- Vardikou, Chrysopigi; Konidaris, Agisilaos; Koustoumpardi, Erato; Kavoura, Androniki (2025). Emojis in Marketing and Advertising: A Systematic Literature Review. Behavioral Sciences. 15(11), 1490. doi:10.3390/bs15111490.