Accessibility Evaluation of AI-Assisted Marketing Interfaces: A Conceptual Protocol for Comprehension, Interaction Quality, and Evidence-Constrained Review

1. Introduction

AI-assisted marketing interfaces now occupy a practical position between customer service, product discovery, and persuasive communication. In banking and related service environments, chatbot systems are used to answer queries, support transactions, and organize information at scale, while domain-aware automation frameworks increasingly aim to make AI-supported analytical workflows more accurate and easier to configure for business stakeholders [He et al., Tian et al.]. In parallel, adoption research in medical management and pharmaceutical marketing reports that perceived usefulness, organizational readiness, and regulatory literacy are associated with AI adoption intention in non-clinical work settings [Boruga et al.]. These strands justify treating AI-assisted marketing interfaces as a distinct evaluation setting, but they do not by themselves establish accessibility criteria for end users. For that reason, the present manuscript uses them only to motivate scope and method design, not as direct evidence that accessibility has been achieved or measured in those systems. [1] [2] [3]

This manuscript addresses a narrower question: how should accessibility be evaluated when the interface itself is AI-assisted and marketing-oriented? The problem differs from ordinary web accessibility because these systems often produce adaptive language, infer intent from partial input, and present answers that mix advice, promotion, service options, and compliance constraints. The result is a dynamic interface in which comprehension depends not only on visual presentation but also on message framing, state transparency, and the recoverability of conversational paths. Here, accessibility is narrower than overall usability and distinct from adoption intention: it concerns whether users can understand the interface state, identify available actions, and complete or repair the interaction without hidden knowledge. Prior review work in digital marketing has concentrated heavily on short-term engagement and purchase-oriented outcomes, while giving limited attention to field settings and ecologically valid evaluation. That pattern is drawn from adjacent marketing research as a methodological analogy rather than as direct evidence about AI-assisted marketing interfaces themselves. This protocol responds to that measurement gap by defining a measurable accessibility construct for AI-assisted marketing interfaces and by specifying how it can be tested without claiming empirical completion. [4]

2. Evidence Scope and Research Question

The evidence base for this protocol is a targeted narrative synthesis rather than a systematic review. The aim is not to enumerate all studies on AI or marketing communication, but to identify methodological elements that can support an accessibility evaluation design. Direct application evidence is limited to AI-mediated marketing and customer-service interfaces, including chatbot-centered service environments and marketing-related automation contexts in which users encounter AI-generated interaction states [He et al.]. Other sources are used only as bounded methodological or contextual supports. Adoption research clarifies why organizations may accept AI tools, not whether those tools are accessible to end users [Boruga et al.]. Review work on emojis in digital marketing documents strong attention to short-term outcomes and limited use of field experiments in an adjacent communication domain, which serves here as a caution about ecological validity rather than as direct evidence about AI-assisted interfaces [Vardikou et al.]. Evidence-structured AI assessment illustrates how a complex system can be decomposed into observable states and decision rules, which the present protocol adapts by analogy [Jing et al.]. [1] [3] [4] [5]

The research question is: how can accessibility of AI-assisted marketing interfaces be defined and evaluated in a way that makes user comprehension and interaction quality observable, comparable, and falsifiable? Two boundary conditions are built into the question. First, the protocol concerns interfaces that mediate marketing or customer-service tasks, not general consumer websites or clinical decision support. Second, accessibility is treated as a property of the interaction between user, interface, and task state, not as a synonym for overall usefulness, trust, or purchase intention. This distinction is important because adoption intention studies may identify correlates of organizational acceptance, but they do not by themselves establish whether users can understand or complete the interaction [Boruga et al.]. [3]

3. Conceptual Basis for Accessibility in AI-Assisted Marketing Interfaces

For the present protocol, accessibility is defined as the extent to which a user can perceive the interface’s communicative intent, infer the available next action, complete the intended task, and detect or repair misunderstanding without depending on hidden system knowledge. This definition incorporates four observable dimensions. Comprehension concerns whether the user can explain what the interface is offering or requesting. Navigability concerns whether the user can identify the next actionable control or utterance. Recoverability concerns whether the interface provides a usable path after an error, ambiguity, or unsupported response. Referential clarity concerns whether the interface’s claims, recommendations, and disclosures remain distinguishable and traceable to visible conditions or controls. In this protocol, accessibility is the overarching interaction property; comprehension and interaction quality are component constructs within it, rather than interchangeable labels. These dimensions are motivated by practical difficulties reported in AI-mediated customer service, where systems must translate uncertain queries into structured action, and by evidence-constrained AI assessment, where system judgment depends on explicit state representation and traceable checks. The latter source is used as a formal analogy, not as prior accessibility evidence [He et al., Jing et al.]. [1] [5]

Marketing interfaces require an additional constraint: persuasive language must not remove the user’s ability to understand system state. In a conversational marketing setting, this means that a recommendation, upsell, or personalized offer should not be merged with a service response in a way that obscures whether the system is informing, persuading, or executing an action. This requirement is not derived directly from prior accessibility studies of marketing AI. Instead, it is adopted as a protocol rule and is methodologically consistent with adjacent digital marketing review evidence showing substantial emphasis on short-term outcomes and limited use of field experiments, both of which can weaken ecological validity [Vardikou et al.]. The protocol therefore treats accessibility as a precondition for interpretable interaction, not as a downstream effect of successful marketing. [4]

4. Proposed Operational Variables and Exact Decision Rules

The proposed evaluation uses seven primary variables. U is user comprehension score; N is navigability score; R is recoverability score; C is referential clarity score; CI is the composite comprehension index; H is harmful ambiguity; and A is accessibility pass/fail at task level. Each task is scored on a 0-2 ordinal scale for U, N, R, and C, where 0 means failure, 1 means partial success with hesitation or limited prompting, and 2 means independent success. For a task t, the total accessibility score is S(t)=U(t)+N(t)+R(t)+C(t), so S(t) ranges from 0 to 8. The separate comprehension index is defined as CI(t)=(U(t)+C(t))/4, which rescales the two comprehension-facing dimensions to the unit interval from 0 to 1; higher values indicate clearer understanding of message purpose and claim meaning. Harmful ambiguity is coded H(t)=1 only when at least one of the following preregistered conditions is met: the interface presents multiple plausible next actions without labeling which one advances the stated task; the response mixes service information with promotional content such that the user cannot determine whether an action has been executed; or a material claim, disclosure, or recommendation cannot be linked to a visible source, condition, or actionable control after one ordinary clarification turn. Otherwise H(t)=0. The default operational rule is preregistered as follows: A(t)=1 only if H(t)=0, S(t)>=6, and no individual dimension equals 0. This rule prevents a high total score from masking complete failure on a single essential dimension. Before any empirical use, the study team should preregister a sensitivity analysis that recalculates pass rates under two adjacent thresholds, S(t)>=5 and S(t)>=7, while keeping the H(t)=0 and no-zero-dimension conditions fixed.

The exact operational rule is: a task is accessible only if the user independently identifies the interface purpose, the next action, and the meaning of the response state, and if any error state offers a visible, non-ambiguous route to continue or exit. In plain text: if a user can perform the task after one normal prompt and no external explanation, the task is provisionally accessible; if the user requires two or more clarification turns, or if the system response conceals whether the user has progressed, the task is not accessible. This rule is intentionally strict because conversational interfaces can appear responsive while still forcing the user to infer hidden structure. The rule mirrors the logic used in structured AI assessment frameworks, where each factual element must be retrievable and verifiable rather than merely implied [Jing et al.]. [5]

A second variable set captures interface context: prompt transparency, presence of explicit control labels, disclosure of AI status, and availability of an accessible fallback. These are coded dichotomously because the aim is to document structural conditions that may explain low comprehension. Where an interface offers both free-text dialogue and button-based shortcuts, the protocol treats the two pathways as separate accessibility channels and compares them within the same task. This within-task separation avoids attributing accessibility to the interface as a whole when success depends on one pathway that compensates for failures in the other.

5. Proposed Evaluation Methodology

The protocol uses a three-stage assessment sequence. Stage 1 is task decomposition. The evaluator defines representative marketing tasks, such as requesting product information, clarifying a promotional claim, finding a return or complaint channel, or asking the system to explain a recommendation. Each task is decomposed into expected user goals and observable interface states. Stage 2 is structured user testing. Eligible participants should be adults who can read the test language and who match the intended user role of the interface, such as prospective customers, current customers, or service staff; exclusion should be limited to prior direct involvement in designing the tested interface, because such involvement would provide hidden system knowledge. Task order should be randomized for each participant, or counterbalanced if the task set is small. Participants are asked to perform the tasks while verbalizing their understanding of what the interface is doing and what action they expect next. One neutral prompt is permitted if a participant falls silent; any additional clarification counts toward failure of independent completion under the task rule in Section 4. Stage 3 is coded judgment. Trained raters score the variables using the operational rules above, then record the exact interface elements associated with confusion, delay, or repair.

The comparator is a rule-based baseline interface that provides the same predefined tasks using static decision trees, fixed FAQ menus, or non-AI scripted replies. The baseline should be evaluated with the same task definitions, the same participant pool, the same task-order procedure, and the same rater criteria used for the AI-assisted interface. The rationale for this comparator is not that rule-based systems are generally superior, but that they make state transitions explicit and therefore provide a transparent reference condition. If the AI-assisted interface is less comprehensible than the rule-based baseline on matched tasks under matched scoring rules, then the added AI layer is reducing accessibility rather than improving it. This comparison is consistent with methodological work showing that domain-aware automation should be judged against simpler baselines rather than against aspirational claims about intelligence [Tian et al.]. [2]

To reduce interpretive drift, raters follow a decision sequence. First, identify whether the user has understood the system’s current role. Second, verify whether the next action is visible without external help. Third, inspect whether the response introduces a new choice, a hidden assumption, or a conflicting claim. Fourth, determine whether the user can recover after an error. Fifth, assign task-level S and A. This ordered sequence is intended to avoid the common problem in marketing-interface studies where a system is judged acceptable because it produces persuasive output, even when the control structure remains opaque.

6. Relation to Prior Methods and Why They Are Insufficient Alone

Existing work offers useful but incomplete precedents. The chatbot study on intelligent customer service demonstrates that AI systems can structure queries through knowledge graphs and neural models, but its emphasis is service efficiency and predictive performance rather than user-facing accessibility criteria [He et al.]. That approach is useful as a technical baseline because it shows how AI mediation can improve handling of complex queries, yet it does not specify whether the user can understand why a response was generated or how to correct it. Similarly, the financial customer analytics framework shows how domain-aware automation can expose natural-language interfaces to business users, but its performance claims concern model quality and development time, not the accessibility of the end-user interaction [Tian et al.]. [1] [2]

The evidence-structured safety assessment framework provides the closest methodological analogy because it converts unstructured evidence into a hierarchical state representation and then uses an AI agent to inspect that state through explicit queries [Jing et al.]. The value of this analogy is formal, not topical. It suggests a transferable design principle: accessibility evaluation can benefit from separating the interface surface from the underlying interaction state and from requiring traceable evidence for each judgment. However, the building-inspection framework operates in a factual environment with measurable geometry, whereas marketing interfaces operate in a linguistic and persuasive environment where ambiguity may be produced by wording, personalization, or hidden action status rather than by missing measurements. The present protocol therefore retains the ideas of state decomposition and tool-assisted verification but replaces geometric measurement with user-centered checks on comprehension, action visibility, and recoverability. [5]

Adoption research in medical management and pharmaceutical marketing is informative but limited for this purpose. It reports that perceived usefulness and organizational readiness are positively associated with AI adoption intention, while data governance concern is a negative correlate, and that regulatory literacy conditions part of this adoption pattern [Boruga et al.]. These findings help explain why organizations may deploy AI-assisted systems and why deployment pressure may persist even under regulatory uncertainty. They do not answer whether a customer or employee can use a specific interface accessibly. The protocol therefore separates organizational adoption from interface accessibility so that willingness to deploy does not substitute for user-centered evidence. [3]

Finally, the systematic review of emojis in marketing and advertising is relevant as a methodological caution. It reports heterogeneous theories and methods, a focus on short-term outcomes, and limited attention to field validity [Vardikou et al.]. That pattern is useful here because AI-assisted marketing interfaces are similarly prone to over-reliance on engagement metrics. The present protocol avoids that limitation by defining accessible performance at the level of task completion, state transparency, and repairability rather than at the level of clicks or persuasion alone. [4]

7. Proposed Validation Design, Failure Cases, and Limits of Generalization

Validation should proceed in three steps. First, content validation by expert panel. The panel should contain at least one reviewer with primary expertise in accessibility evaluation, one in human-computer interaction, and one in marketing communication or promotional compliance; if one person covers multiple domains, the panel should still include no fewer than three reviewers. Experts should review the task set, the coding manual, and the harmful-ambiguity rule independently before any meeting. A paragraph, task, or coding definition should be revised when at least two reviewers identify the same ambiguity or when consensus cannot be reached after a recorded discussion of reasons. Second, construct validation. Compare the proposed scores with prospectively collected indicators such as clarification turns, abandonment, time to visible task completion, and user-reported confusion gathered immediately after each task. Third, contrastive validation. Apply the protocol to both an AI-assisted interface and a rule-based baseline under matched task conditions. A useful test is whether the protocol discriminates interfaces that are equally persuasive but differently transparent. Because no empirical study is reported here, these steps remain prospective design requirements rather than results. Any later empirical study should preregister participant inclusion criteria, task-order randomization, rater training, and a disagreement-resolution rule before data collection.

The main failure cases are predictable. Over-personalized language may appear helpful while actually increasing ambiguity about data use or recommended next steps. A system may also present compliant-looking explanations that are too dense to support comprehension. Another failure case is hidden state change, where the interface silently switches from advisory to transactional mode without clear notice. A final case is recoverability failure: the interface may offer a fallback option but make it visually or verbally subordinate to the AI-generated path. These cases matter because accessibility can fail even when the model produces fluent text. The protocol therefore treats fluency as insufficient evidence.

Generalization is limited in four ways. It is strongest for conversational or semi-conversational marketing interfaces that contain AI-generated content, recommendation logic, or service routing. It is weaker for static promotional pages without interaction states. It also requires adaptation for multilingual settings, regulated claims, and interfaces that serve both marketing and formal service functions. The proposed threshold values are hypothetical and should not be transferred unchanged across product categories or user groups. In particular, a score boundary that may be acceptable for a simple FAQ bot may be too lenient for a high-stakes product disclosure flow. This boundedness is deliberate, because the protocol is intended to support evaluation rather than to impose a universal accessibility standard.

8. Discussion

The principal conceptual contribution of this manuscript is to separate accessibility from adoption, engagement, and output quality in AI-assisted marketing interfaces. Prior studies indicate that AI is spreading into customer-service and marketing-related workflows, that domain-aware automation can support efficiency in financial customer analytics, and that organizational adoption is associated with usefulness and readiness [He et al., Tian et al., Boruga et al.]. Those strands motivate the relevance of the problem, but they do not define a user-centered accessibility test for marketing interactions. By introducing explicit operational rules, a task-level scoring structure, and a matched comparison against a rule-based baseline, the proposed protocol makes accessibility observable without reducing it to satisfaction, conversion, or organizational acceptance. [1] [2] [3]

The protocol also addresses a methodological weakness visible in adjacent marketing research: many studies emphasize behavioral outcomes while leaving the communicative mechanism insufficiently specified. For AI-assisted marketing, that omission is more serious because the interface itself is often the persuasive medium. A system may increase response volume or perceived relevance while still concealing state changes, overstating certainty, or blocking repair. The proposed method therefore treats ambiguity as a first-order accessibility defect rather than as a secondary usability inconvenience, and it requires raters to document the interface state that produced the failure rather than inferring accessibility from engagement alone. [4]

At the same time, the protocol does not claim that all accessible interfaces are ethically acceptable or commercially effective. Accessibility is necessary for equitable use, but it does not resolve questions of manipulation, data governance, or promotional compliance. Nor does it guarantee adoption. Those broader questions remain connected to, but distinct from, the accessibility problem addressed here. [3]

9. Conclusion

Conceptually, this manuscript establishes a bounded protocol for evaluating accessibility in AI-assisted marketing interfaces. It defines accessibility as a joint property of comprehension, navigability, recoverability, and referential clarity; proposes an explicit task-level scoring rule; and situates the evaluation against a non-AI baseline to distinguish transparent interaction from merely fluent AI output. It also specifies how prior work on customer-service chatbots, domain-aware automation, adoption intention, and evidence-structured AI assessment can be adapted to this topic without collapsing accessibility into usefulness or engagement. [1] [2] [3] [5]

What remains untested is the protocol’s empirical performance. The score thresholds, coding reliability, participant burden, and sensitivity to language, domain, and user expertise all require prospective validation. Until that work is done, the manuscript supports a precise evaluation design, not an empirical verdict on any particular marketing system.

Contribution

This manuscript offers a topic-specific research protocol for accessibility evaluation of AI-assisted marketing interfaces. The contribution is methodological and deliberately bounded. First, it defines accessibility for this interface class as a joint property of comprehension, navigability, recoverability, and referential clarity, rather than as a proxy for satisfaction, trust, adoption intention, or sales performance. Second, it translates that definition into a task-based scoring procedure with an explicit default pass rule, a harmful-ambiguity override, and preregistered sensitivity thresholds. Third, it specifies comparator logic against a rule-based baseline so that the added AI layer can be judged for transparency rather than presumed beneficial. Fourth, it outlines a prospective validation design based on expert review, user testing, and coded evidence of confusion and repair. The manuscript does not claim completed experimentation, population-wide generalizability, or established superiority over alternative protocols.

Validation status

Conceptual protocol only; no empirical validation has been performed in this manuscript. All variables, thresholds, and decision rules should therefore be read as prospective specifications for later testing rather than as verified properties of any deployed system. In particular, the proposed composite score S(t), the separate comprehension index CI(t), the harmful-ambiguity override, the comparator against a rule-based interface, and the suggested validation sequence have not yet been examined for reliability, construct validity, sensitivity to task order, or cross-context stability. Any future study using this protocol should preregister the task set, participant inclusion criteria, rater training procedure, threshold sensitivity analysis, and criteria for disagreement resolution before collecting user data. Reports should also distinguish exploratory refinements from confirmatory tests so that later revisions of the protocol remain auditable.

References

1. 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.
2. Tian, Ye; Shao, Wenqian; Deng, Zihan (2026). Marketing-AutoM3L: domain-aware automated machine learning for financial customer analytics. Frontiers in Artificial Intelligence. 9, 1726900. doi:10.3389/frai.2026.1726900.
3. Boruga, Veronica Madalina; Bratu, Melania Lavinia; Puenea, George; Popa, Daniel; Popa, Cristina Annemari; Bogdan, Iulia Georgiana; Savencu, Cristina Elena (2026). Organizational Readiness, Perceived Usefulness, and Determinants of Artificial Intelligence Adoption in Romanian Medical Management and Pharmaceutical Marketing. Healthcare. 14(12), 1714. doi:10.3390/healthcare14121714.
4. 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.
5. Jing, Zijian; Zhu, Liyi; Chen, Tianyi; Hovestadt, Ludger; Li, Li (2026). Vision-Based Digital Twin and AI Agent Framework for Low-Cost, Explainable Indoor Building Inspection and Safety Assessment. Sensors. 26(15), 4992. doi:10.3390/s26154992.