Why AI Needs to “Speak with Data”: The Impact Mechanism of Digitalized Descriptions by Virtual eWOM Senders on eWOM Effectiveness

Abstract

Based on the Persuasion Knowledge Model (PKM), this research investigates how virtual electronic word-of-mouth (eWOM) senders’ message framing—numerical versus experiential—influences eWOM effectiveness across three experiments. We find that: (1) numerical descriptions from virtual eWOM senders significantly enhance eWOM effectiveness compared to experiential descriptions, while this effect does not emerge for human senders; (2) perceived diagnosticity mediates the relationship between message framing and eWOM effectiveness; and (3) product type moderates this effect pathway, with numerical descriptions showing stronger positive effects for search products than for experience products. This research enriches theoretical understanding of eWOM communication in interactive marketing and provides practical guidance for e-commerce companies to optimize their content marketing strategies.

1. Introduction

The rapid development of digital technologies and the proliferation of social media have transformed brand–consumer relationships from unilateral communication to bidirectional interaction, making interactive marketing one of the fastest-growing fields in contemporary business environments [1,2]. On e-commerce platforms such as Amazon, Taobao, and JD.com, consumers not only enjoy convenient and efficient shopping experiences but also form diverse interaction networks with brands and other consumers through product reviews, community discussions, and livestream interactions. These interactions have reshaped information dissemination pathways and profoundly influenced consumer purchase decision processes [3].

Consumers increasingly rely on Electronic Word-of-Mouth (eWOM) as a powerful driver of their purchasing decisions [4]. Research indicates that over 90% of consumers check reviews before making online purchases, and approximately 50% report that online reviews directly impact their purchasing decisions [5,6]. Against this backdrop, advances in machine learning, natural language processing, and other forms of artificial intelligence have led many brands to introduce virtual eWOM senders who participate in word-of-mouth communication as reviewers. The application of these virtual influencers has expanded brand communication channels, significantly impacted retail formats and sales models, and further enhanced consumer shopping experiences and brand interactions [7]. This AI-driven interactive marketing approach has been show to significantly improve consumer experience and interaction efficiency [8]. Compared to human reviewers, virtual influencers offer greater content control and can provide precise descriptions of specific product attributes without the limitations of real individual experiences. However, as “non-human agents”, consumer perception and acceptance of their recommendations involve more complex cognitive processes [9,10,11].

While numerous studies have focused on how characteristics of eWOM senders and receivers influence eWOM effectiveness, including sender identity, expertise, similarity to receivers, and social familiarity [12,13], as well as receivers’ product knowledge, involvement levels, and information processing motivations [14,15,16], most existing research has centered on human reviewers. Additionally, review characteristics such as volume, text features, readability, errors, and emotional cues have been shown to affect consumer attitudes and purchase decisions [17,18,19,20]. However, these discussions of online reviews have largely been built on users’ authentic identities and genuine experiences. In contrast, virtual eWOM senders as a new practice in e-commerce review scenarios have not received sufficient attention, with current literature on this topic remaining quite limited.

Present research on virtual influencers primarily focuses on their image design or social interaction capabilities, such as how their anthropomorphism level, visual appeal, and personality traits influence user cognition and brand attitudes [21,22]. Few studies have addressed their content generation and expression mechanisms when participating in eWOM communication as “reviewers” in e-commerce contexts. What review approaches should virtual eWOM senders adopt to gain consumer acceptance? Compared to the emotional and experiential expression styles based on subjective experiences typical of real user reviews [23], would a more objective, quantified “numerical” description framework from virtual eWOM senders yield better communication outcomes? As novel information carriers that simulate user identities to generate product reviews, virtual eWOM senders create an entirely new communication context through their functional identity and non-human qualities. Traditional research perspectives on virtual influencers struggle to fully explain how consumers process and evaluate content from these information sources. Based on these considerations, this study attempts to fill this theoretical gap by exploring content generation strategies for virtual eWOM senders from the perspective of message framing.

This research employs the Persuasion Knowledge Model (PKM) as its theoretical foundation [24]. This theory proposes that when consumers encounter persuasive information, they activate three knowledge structures—agent knowledge, topic knowledge, and persuasion knowledge—to evaluate the communicator’s motives and credibility, and subsequently select and execute effective and appropriate response strategies. When information presentation does not align with consumer expectations about the information source, persuasion knowledge may be activated, triggering resistance or skepticism. Building upon this established model, this study introduces a novel perspective by applying it to the context of virtual eWOM senders. Specifically, consumers generally perceive artificial intelligence as having significant advantages in analyzing and processing structured data, while relatively lacking in expressing subjective emotions and individual experiences [25]. Therefore, when they use numerical descriptions, they better align with consumers’ expectations, enhancing information credibility and persuasiveness. Conversely, when they employ experiential descriptions, the violation of expectations may trigger cognitive conflict, activate the persuasion knowledge system, and consequently weaken persuasive effects and information adoption intentions.

Based on these theoretical hypotheses, this study constructs a mediation model to explore how different message framing by virtual eWOM senders influences eWOM effectiveness through consumers’ perceived diagnosticity. Product type is also introduced as a moderating variable. This research includes three experiments: Experiment 1 verifies the main effect, showing that numerical descriptions significantly enhance eWOM effectiveness compared to experiential descriptions. Through comparative experiments, we find this effect does not exist for human reviewers, confirming the uniqueness of virtual agents. Experiment 2 further examines the mediating role of perceived diagnosticity, revealing that numerical descriptions lead to higher perceived diagnosticity among consumers, thereby enhancing eWOM effectiveness. Experiment 3 explores the boundary conditions of product type, with results showing that the advantage of numerical descriptions is more significant for search products, while the difference between description frameworks is smaller for experience products.

This study extends the Persuasion Knowledge Model by applying it to virtual eWOM senders, revealing how message framing influences consumer cognition and behavior, thereby bridging a literature gap. Practically, it guides online e-commerce platforms in developing AI content strategies and helps e-commerce businesses optimize messaging by product type to enhance effectiveness and trust.

2. Literature Review and Hypothesis Development

2.1. Virtual eWOM Senders

Electronic word-of-mouth (eWOM) refers to informal information about products or services that consumers publish or disseminate online [26], representing a bidirectional social communication form between consumers. On one hand, consumers actively share their product experiences and evaluations, playing roles as reviewers and influencers across various digital platforms and brand communities [27]. On the other hand, potential consumers gather feedback information by examining others’ reviews to support their decision-making processes [3]. According to Shannon’s communication theory framework [28], any communication act comprises three fundamental elements: sender, message content, and receiver. As the starting point of the communication chain, the sender directly influences communication effectiveness.

Unlike traditional offline word-of-mouth communication, eWOM senders are no longer restricted by factors such as appearance, background, social status, or professional credentials [29]. The internet environment provides diverse information sources, allowing consumers to access product opinions and experience sharing from different perspectives [30,31]. Among these, online reviews represent the most representative and widely used form of eWOM, becoming increasingly important in today’s information-overloaded digital environment. Consumers generally view online reviews as crucial references for understanding product information and making purchase decisions, especially when other reliable information sources are lacking users on social media and digital platforms to promote brands and disseminate information [30]. Compared to human influencers, virtual influencers offer advantages including more consistent content output, stronger controllability, and lower operational risks, effectively avoiding brand risks associated with inappropriate behavior or image crises of human influencers [3,32].

Traditional review publishers are generally authentic individuals based on personal experience, whose persuasiveness stems from their genuine personal experiences and emotional resonance. Their review content typically carries strong personal subjective feelings and experiential descriptions, thus possessing high credibility and impact [33,34]. In recent years, the development of artificial intelligence technology and the transformation of the digital marketing environment have given rise to virtual influencers as emerging marketing communication entities. Virtual influencers are typically driven by algorithms and artificial intelligence technologies, functioning through content creation and user interaction on social media and digital platforms to promote brands and disseminate information. Compared to human influencers, virtual influencers offer advantages including more consistent content output, stronger controllability, and lower operational risks, effectively avoiding brand risks associated with inappropriate behavior or image crises of human influencers [35,36,37]. Building on prior work, this study introduces Virtual eWOM sender, which is a subdivision of virtual influencers, focusing on the creation and dissemination of text products and user experience feedback. These virtual entities do not have personal experience using products, but instead rely on technology to simulate consumers’ expression styles and behavioral patterns, publishing product evaluations in comment sections of various online platforms through preset text content. Therefore, the presentation format of their information largely determines the persuasiveness and effectiveness of word-of-mouth communication. This research focuses on how the information description framework of virtual eWOM senders affects eWOM effectiveness.

2.2. Message Framing

Framing effect refers to the phenomenon where different presentation methods of same information influence people’s attitudes, preferences, or decision-making behaviors [38]. This effect stems from cognitive biases that individuals develop when the same information is presented in different ways [39]. In marketing communication, the reasonable application of framing effects can effectively guide consumers’ decision tendencies by influencing their cognitive responses [40,41]. For instance, different researchers have proposed differentiated conclusions regarding the effects of gain-framed versus loss-framed expressions: gain-framed messaging can enhance consumers’ perception of information effectiveness [42], while loss-framed messaging more easily triggers consumer attention and purchasing behavior [43,44]. Therefore, clarifying the characteristics and applicable contexts of different frames is of significant importance for improving communication effectiveness.

As an important basis for consumer decision-making, the communication effect of word-of-mouth information similarly depends on the structure and method of information presentation. Existing research has categorized word-of-mouth information into factual and evaluative types based on linguistic expression characteristics [45]. Factual information refers to describing product features with objective, rational, and verifiable language, while evaluative information uses subjective, emotional language to express personal feelings and experiences. Correspondingly, online review information is also divided into objective factual and subjective evaluative types. Factual information refers to describing product characteristics using objective, logical language that can be verified; evaluative information refers to describing product characteristics using subjective, emotional language. Additionally, in research on e-commerce platform reviews, Chen (2016) proposed that online reviews can be further subdivided into digital ratings and open-text comments [46]. Building on the above literature, this study proposes an information description framework that refines existing classification standards by distinguishing between numerical and experiential descriptions. Numerical descriptions contain quantified, objective, and verifiable data content, such as product parameters, performance indicators, ranking data, and price comparisons; experiential descriptions focus on expressing subjective feelings and personalized experiences, such as comfort of use, emotional satisfaction, and personal preferences. Based on the Persuasion Knowledge Model theory, this research aims to analyze how these two information description frameworks used by virtual eWOM senders differentially influence consumers’ cognitive responses and word-of-mouth communication effectiveness.

2.3. The Persuasion Knowledge Model

The Persuasion Knowledge Model (PKM), proposed by Friestad and Wright, explains consumers’ cognitive response mechanisms when faced with marketing persuasion attempts [24]. This model posits that consumers are not merely “passive recipients “of persuasive information. According to PKM theory, individuals possess three interconnected knowledge structures: first, topic knowledge, referring to individuals’ understanding of products or services; second, persuasion knowledge, indicating consumers’ comprehension of marketing intentions and expression strategies behind information; and third, agent knowledge, representing individuals’ perceptions of information providers’ identity, motives, and credibility. When consumers identify a persuasion attempt, they activate relevant persuasion knowledge and use it to interpret and respond to persuasive information, thereby influencing persuasion effectiveness [47].

In AI marketing practice, persuasion mechanisms are undergoing profound transformation. Algorithmic persuasion refers to the process where AI systems serve as intermediaries delivering content that influences user attitudes and behaviors. Here, brands or organizations act as persuaders, and effectiveness depends on their strategic choices and how they deploy these algorithmic tools [48,49]. With the emergence of virtual influencers, algorithmic persuasion is evolving into a new paradigm, where algorithms themselves become the subject of persuasion, with an “image” to cling to. Algorithms no longer function merely as backend information filters or recommendation engines [50]; instead, they act as highly personified, generative “recommenders” on the frontend. Their core task shifts from transmitting information to creating new persuasive content, thereby achieving persuasion through entirely novel forms of expression. Despite the accuracy advantages of statistical models in prior recommender system research, people remain skeptical about accepting AI recommendations due to resistance, anxiety, or ambivalence toward this emerging technology [51].

When facing recommendations from statistical models, some people prefer human recommendations [52], as evidenced in AI applications for book recommendation systems [53]. In this emerging context, the question consumers confront transforms into “how to establish dynamic trust with an algorithm agent that possesses persuasive capabilities”. This trust operates dynamically because consumers must continuously evaluate the AI “persuader’s” reliability, intentions, and the authenticity of its generated content [54], then adjust their interaction and adoption behaviors accordingly.

Overall, the traditional PKM framework requires expansion and deepening in this new context. This study argues that in contexts where virtual eWOM senders provide product recommendations online, their published review content essentially constitutes a persuasion attempt to guide consumer purchases, with virtual information senders thus playing the role of persuasion agents. Based on the PKM theoretical framework, consumers’ agent knowledge about artificial intelligence—their understanding of the cognitive abilities, limitations, and potential commercial motives of AI—directly forms the cognitive basis of the “AI trust framework” and significantly influences their evaluation of recommended information.

Generally, consumers tend to view artificial intelligence as data-driven, logically rigorous, objective, and lacking subjective emotional experiences. This cognitive bias stems from AI’s operational principles based on big data analysis and algorithmic computations, upon which consumers form stereotypes about its superior computational capabilities and objective neutrality in processing information [55,56,57]. When considering issues involving deep logic, users more readily accept conclusions from artificial intelligence, and in certain decision scenarios, people prefer recommendations from algorithm-based AI [58]. Logg et al. found that across various prediction tasks, compared to human opinions, users tend to adopt recommendations and suggestions from algorithms, demonstrating an AI preference [59]. This “rationalized” perception of artificial intelligence also leads consumers to have clear consistency expectations for AI-generated content. When artificial intelligence exhibits characteristics “incongruent with its identity”, such as designing emotion-driven products (artworks), consumers experience cognitive dissonance, questioning its credibility and authenticity [60,61]. Language Expectancy Theory suggests that people hold pre-established expectations about what language style specific types of communicators should use [62]. Expectancy Violation Theory further clarifies that violating these expectations triggers cognitive evaluations of the communicator’s social competence [63], which ultimately determines whether the outcome is positive or negative. Consumer responses to artificial intelligence largely depend on whether AI performance aligns with expected rational characteristics; when its behavior deviates from its “rational tool” role, it often triggers negative emotions and resistance from consumers. This agent knowledge about artificial intelligence forms the foundational framework for consumers’ interpretation of eWOM information.

In the context of this study, consumers’ agent knowledge of virtual eWOM senders forms expectations about their behavioral patterns. When the description framework of eWOM information aligns with these expectations, consumers perceive cognitive congruence, thereby enhancing trust in recommenders and acceptance of eWOM information [64]. Specifically, when virtual eWOM senders employ numerical descriptions, this information presentation method aligns with consumers’ agent knowledge of artificial intelligence as “objective, rational, and data-driven”. In this case, consumers consider the provision of eWOM information to be based on AI’s proficient data processing and analytical capabilities, highly consistent with the essence of artificial intelligence. This consistency strengthens consumers’ perception of virtual eWOM senders’ credibility and reduces the likelihood of activating defensive persuasion knowledge. Consumers are more inclined to view numerical recommendations as objective, useful information provision, increasing their perceived diagnosticity of eWOM information, rather than as “selling” with strong persuasive intent. Perceived diagnosticity refers to consumers’ subjective judgment of an information source’s effectiveness in conveying product-related information [65]. Information with higher diagnosticity better helps consumers differentiate between options, evaluate product attributes, and make decisions, effectively reducing decision uncertainty [66]. Based on the Persuasion Knowledge Model theory, when evaluating eWOM recommendations, consumers judge information credibility and usefulness according to their relevant agent knowledge. When virtual eWOM senders use numerical descriptions, consumers perceive that the eWOM information aligns with their agent knowledge and thus interpret it as more reliable and valuable decision reference [67], thereby perceiving the information as having higher diagnosticity. For consumers, reviews from virtual eWOM senders represent a process of information input, extraction, and integration [68]. The higher the diagnosticity of eWOM information, the more likely it will be adopted by consumers for subsequent behaviors and decisions, thereby enhancing eWOM effectiveness.

When virtual eWOM senders employ experiential descriptions, since artificial intelligence is generally considered to lack human subjective emotions and experiential capabilities, consumers may question how it can “feel” or “experience” products. This description method’s mismatch with consumers’ relevant agent knowledge leads them to doubt information authenticity and reliability, perceive potential manipulative strategies, generate resistance emotions, and activate defensive persuasion knowledge.

When virtual eWOM recommenders use experiential descriptions, based on agent knowledge, consumers recognize AI’s lack of genuine experiential capabilities. This cognitive dissonance makes consumers aware of the low value of such information in helping them evaluate products and make decisions, thereby weakening perceived diagnosticity of eWOM information and reducing eWOM effectiveness.

Based on the theoretical analysis above, this study proposes the following hypotheses:

H1: 

The message framing (numerical vs. experiential) of virtual eWOM senders significantly influences eWOM effectiveness. Specifically, numerical descriptions by virtual eWOM senders significantly enhance eWOM effectiveness compared to experiential descriptions.

H2: 

Perceived diagnosticity mediates the relationship between message framing of virtual eWOM senders and eWOM effectiveness.

2.4. The Moderating Effect of Product Type

In e-commerce research, product type has been widely confirmed to significantly influence consumers’ focus of attention and evaluation dimensions regarding information content [69]. According to Nelson’s classification, products can be categorized as search products and experience products [70]. Purchase decisions for search products primarily rely on objective attribute information (such as specifications, technical indicators, etc.), which possesses verifiability; experience products emphasize subjective usage impressions (such as emotional experiences, personalized effects), with evaluations featuring individual differences [71]. According to the Persuasion Knowledge Model, consumers’ topic knowledge influences their identification and degree of recognition of persuasive intent. Therefore, we believe that product type, as an important component of topic knowledge, moderates consumers’ acceptance of virtual recommenders’ language styles and the degree of persuasion knowledge activation.

In search product contexts, consumers, based on their existing topic knowledge, typically expect to obtain objective, data-oriented information to pre-evaluate products. According to Language Expectancy Theory (LET) [62], when virtual recommenders employ numerical language styles, their expression meets consumers’ language expectations, maintaining persuasive power and enhancing consumers’ perceived diagnosticity of eWOM information and eWOM effectiveness [47]. Conversely, when virtual recommenders use experiential language for search products, this subjective expression style conflicts with consumers’ language expectations formed based on topic knowledge, thereby reducing perceived diagnosticity of eWOM information and eWOM effectiveness.

In experience product contexts, although consumers’ existing topic knowledge directs them to focus on subjective usage impressions in product evaluations, they also recognize that product quality in this category is difficult to judge definitively before purchase. Therefore, consumers’ consideration of information source credibility becomes particularly crucial. Research indicates that trust is a priority judgment factor in consumer information adoption and a decisive antecedent for users’ final decisions [72]. When consumers face difficult judgments or ambiguous evaluation criteria, they more cautiously assess the credibility of eWOM information sources [60,73]. Virtual eWOM senders, due to their lack of genuine experiences, have inherently limited trust foundations for “experiential narratives”. Therefore, in experience product contexts, when virtual senders adopt numerical language styles, this information content is inconsistent with consumers’ expected information for experience products and fails to help consumers effectively judge the value of experience products, resulting in lower perceived diagnosticity and eWOM effectiveness. However, when virtual recommenders adopt experiential language styles that meet consumer expectations, consumers also question the authenticity and credibility of their expressions, leading to lower perceived diagnosticity and eWOM effectiveness. Consequently, in this context, regardless of which message framing virtual eWOM senders adopt, it is difficult to significantly influence eWOM effectiveness through consumers’ perceived diagnosticity of eWOM information.

Based on the theoretical analysis above, this study proposes the following hypothesis:

H3: 

Product type moderates the influence of message framing adopted by virtual eWOM senders on eWOM effectiveness. When the product is a search good, numerical descriptions by virtual eWOM senders enhance eWOM effectiveness more than experiential descriptions. When the product is an experience good, message framing by virtual eWOM senders does not significantly affect eWOM effectiveness.

3. Overview of Experiments

This research tests the hypotheses through three experiments, each building upon the previous to comprehensively explore the influence of message framing by virtual eWOM senders on consumer behavior. Experiment 1 aims to establish the main effect by demonstrating that virtual eWOM senders who employ numerical descriptions are more effective in enhancing eWOM effectiveness compared to those using experiential descriptions. This provides the foundation for our investigation. Experiment 2 extends the understanding by refining the causal chain, identifying the mediating role of perceived diagnosticity in how virtual eWOM senders’ message framing influences eWOM effectiveness. The results confirm that numerical descriptions enhance eWOM effectiveness by increasing consumers’ perceived diagnosticity of information. Finally, Experiment 3 investigates the moderating role of product type. It reveals that for search products, numerical language styles adopted by virtual eWOM senders enhance eWOM effectiveness more than experiential language styles, while for experience products, the message framing type has little effect on eWOM effectiveness. Together, these experiments form a cohesive analysis demonstrating the specific conditions under which message framing by virtual eWOM senders impacts consumer behavior.

3.1. Experiment 1

Experiment 1 aims to validate H1, demonstrating that virtual eWOM senders using numerical descriptions enhance eWOM effectiveness more than those using experiential descriptions. This effect does not exist in human eWOM sender contexts.

3.1.1. Participants

We employed a 2 (eWOM sender type: virtual vs. human) × 2 (message framing: numerical vs. experiential) between-subjects experimental design. We recruited 196 participants through a professional survey platform and randomly assigned them to four conditions evaluating a SmartWatch product. After excluding participants who failed attention checks and manipulation checks, we retained 185 valid responses (ages 19–56, M = 32.80, SD = 8.25, with 52.97% female participants). The sample sizes for each condition were: n_{virtual,numerical} = 45, n_{virtual,experiential} = 47, n_{human,numerical} = 46, and n_{human,experiential} = 47.

3.1.2. Method

To ensure the validity of our experimental manipulations, we recruited 126 participants from a university (ages 19–25, M = 21.56, SD = 2.07, with 53.05% female participants) and randomly assigned them to the 2 × 2 experimental design to verify the effectiveness of our message framing (numerical vs. experiential) and eWOM sender type (virtual vs. human) manipulations. We created a virtual eWOM sender named “Nova” to introduce SmartWatch products to consumers.

The eWOM sender type manipulation was as follows:

Participants in the virtual condition viewed comments from a user with a robot avatar and the ID “Nova”. They read the prompt: “Imagine you are shopping online and browsing reviews for a smartwatch. You see the following comment posted by a virtual eWOM sender”.

Participants in the human condition viewed comments from a user with a default avatar and the ID “Zhang San”. They read the prompt: “Imagine you are shopping online and browsing reviews for a smartwatch. You see the following comment posted by another user”.

The message framing manipulation was as follows:

Participants in the numerical condition read: “This smartwatch offers 18-h battery life, a 35% improvement over the previous generation. Its heart rate monitor has only ±2% error margin and tracks your health in real-time. It’s waterproof to 50 m, suitable for various daily scenarios. The system response time is under 0.3 s with less than 0.5-s delay. Sleep monitoring analyzes 6 sleep states with 92% accuracy. Overall, this smartwatch combines cutting-edge technology with perfect design—it’s worth having!”

Participants in the experiential condition read: “This smartwatch has excellent battery life, a significant improvement over the previous generation. Its heart rate monitor is highly accurate and helps you track your health in real-time. The waterproofing is outstanding—no need to worry during daily activities. The system runs smoothly with no lag. The sleep monitoring feature helps you understand various sleep states with reliable accuracy. Overall, this smartwatch combines cutting-edge technology with perfect design—it’s worth having!”

Participants then rated the numerical and experiential nature of the content: “The reviewer’s description contains specific numerical data” and “The reviewer describes product features through feelings and emotions” (7-point scale, 1 = strongly disagree, 7 = strongly agree) [74]. They also assessed the eWOM sender type (7-point scale: 1 = human influencer, 7 = virtual influencer) [75].

Results confirmed our manipulations were effective: The numerical condition scored significantly higher on numerical content than the experiential condition (M_numerical = 5.79, SD = 1.28; M_experiential = 2.34, SD = 1.29, t = 15.05, df = 124, p < 0.001, d = 2.682). The experiential condition scored significantly higher on experiential content than the numerical condition (M_numerical = 3.10, SD = 1.50; M_experiential = 4.92, SD = 1.55, t = 6.72, df = 124, p < 0.001, d = 1.198). The virtual eWOM sender was rated as significantly more virtual than the human eWOM sender (M_virtual = 5.28, SD = 1.44, M_human = 2.06, SD = 1.01, t = 14.56, df = 124, p < 0.001, d = 2.581).

In the main experiment, participants were told the purpose was to assess consumer attitudes toward new products. They were instructed to evaluate the product carefully. Participants viewed a product review community featuring one of four different versions of product recommendations based on our 2 × 2 design. Previous research has shown that the quantity of word-of-mouth recommendations significantly affects the effectiveness of word-of-mouth communication [76]. Additionally, product involvement level also influences consumers’ processing methods and acceptance of word-of-mouth information [77]. To control for potential confounds, all participants were told this was their first exposure to reviews about the product and were asked to evaluate from an observer’s perspective. We established a uniform product usage scenario to control for involvement level: participants imagined they were considering purchasing a smartwatch, had basic knowledge of standard features (heart rate monitoring, sleep tracking, message alerts), had browsed similar products online without researching specific brands, and had a budget of approximately 1000 yuan. After the experiment, participants evaluated the product recommendation by rating eWOM effectiveness: “I think this review is authentic”, “I think this review is accurate”, “I think this review is credible” [78]. They also completed measures of emotional state [79], product involvement [77], perceived review quantity [77], message framing [74] and eWOM sender type [75]. Finally, participants guessed the study’s purpose.

3.1.3. Results and Discussion

Manipulation checks: We excluded data from 11 participants (5 who failed attention checks and 6 who correctly guessed the study’s true purpose). The numerical condition scored significantly higher on numerical content than the experiential condition (M_numerical = 5.74, SD = 0.89; M_experiential = 2.32, SD = 1.32, t = 20.67, df = 183, p < 0.001, d = 3.021). The experiential condition scored significantly higher on experiential content than the numerical condition (M_numerical = 2.87, SD = 1.17; M_experiential = 4.76, SD = 1.27, t = 10.53, df = 183, p < 0.001, d = 1.549). Virtual eWOM senders were rated as significantly more virtual than human eWOM senders (M_virtual = 5.16, SD = 1.18; M_human = 3.75, SD = 1.45; t = 7.26, df = 183, p < 0.001, d = 1.067). These results confirmed our manipulations were effective.

eWOM effectiveness: Results revealed a significant interaction between eWOM sender type and message framing on eWOM effectiveness (F = 26.86, p < 0.001). In the virtual eWOM sender context, the numerical condition yielded significantly higher eWOM effectiveness than the experiential condition (M_numerical = 5.59, SD = 1.00; M_experiential = 4.58, SD = 1.12; t = 5.10, df = 90, p < 0.001, d = 1.063). In the human eWOM sender context, the experiential condition yielded significantly higher eWOM effectiveness than the numerical condition (M_numerical = 5.19, SD = 0.85; M_experiential = 5.61, SD = 0.77; t = 2.15, df = 91, p < 0.05, d = 0.447). These findings support H1.

Control variables: We found no significant differences across the four conditions in perceived review quantity (F(3,181) = 0.70, p = 0.550; M_{virtual,numerical} = 4.50, SD = 0.59; M_{virtual,experiential} = 4.58, SD = 0.86; M_{human,numerical} = 4.38, SD = 0.64; M_{human,experiential} = 4.55, SD = 0.67), product involvement (F(3,181) = 0.08, p = 0.971; M_{virtual,numerical} = 4.89, SD = 0.83; M_{virtual,experiential} = 4.91, SD = 0.96; M_{human,numerical} = 4.96, SD = 1.02; M_{human,experiential} = 4.87, SD = 0.90), or emotional state (F(3,181) = 0.46, p = 0.708; M_{virtual,numerical} = 3.90, SD = 0.82; M_{virtual,experiential} = 3.82, SD = 0.55; M_{human,numerical} = 3.94, SD = 0.66; M_{human,experiential} = 3.96, SD = 0.75). This rules out potential interference from these factors on our experimental results.

The data from Experiment 1 confirm H1, providing a foundation for exploring the underlying mechanisms and boundary conditions that explain why numerical descriptions enhance eWOM effectiveness for virtual eWOM senders.

3.2. Experiment 2

Experiment 2 aims to verify H2, demonstrating that numerical descriptions by virtual eWOM senders enhance eWOM effectiveness by increasing consumers’ perceived diagnosticity.

3.2.1. Participants

We recruited 136 participants through a professional survey platform to complete a series of questionnaires about product recommendations from virtual eWOM senders. Participants were randomly assigned to two conditions (numerical vs. experiential). After excluding those who failed attention checks and manipulation checks, we retained 131 valid responses (ages 18–54, M = 32.60, SD = 8.28, with 48.09% female participants). The sample sizes for each condition were: n_numerical = 65 and n_experiential = 66.

3.2.2. Method

We designed a smart speaker named “Vox” as the experimental product and created a virtual eWOM sender named “Nova” to recommend this product. To ensure the effectiveness of our message framing manipulation (numerical vs. experiential), we recruited 92 participants from a university (ages 20–29, M = 23.42, SD = 2.55, with 46.91% female participants) for a pretest, randomly assigning them to two conditions. Participants were told they would read a product review from a virtual eWOM sender.

The numerical condition read:

“Hello everyone! I’m Nova! The Vox smart speaker has reached 500,000 units in sales with 15% market share. Its bass extends to 20 Hz with 96% voice recognition accuracy. Battery tests show 12-h usage with 20% faster charging time. It connects to 64 smart devices with 40% improved transmission distance. I believe it will be a great assistant for your smart lifestyle—highly recommended!”

The experiential condition read:

“Hello everyone! I’m Nova! The Vox smart speaker is widely loved by users with excellent sales performance. It delivers powerful bass and nearly flawless voice recognition. The battery life is impressively long with significantly reduced charging time. It connects seamlessly with all your home smart devices and offers extensive signal coverage. I believe it will be a great assistant for your smart lifestyle—highly recommended!”

Participants then rated the numerical nature of the virtual eWOM sender’s description [74]. Results showed that the numerical condition scored significantly higher on numerical content than the experiential condition (M_numerical = 5.82, SD = 0.96; M_experiential = 2.21, SD = 1.25, t = 15.48, df = 90, p < 0.001, d = 3.229). The experiential condition scored significantly higher on experiential content than the numerical condition (M_numerical = 2.87, SD = 1.36; M_experiential = 4.81, SD = 1.17, t = 7.35, df = 90, p < 0.001, d = 1.533), confirming the effectiveness of our message framing manipulation for Experiment 2.

In the main experiment, participants were told the study aimed to investigate consumer attitudes toward new product recommendation information and were asked to carefully evaluate the review content. To control for perceived review quantity, participants were told this was their first exposure to reviews about the product. To control for potential product involvement effects, we established a uniform usage scenario: participants imagined they were considering purchasing a smart speaker, had some prior interest in such products, but hadn’t decided on a specific model. Participants were also informed that the study focused on their opinions about the information content rather than actual purchase decisions. After the experiment, participants rated the numerical and experiential nature of the virtual eWOM sender’s description [74]. Participants then reported their perceived diagnosticity of the eWOM information: “These reviews help me systematically and effectively search and compare different product evaluation information to find the most suitable one”, “These reviews help me effectively evaluate the products I browse based on information provided by other consumers” and “These reviews provide me with the opportunity to systematically and effectively search and evaluate numerous products” [80]. They also completed measures of eWOM effectiveness. Finally, we measured potential confounding factors, including participants’ emotional state [79], product involvement [77], and perceived review quantity [77]. Participants were then asked to guess the study’s purpose.

3.2.3. Results and Discussion

Manipulation checks: We excluded data from 6 participants (5 who relied on past usage experience and 1 who correctly guessed the study’s true purpose). The numerical condition scored significantly higher on numerical content than the experiential condition (M_numerical = 5.86, SD = 0.83; M_experiential = 2.23, SD = 1.13, t = 20.98, df = 129, p < 0.001, d = 3.658). The experiential condition scored significantly higher on experiential content than the numerical condition (M_numerical = 2.63, SD = 1.05; M_experiential = 4.91, SD = 1.06, t = 12.32, df = 129, p < 0.001, d = 2.152). These results confirmed the effectiveness of our message framing manipulation in Experiment 2.

Perceived diagnosticity: Results showed significant differences between conditions in perceived diagnosticity of information provided by virtual eWOM senders.

Participants in the numerical condition reported significantly higher perceived diagnosticity of the product information than those in the experiential condition (M_numerical = 5.73, SD = 0.70; M_experiential = 4.99, SD = 0.78, t = 5.703, df = 129, p < 0.001, d = 0.997).

eWOM effectiveness: Results revealed significant differences in eWOM effectiveness between conditions, with the numerical condition showing significantly higher eWOM effectiveness than the experiential condition (M_numerical = 5.59, SD = 0.55; M_experiential = 4.71, SD = 0.70, t = 8.051, df = 129, p < 0.001, d = 1.404).

Control variables: We found no significant differences between conditions in perceived review quantity (M_numerical = 4.49, SD = 0.87; M_experiential = 4.51, SD = 0.86, t = 0.11, df = 129, p = 0.910, d = 0.020), product involvement (M_numerical = 4.71, SD = 0.68; M_experiential = 4.80, SD = 0.74, t = 0.73, df = 129, p = 0.469, d = 0.127), or emotional state (M_numerical = 3.87, SD = 0.93; M_experiential = 3.98, SD = 0.91, t = 0.70, df = 129, p = 0.484, d = 0.123). This rules out potential interference from these factors on our experimental results.

Mediation analysis: To further examine the relationship among message framing, perceived diagnosticity, and eWOM effectiveness, we conducted a mediation analysis using the Bootstrapping method (PROCESS Model 4) [81] to assess the role of perceived diagnosticity. Results confirmed that perceived diagnosticity mediated the relationship between message framing by virtual eWOM senders and eWOM effectiveness (95% confidence interval β = 0.16, CI = [0.03, 0.18]), supporting H2 (see Figure 1).

Experiment 2 verified the mediating role of perceived diagnosticity in the relationship between message framing and eWOM effectiveness. Experiment 3 will further examine whether product type moderates this mechanism, specifically exploring how product type influences the process by which message framing affects eWOM effectiveness through perceived diagnosticity.

3.3. Experiment 3

Experiment 3 aims to investigate the moderating role of product type in the relationship between message framing by virtual eWOM senders and eWOM effectiveness, thereby testing H3.

3.3.1. Participants

The experiment employed a 2 (product type: search products vs. experience products) × 2 (message framing: numerical vs. experiential) between-subjects design. We recruited 200 participants and randomly assigned them to four conditions to complete a series of product evaluation surveys. After excluding those who failed attention checks and manipulation checks, we retained 192 valid responses (ages 18–44, M = 24.71, SD = 5.33, with 52.80% female participants). The sample sizes for each condition were: n_{search,numerical} = 47, n_{search,experiential} = 49, n_{experience,numerical} = 48, and n_{experience,experiential} = 48.

3.3.2. Method

To ensure the effectiveness of our experimental manipulations, we recruited 135 participants from a university (ages 18–25, M = 23.84, SD = 1.92, with 57.10% female participants) for a pretest, randomly assigning them to a 2 × 2 experimental design to verify our message framing (numerical vs. experiential) and product type (search products vs. experience products) manipulations. We selected a smart sports watch (SmartWatch) as the search good and a premium chocolate gift box (Choco) as the experience good. Participants were randomly assigned to one of four experimental conditions and received product recommendations corresponding to their assigned product type and message framing. The two description styles for SmartWatch were the same as those used in Experiment 1. The two description styles for Choco were as follows:

Numerical description:

“Choco chocolate contains 72% cocoa ingredients, 12 flavors in individual packaging, with a net weight of 200 g per box. Sugar content is 5 g/100 g, with a three-layer structure (65% base + 25% filling + 10% coating). Packaging dimensions are 15 cm × 8 cm × 4 cm, with nitrogen-filled inner bag. Aroma intensity: 8.7/10, melting speed: 9.1/10. Highly recommended!”

Experiential description:

“Choco chocolate has a rich cocoa aroma, with distinctive multiple flavors, and the whole box feels substantial. The sweetness is just right, with three layers of flavor unfolding when you bite—deep base, smooth filling, and thin coating. The exquisite gift box keeps it fresh, and when opened, the aroma is enticing, with a soft, silky texture. Highly recommended!”

Participants then rated the numerical nature of the virtual eWOM sender’s description [74]. Additionally, participants evaluated the product type: “Search products are products or services whose quality can be evaluated using objective standards before purchase; experience products are products whose quality can only be accurately evaluated after trying or purchasing them” (7-point scale, 1 = search products, 7 = experience products) [70]. Results showed that the numerical condition scored significantly higher on numerical content than the experiential condition (M_numerical = 5.72, SD = 0.95; M_experiential = 2.32, SD = 1.30, t = 17.31, df = 133, p < 0.001, d = 2.979). The experiential condition scored significantly higher on experiential content than the numerical condition (M_numerical = 2.91, SD = 1.25; M_experiential = 4.93, SD = 1.51, t = 8.44, df = 133, p < 0.001, d = 1.453). The experience product condition scored significantly higher on the product type rating than the search product condition (M_experience = 5.29, SD = 1.45; M_search = 3.83, SD = 1.83, t = 5.12, df = 133, p < 0.001, d = 0.886), confirming the effectiveness of our message framing and product type manipulations in Experiment 3.

In the main experiment, participants were told the study aimed to investigate consumer attitudes toward new products and were asked to carefully evaluate a product review from a “virtual eWOM sender Nova”/“other reviewer”. To enhance experimental realism and consistency, all participants were guided to a uniform product usage scenario. Participants were asked to imagine they were considering purchasing the relevant product and told they had some basic product knowledge but hadn’t researched specific brands or models, thereby controlling for product involvement and perceived review quantity. Participants then read product recommendations corresponding to their condition: those in the search product condition read reviews about SmartWatch, while those in the experience product condition read reviews about Choco. At the same time, participants in the numerical condition saw product information described with objective data, while those in the experiential condition saw descriptions emphasizing subjective feelings and experiences, with specific content as described in the pretest. After the experiment, participants first rated the numerical and experiential nature of the virtual eWOM sender’s description [74] and evaluated the product type [70]. They then reported their perceived diagnosticity of the eWOM information and eWOM effectiveness. To further control for potential confounding variables, we measured product involvement [77], perceived review quantity [77], and participants’ emotional state [79]. Finally, participants were asked to guess the study’s purpose.

3.3.3. Results and Discussion

Manipulation checks: Seven participants failed attention checks, and one correctly guessed the study’s true purpose, resulting in eight participants being excluded. The numerical condition scored significantly higher on numerical content than the experiential condition (M_numerical = 5.67, SD = 1.42; M_experiential = 2.26, SD = 1.61, t = 15.60, df = 133, p < 0.001, d = 2.251). The experiential condition scored significantly higher on experiential content than the numerical condition (M_numerical = 2.65, SD = 1.41; M_experiential = 4.86, SD = 0.96, t = 12.63, df = 133, p < 0.001, d = 1.829). The experience product condition scored significantly higher on the product type rating than the search product condition (M_experience = 5.34, SD = 1.54; M_search = 3.31, SD = 2.12, t = 7.58, df = 190, p < 0.001, d = 1.095). These results confirmed the effectiveness of our message framing and product type manipulations in Experiment 3.

Perceived diagnosticity: Results showed a significant interaction effect between product type and message framing on perceived diagnosticity (F = 16.23, p < 0.001). For search products, the numerical condition yielded significantly higher perceived diagnosticity than the experiential condition (M_numerical = 5.83, SD = 0.82, M_experiential = 5.14, SD = 0.82; t = 4.34, df = 190, p < 0.001, d = 0.882). For experience products, no significant difference was found in perceived diagnosticity between numerical and experiential descriptions (M_numerical = 4.92, SD = 0.77, M_experiential = 5.12, SD = 0.73; t = 1.26, df = 190, p = 0.210, d = 0.256). This demonstrates that numerical descriptions by virtual eWOM senders enhance perceived diagnosticity in search product contexts, while this effect does not exist in experience product contexts.

eWOM effectiveness: Results revealed a significant interaction effect between product type and message framing on eWOM effectiveness (F = 24.97, p < 0.001). For search products, the numerical condition yielded significantly higher eWOM effectiveness than the experiential condition (M_numerical = 5.61, SD = 0.63; M_experiential = 4.71, SD = 0.81; t = 5.60, df = 190, p < 0.001, d = 1.138). For experience products, no significant difference was found in eWOM effectiveness between the two conditions (M_numerical = 4.78, SD = 0.75; M_experiential = 5.03, SD = 0.92; t = 1.56, df = 190, p = 0.120, d = 0.317). These results support H3.

Control variables: We found no significant differences across the four conditions in perceived review quantity (F(3,188) = 0.25, p = 0.86; M_{search,numerical} = 4.74, SD = 0.76; M_{search,experiential} = 4.69, SD = 0.81; M_{experience,numerical} = 4.66, SD = 0.85; M_{experience,experiential} = 4.60, SD = 0.82), product involvement (F(3,188) = 0.39, p = 0.86; M_{search,numerical} = 4.99, SD = 0.90; M_{search,experiential} = 5.04, SD = 0.83; M_{experience,numerical} = 5.07, SD = 0.89; M_{experience,experiential} = 4.90, SD = 0.76), or emotional state (F(3,188) = 1.95, p = 0.12; M_{search,numerical} = 3.75, SD = 0.56; M_{search,experiential} = 3.84, SD = 0.68; M_{experience,numerical} = 4.03, SD = 0.57; M_{experience,experiential} = 3.82, SD = 0.59). This rules out potential interference from these factors on our experimental results.

Moderated mediation analysis: Using the Bootstrapping method (PROCESS Model 8) [81], we found that the interaction between message framing (numerical/experiential) by virtual eWOM senders and product type (search/experience) significantly influenced perceived diagnosticity (95% confidence interval β = 0.92; CI [0.47, 1.37]). Furthermore, perceived diagnosticity significantly influenced eWOM effectiveness (95% confidence interval β = 0.51; CI [0.39, 0.64]). For search products, different message framing by virtual eWOM senders significantly influenced eWOM effectiveness through perceived diagnosticity (conditional indirect effect, 95% confidence interval β = 0.354; CI [0.176, 0.556]). For experience products, the effect of message framing by virtual eWOM senders on eWOM effectiveness through perceived diagnosticity was not significant (conditional indirect effect, 95% confidence interval β = −0.117; CI [−0.284, 0.035]). In conclusion, product type effectively moderated the relationship between message framing by virtual eWOM senders and eWOM effectiveness, and this relationship was mediated by perceived diagnosticity (Index = 0.471; 95% Bootstrap CI [0.23, 0.75]). See Figure 2 for details.

4. Discussion

4.1. Conclusions

Through three experimental designs, this research investigated the influence of message framing (numerical vs. experiential) by virtual eWOM senders on eWOM effectiveness and analyzed the mediating role of perceived diagnosticity and the moderating effect of product type. The following are the conclusions of this research.

First, Experiment 1 validated our main effect hypothesis, showing that virtual eWOM senders using numerical descriptions significantly enhance eWOM effectiveness compared to those using experiential descriptions. Numerical descriptions convey product core features through objective data and quantitative indicators, while experiential descriptions emphasize subjective feelings and emotional experiences [69]. Numerical descriptions focus on objective data and facts, which align with consumers’ agent knowledge of artificial intelligence as a “rational, data-driven” entity, reducing perceived persuasion intent and thereby increasing information credibility and effectiveness. In contrast, experiential descriptions contradict consumers’ perception that AI lacks genuine experiences, thus triggering doubt and resistance. Furthermore, this research compared the effectiveness of different description frameworks between virtual and human eWOM senders, finding that the advantage of numerical descriptions is significant only in virtual eWOM sender contexts. For human eWOM senders, experiential descriptions can better resonate with consumers, thereby more effectively enhancing the effectiveness of eWOM communication.

Second, Experiment 2 revealed the mediating role of perceived diagnosticity in the process by which message framing of virtual eWOM senders influences eWOM effectiveness, validating H2. Numerical descriptions, due to their objective and verifiable characteristics, enhance consumers’ perceived diagnosticity of eWOM information, meaning they consider the information more effective in aiding decision-making. Information with high perceived diagnosticity is more likely to be adopted by consumers, thereby enhancing eWOM effectiveness. Experiential descriptions, however, due to their inconsistency with agent cognition, reduce perceived diagnosticity, weakening eWOM effectiveness.

Finally, Experiment 3 demonstrated that product type significantly moderates the influence of message framing by virtual eWOM senders on eWOM effectiveness, validating H3. Specifically, in search product contexts, where consumers have higher demand for objective attribute information, numerical descriptions highly align with consumer expectations for search product information, significantly enhancing consumer perceived diagnosticity and eWOM effectiveness. In experience product contexts, although consumers expect experiential information, their insufficient trust in virtual recommenders who lack genuine experiences means that virtual eWOM senders, whether using numerical or experiential descriptions, struggle to effectively enhance consumers’ perceived diagnosticity of eWOM information and consequently cannot effectively influence eWOM effectiveness.

4.2. Theoretical Implications

First, this research expands the research perspective on virtual eWOM senders, providing new insights into how AI-published content influences consumer decision processes. As marketing practices shift from unidirectional communication to bidirectional value co-creation and real-time interactive platform ecosystems [82], existing research on virtual influencers has primarily focused on factors affecting their market acceptance, such as credibility [35], anthropomorphism level [21], and social interaction capabilities [22], while rarely exploring how their textual expression strategies as “reviewers” influence consumer behavior. This research systematically compares for the first time the differential effects of numerical versus experiential description frameworks used by virtual eWOM senders, discussing their influence mechanisms and providing a novel theoretical framework for choosing communication strategies for virtual agents.

Second, this research innovatively extends the application of PKM to eWOM and virtual agent contexts. PKM theory originally focused on how consumers identify, evaluate, and respond to marketing strategies in traditional marketing scenarios. This research identifies virtual eWOM senders as essentially constituting a new type of persuasion scenario, integrating consumers’ common perceptions of artificial intelligence—its advantages in data processing and limitations in emotional experience—as core components of “agent knowledge”. It reveals how when message framing by virtual eWOM senders does not match consumers’ expectations of their attributes, it activates persuasion knowledge, triggering cognitive conflict, doubt, and resistance, thereby reducing perceived diagnosticity and eWOM effectiveness. This extends the applicable boundaries of PKM, providing a more refined theoretical explanation for understanding how consumers process and evaluate information in AI-mediated communication contexts.

Additionally, by introducing perceived diagnosticity as a mediating variable and examining the moderating role of product type, this research provides deeper theoretical insights into eWOM effectiveness mechanisms. Perceived diagnosticity as a key mediator explains why numerical descriptions can enhance virtual eWOM effectiveness. This occurs because numerical descriptions strengthen consumers’ perception of information usefulness and overall decision value. The moderating role of product type reveals boundary conditions for virtual eWOM senders’ description strategies. This research discusses the complex matching mechanism among information source (virtual eWOM senders), information content (numerical/experiential), and product characteristics (search products/experience products), enriching the theoretical framework of information processing and adoption in eWOM research.

5. Marketing Implications

This research provides important marketing implications and practical guidance for e-commerce companies using virtual eWOM senders for recommendations and product promotion. First, when selecting virtual eWOM senders for promotion, e-commerce companies should fully consider the match between their message framing and product characteristics, leveraging virtual eWOM senders’ advantages in processing and presenting objective, rational information. For example, virtual senders can create comparison tables that benchmark two or three core competing products in the market by providing detailed specifications and technical parameters. These tables should highlight the advantages of their own products while honestly acknowledging any disadvantages. This approach closely aligns with consumers’ information needs for search products, effectively enhances the perceived diagnosticity of eWOM information, and thus facilitates purchase decisions.

Second, e-commerce companies need to recognize virtual eWOM senders’ limitations in emotional expression and authentic experience. When promoting experience products, should Secondly, e-commerce companies must recognize the limitations of online virtual influencers in emotional expression and authentic experience sharing. When promoting experiential products, it is advisable to exercise caution in deploying virtual eWOM senders and instead adopt alternative marketing approaches to deliver experiential information. These may include redefining the virtual sender’s persona—for instance, positioning them as an “experience collector” rather than a direct product user—creating immersive short videos without overt promotional language, and using dynamic posters to visually showcase product texture. These approaches effectively compensate for the inherent shortcomings in authentic experience communication.

Third, this research emphasizes the key role of perceived diagnosticity in eWOM communication. When designing content for virtual eWOM senders, e-commerce companies should strive to enhance information diagnosticity, ensuring the information effectively helps consumers distinguish between different products, evaluate product attributes, and make decisions. This means content should not merely be general praise but should include specific, verifiable numerical information that practically assists consumer decision-making. By providing highly diagnostic numerical information, companies can effectively strengthen consumer trust and adoption intentions toward virtual eWOM senders.

Finally, e-commerce companies should strive to establish a dynamic and transparent virtual eWOM management system, develop ethical guidelines adapted to AI-generated content, and continuously monitor the effectiveness of different description strategies. Specifically, platforms should proactively label AI-generated reviews with clear identifiers—such as “AI-Generated” or “Virtual User Review”—as a key marketing tactic to enhance technical transparency and actively shape consumer expectations. This approach also provides a compliant and sustainable framework for leveraging AI in marketing innovation. In response to the evolving market landscape, companies should regularly analyze data and gather consumer feedback to flexibly adjust the content strategies of virtual eWOM senders. Through ongoing optimization, e-commerce companies can more effectively harness AI technologies to strengthen their influence and market competitiveness. This aligns with interactive marketing’s future development direction of providing more personalized, dynamic content based on technological advancements [82].

6. Limitations and Future Directions

This research provides new insights into how message framing strategies by virtual eWOM senders influence eWOM effectiveness, but several aspects warrant further exploration in future research. First, this research primarily focuses on two core message framing approaches by virtual eWOM senders: numerical descriptions and experiential descriptions. However, this design has limitations in ecological validity. Given that the study participants were all from China, the cultural generalizability of the findings may be constrained. Future research could extend to populations from diverse cultural backgrounds to examine the cross-cultural robustness of the findings. Additionally, in actual online review environments, the presentation of information is diverse and may also include comparative descriptions, storytelling descriptions, as well as reviews that contain both specific parameter data and integrate subjective usage impressions. Future research could refine the stimulus materials by incorporating videos or images featuring virtual influencers to better align with the content formats prevalent on mainstream social media platforms. Additionally, it would be valuable to explore the efficacy of a sequential presentation strategy—one that prioritizes quantifiable information initially to align with user cognition and build trust, followed by descriptions of product experience and emotional benefits. Investigating the differential effects of this approach compared to a singular framing would provide enterprises with more nuanced guidelines for content strategy.

Second, this research examined product type (search products vs. experience products) as an important moderating variable. Future research could further refine product classification, for example, based on product price, durability, technical complexity, or other dimensions, or consider differences in consumer product and information processing across different cultural backgrounds, to explore boundary conditions for virtual eWOM senders’ description strategies.

Third, this research primarily used experimental methods to verify the influence of description strategies on eWOM effectiveness, but real-world interactions between virtual eWOM senders and consumers are more complex and diverse. Future research could adopt longitudinal designs to examine the effects of repeated consumer exposure to AI-generated reviews, thereby investigating the long-term trajectory of its impact on consumer cognition and decision-making. Simultaneously, with the advancement of technologies such as multimodal virtual influencers and generative conversational agents, an important direction worthy of attention is how the deep integration of information source presentation and content generation logic reshapes the boundaries of consumer perception regarding information authenticity.

Overall, research on the effectiveness of virtual eWOM senders as an emerging marketing tool is still in its early stages. This study provides a foundation for understanding the effectiveness of virtual eWOM senders’ description strategies. Future research can expand on this basis to explore more dimensions, building more complete and refined theoretical frameworks to provide more targeted guidance for companies using virtual eWOM senders for marketing communications.