AI Digital Human Responsiveness and Consumer Purchase Intention: The Mediating Role of Trust

Abstract

This study investigates how AI-driven virtual anchors affect consumers’ purchase intentions by identifying their key attributes, underlying mechanisms, and configurational interplay. We integrate latent Dirichlet allocation (LDA), structural equation modeling (SEM), and fuzzy-set qualitative comparative analysis (fsQCA) into a unified methodological framework. Empirical evidence demonstrates that the public visibility of virtual anchors exerts a significant positive impact on purchase intention, whereas professionalism, responsiveness, and personalization primarily cultivate consumer pleasure and trust, yet exert limited direct influence on purchase decisions. Emotional states—arousal, pleasure, and trust—mediate the relationship between anchor characteristics and purchase intention. fsQCA further reveals that high purchase intention emerges when responsiveness serves as a necessary condition, trust operates as a pivotal hub, and arousal/pleasure function as emotional conduits; conversely, low purchase intention is chiefly attributable to deficiencies in visibility, responsiveness, and trust. By synthesizing the SOR (stimulus-organism-response) model with the PAD (Pleasure-Arousal-Dominance) emotion theory, this research extends theoretical insights into consumer behavior within e-commerce live-streaming contexts and provides actionable guidance for optimizing virtual anchor strategies, thereby advancing both standardization and innovation in the industry.

1. Introduction

Driven by rapid advances in Internet infrastructure, the live-streaming e-commerce sector has undergone three evolutionary phases: unregulated expansion, traffic wars among superstar anchors, and the current era of brand-controlled self-broadcasting. In 2022, generative-AI-powered virtual anchors suddenly emerged, injecting new vitality into live commerce. Customizable avatars, precise product narratives, and real-time interactivity jointly create an immersive shopping experience; 24 × 7 non-stop broadcasting extends sales windows, mitigates the risk of human-anchor scandals, and significantly reduces operating costs. The outstanding performance of JD Supermarket’s “AI procurement anchor” during targeted campaigns exemplifies this potential [1]. IDC’s latest report shows that China’s AI digital human market reached RMB 4.12 billion in 2024, an 85.3% year-on-year increase, providing robust market support for AI-driven live selling [2].

Yet practice has outpaced theory. Although prior studies have identified the positive effects of interactivity and avatar cuteness on purchase intention [3,4], they have not systematically examined how professionalism, visibility, responsiveness, and personalization influence consumer decisions through deeper psychological pathways. The abandonment of virtual anchors by several brands after trial runs underscores the urgency of clarifying these mechanisms. High-profile avatars such as “Liu Qiangdong” and “Luo Yonghao” quickly secure consumer trust and enhance conversion through expert explanations, natural interactions, and hyper-realistic designs [5]; however, technical shortcomings—robotic expressions and synthetic voices—remain. Advances in voice cloning and lip-sync are progressively resolving these issues, thereby reshaping the e-commerce ecosystem. Consequently, this paper poses three research questions:

(RQ1) What are the core attributes of AI digital human anchors that influence purchase intention?

(RQ2) Through which mechanisms do these attributes operate?

(RQ3) How do different attributes interact or substitute for one another?

To address these questions, we adopt the SOR-PAD integrative framework. The Pleasure-Arousal-Dominance (PAD) emotion theory decomposes emotional states into pleasure, arousal, and dominance [6], which align with the organism component of the Stimulus-Organism-Response (SOR) model [7]. We conceptualize professionalism, visibility, responsiveness, and personalization of AI digital human anchors as external stimuli that transmit through emotional states—pleasure, arousal, and trust—ultimately affecting purchase behavior. Existing literature predominantly focuses on direct effects of attributes on intention [8], leaving the chain of emotional mediation underexplored. By embedding PAD emotion theory, this study systematically unpacks these mediated pathways and fills the theoretical gap.

Our contributions are threefold. Methodologically, we chose the LDA-SEM-fsQCA method combination for this study because each method offers unique advantages and complements the others in processing data and analyzing problems. LDA can extract potential themes from large-scale text data, which helps us quickly understand consumers’ intuitive feedback and focus on AI digital human anchors, thereby guiding subsequent in-depth analysis. SEM excels at handling complex variable relationships and can verify the hypothetical paths between anchor features and purchase intentions. It enables us to clearly see the direct impact of each factor by verifying the hypothetical paths between anchor features and variables like purchase intent. Meanwhile, fsQCA can reveal the sufficiency and necessity of condition combinations for the results, helping us understand how different feature combinations influence purchase intention across contexts and thereby providing a more comprehensive grasp of the influencing mechanism. Theoretically, the cross-fertilization of SOR and PAD extends the boundaries of consumer-behavior research in live-streaming e-commerce. Practically, the findings provide actionable guidelines for merchants and designers to optimize avatar images and interactive effects, thereby enhancing purchase intention and brand loyalty.

2. Literature Review

2.1. Consumer Behavior in Live Shopping

The evolution of e-commerce can be conceptualized as a three-stage leap—from static image–text formats to real-person live streaming and, more recently, to AI digital human live commerce—each transition reshaping consumer behavior. In the traditional picture-and-text era, consumers relied on static images and product descriptions, lacked real-time interaction, and faced monotonous experiences, slow decisions, and pronounced information asymmetry [9]. Live streaming overcame these limitations through immediacy and limited-time incentives, significantly elevating purchase incidence [9], yet scholarly consensus on the underlying triggers remains elusive. Liu et al. experimentally demonstrated that product presentation and anchor language directly affect purchase likelihood, but they did not unpack the role of anchor personality [10]. Meng Fei emphasized the centrality of charisma and professionalism, yet the precise pathways remain underspecified [11]. AI digital human live commerce represents the next frontier, amplifying the interplay between technology and context: 24 h uninterrupted broadcasting and algorithm-driven personalized recommendations markedly enhance convenience and human–avatar interaction depth4. Cai (2024) examined virtual anchors’ impact on impulse buying from a technology-availability lens [12], whereas Li et al. (2025) highlighted the mediating role of positive emotions [3]. In recent years, related research has continued to deepen. For example, Zhong et al. (2025) conducted an empirical study on the mediality of AI anchors and the construction of human–computer trust relationships, emphasizing the role of anthropomorphism and intelligence in building trust [1]. Li et al. (2025) explored the impact of host interactivity in live streaming on consumer purchasing behavior [3], and Sun and Zhu (2025) studied the mechanism of action of AI digital human live streaming on consumer purchase intention [4]. In practice, when stores are promoted off-site through activities such as collaborating with social media platforms for advertising or conducting promotional activities on other e-commerce platforms, consumers may be influenced by these external factors to visit the store for browsing and purchasing. This highlights the importance of understanding consumer behavior in the context of various promotional strategies. Although extant literature consistently underscores trust as a critical determinant of purchase decisions [3], the full psychological chain—“AI anchor attributes → emotional arousal → trust formation → purchase behavior”—has yet to be systematically elucidated, calling for further investigation.

2.2. AI Digital Human Live Streaming

As a convergence of artificial intelligence and virtual reality, digital-human live streaming is rapidly emerging as a new infrastructure for e-commerce. Its core value is reflected in three advantages. First, by rendering high-fidelity voice, micro-expressions, and body movements in real time, digital humans can replicate the nuanced interactions of human anchors, thereby enhancing viewer immersion and participation, and ultimately improving the shopping experience [4]. Second, leveraging algorithmic user profiling, digital humans deliver personalized product recommendations that precisely match individual needs and effectively increase conversion rates. Third, 24/7 uninterrupted broadcasting extends the reach window—particularly suitable for fast-moving consumer goods, science education, and tourism consulting—while operating costs are only 10% of those of human anchors, offering merchants a more economical option [13]. In addition, digital-human technology compensates for the emotional interaction deficit in traditional e-commerce; real-time feedback and emotional expressions shorten the psychological distance between brands and consumers, making brand images more vivid and stimulating purchase willingness [14]. Nevertheless, technological limitations such as templated responses and insufficient emotional expression may undermine streaming effectiveness, leading to polarized platform attitudes—JD.com and Baidu actively promote adoption, whereas Tencent Video imposes access restrictions [15]. Although industry revenue is projected to reach 640.27 billion yuan by 2025, most merchants’ performance falls short of expectations, and low consumer trust yields a return rate 30% higher than that of human-anchor live streams [16]. Consequently, identifying how digital-human anchors establish emotional connections with audiences and enhance trust has become a pivotal issue for improving live-streaming effectiveness.

2.3. SOR-PAD Theoretical Framework

The Stimulus–Organism–Response (SOR) framework posits that external stimuli (S) trigger internal psychological states (O), which in turn lead to behavioral responses (R) such as purchase decisions [17]. In consumer research, product attributes and marketing messages serve as S, eliciting cognitive, emotional, and physiological O that ultimately drive R. Features of AI digital human anchors can likewise be conceptualized as external stimuli: professional narrations enhance product understanding and interest, thereby increasing purchase intention, while Zhang Lu empirically confirms that interactive design and product displays in live streams significantly shape purchase pathways [18].

To dissect the underlying psychological mechanisms, we integrate the PAD emotion theory, which decomposes affect into pleasure, arousal, and dominance [7]. Weinberg et al. [19] demonstrate that pleasure and surprise during live shopping positively predict purchase intention; heightened arousal facilitates impulsive purchases, and elevated dominance strengthens perceived control and confidence. Subsequent studies reveal inter-dimensional interactions: Hui & Bateson [20] show dominance directly increases pleasure; Chang et al. [21] find dominance indirectly enhances pleasure via arousal; Koo & Lee [22] illustrate that trust—an expression of dominance—directly boosts arousal, which in turn elevates pleasure. These findings provide a theoretical basis for examining the interplay of the three PAD dimensions in AI digital human live streaming.

By synthesizing SOR and PAD, we construct a unified analytical framework in which professionalism, visibility, responsiveness, and personalization of AI digital human anchors function as external stimuli (S) that evoke consumers’ pleasure, arousal, and trust (O), ultimately influencing purchase intention and behavior (R). Professional anchors who deliver accurate and detailed information can increase viewer pleasure and trust; highly responsive anchors enhance interactive experience and perceived dominance; visible anchors more readily attract attention and credibility [21,23]. Chen [24] corroborates that anchor characteristics influence purchasing behavior through psychological mediation. Recent studies further demonstrate that digital humans’ emotional expressions, interactive styles, and personalized recommendations significantly affect emotional experience and purchase intention [3,4], offering direct support for our framework.

Nevertheless, two gaps remain. First, extant research predominantly focuses on surface attributes and direct effects, overlooking the mediating chain “attributes → emotions → intention.” Second, prior data collection relies heavily on retrospective surveys, failing to capture real-time emotional dynamics embedded in live-stream bullet comments. Addressing these gaps, the present study integrates SOR-PAD theory and employs a mixed-method design combining real-time bullet-screen text mining, questionnaire surveys, and fsQCA configurational analysis to explicate how AI digital human anchors influence purchase intention through emotional mechanisms, thereby enriching theoretical understanding and offering strategic guidance for live e-commerce practice.

3. Research Design

To rigorously address the research questions, we devised a three-stage mixed-methods workflow (see Figure 1). In Stage 1, we employed customized web crawlers to harvest real-time bullet-screen comments from AI digital-human live-stream shopping rooms across three major platforms. After de-duplication, de-emphasis, tokenisation, and stop-word removal, we applied latent Dirichlet allocation (LDA) to distill underlying topic dimensions from the cleaned corpus. In Stage 2, guided by the extracted topics and grounded in SOR-PAD theory, we constructed a structural equation model (SEM) and derived corresponding hypotheses. We then administered an online questionnaire to active consumers recruited from the private-domain communities of AI digital human anchors, ensuring high contextual relevance. The survey data were utilized to empirically test the hypothesized paths between anchor attributes, emotional mediators, and purchase intention. In Stage 3, we employed fuzzy-set qualitative comparative analysis (fsQCA) to examine configurational combinations of conditions that lead to high (or low) purchase intention, thereby complementing the SEM results with a set-theoretic perspective.

4. Study 1: LDA Thematic Modeling to Identify Key Variables

4.1. Data Sources

Study 1’s data comes from three popular Chinese e-commerce platforms: Jingdong, Meituan, and Baidu. We used web crawler technology to collect live-streaming pop-up-screen text from these platforms. As shown in Figure 2, the sample selection criteria included the live-room ID, the follower–live–broadcast–data–list ranking, AI–digital–person labeling, and pop-up–comment activity. We selected 25 AI digital–person anchors as research objects, as displayed in

Table 1. Sample of anchors (partial).

Platform Source	Anchor Name	Fan Volume	Number of Pop-ups (Bars)
Jingdong	JD Supermarket Procurement and Sales	5.473 million	5062
	Fresh JD Self operated Zone	5.982 million	2310
	ASIA SYMBOL JD Flagship Store	513 thousand	2210
	JD is really cheap live room	9.121 million	1835
	Jierou Official Flagship Store	776 thousand	1641
	Lenovo JD self operated flagship store	1.571 million	1524
	Okamoto Overseas JD.com	538 thousand	1658
	Tongrentang JD Flagship Store	4.217 million	1432
Baidu	BE FRIENDS HLDG	2.091 million	5241
	Jihui Jade Jewelry	11 thousand	1654
	Chinese Medicine Wang Wei	273 thousand	1568
	Angel Yeast	58,000	1235
	Chinese Medicine Zheng Yuqin	127 thousand	1521
Meituan	Zhengxin Chicken Chop	27 thousand	1547
	JD Car Maintenance	95 thousand	1374
	Yihetang Official Group Buying Live Room	64 thousand	1257

. These anchors have high fan volumes, high live-room interactivity, and strong sales ability, making them highly representative and typical. We collected pop-up texts from 52 live broadcasts, totaling over 41,000 items.

4.2. Data Pre-Processing

We conducted preprocessing operations on the collected AI digital human live pop-up texts to enhance data analysis accuracy and efficiency. First, we removed special characters and emoticons that might interfere with text analysis, such as “@ symbols” and emoticons, to make the text more concise. Second, we removed repetitive phrases and short texts, which often lack substantive information and are of limited use for theme extraction and sentiment analysis, e.g., removing repetitive phrases like “hahahahaha”. Next, we performed filtering based on text length, setting a two-character threshold to remove very short text, e.g., single words like “good”. Then, we used the Jieba segmentation tool to divide the text into independent vocabulary units. Finally, we removed deactivated words—common meaningless words—by creating a stop-word list and removing them from the text to focus on key information. After the above preprocessing, the text is cleaner and more useful, providing a solid foundation for subsequent topic modeling and sentiment analysis.

4.3. LDA Topic Modeling

The preprocessed text data was utilized for topic modeling. In this study, the LDA (Latent Dirichlet Allocation) model was employed for topic extraction. As a key tool for uncovering potential topics in large-scale text data, LDA can reveal hidden topic structures, making it highly suitable for analyzing complex text data, such as live pop-up texts.

Predetermining the number of topics is a crucial step when using LDA models. Although traditional methods primarily use perplexity and consistency to evaluate the reasonableness of the number of topics [25], the developers of the gensim library have questioned perplexity’s reliability. Therefore, this study adopted the local enumeration method to determine the optimal number of topics, aligning with the natural emergence principle of the rootedness theory [26,27]. In this study, pop-up comments served as initial coding material. The clustering effects of topic numbers from 1 to 15 were carefully compared, and 9 topics were ultimately determined to be optimal. After removing low-frequency words (frequency < 10), the model underwent 10 iterations of online training, completing model fitting and topic extraction. Specific topic classifications and their weights were derived. Finally, eight core categories were extracted through manual coding and theme fusion (see

Table 2. Pop-up text theme analysis results.

Artificial Coding		LDA Topics	Keywords
AI digital human anchor features	Professionalism	Anchor features	false, accent, Putonghua, AI human, digital human, AI haunted, mute, mechanical, lip-sync, clone
		Product explanation	reply, professional, interactive, great, doctor, endorsement, false live, counterpoint, simulation, memorized words, tone of voice
	Visibility	Brand stars	Liu Qiangdong, Luo Yonghao, Lao Luo, celebrities, Meituan, Jingdong, Telenor, Okamoto, popularity, idol, goddess
	Responsiveness	Comments and Replies	reply, please, beg to differ, how to buy, what, script, raw, expect, how much, please
	Personalization	Big data Recommendations	wanted, often bought, repurchased, worn, repurchased, have used, great, finally, big data, recommended, again, revisited, algorithms
Purchase emotion	Pleasure	Live streaming effect	events, raffles, games, red envelopes, red envelope rain, goody bags, eggs, interaction, satisfaction, humor
	Arousal	Product needs	explain, Maotai, milk powder, toothpaste, try, stock up, cheap, revisit, repurchase, special offer, experience, jacket
	Trust	Positive product Reviews	awesome, reassuring, authentic, tested, big name, tasty, good, strong, comfortable, responsible, recommended, assured
Purchase intention		Price offer and Purchase	grabbed, regret, open price, cheap, paid, bargain, still want, didn’t grab, add goods, support, have shot, can’t grab

).

To better observe the effectiveness of topic modeling, this study uses topic 9 as a case study and intuitively visualizes it using the pyLDAvis library. Figure 3’s left circular graph shows the distribution of different topics. The circle’s size and interval reflect the topics’ proportion in the corpus and their inter-topic similarity, respectively. The right bar chart lists the keywords under each topic. The column’s height and color indicate the word’s importance and its frequency in the corpus, respectively. By adjusting the parameter λ (0 ≤ λ ≤ 1) in the upper-right corner, the display of keyword–topic relevance can be dynamically adjusted. When λ approaches 0, it highlights topic-specific words; when λ approaches 1, it highlights high-frequency words. This visualization makes the topic modeling results more intuitive and effectively reveals consumer concerns in AI digital human live broadcasts [28,29].

5. Study 2: SEM Structural Equation Modeling

5.1. Questionnaire

5.1.1. Questionnaire Design

We designed a questionnaire to validate the theme-modeling results and further collect users’ perceptions and purchase intentions regarding these themes. The questionnaire has two parts. Part one covers eight latent variables: AI digital human anchors’ behavioral characteristics (professionalism, visibility, responsiveness, and personalization), users’ pleasure, arousal, and trust during live streaming, and purchase-intention factors. To ensure the scale’s reliability and validity, the measurement items are adapted from existing studies and mature domestic and international research scales, as shown in

Table 3. Scale Measures.

Variable	Items	Measurement Items	References
Professionalism	A1	This AI digital human anchor presented the product in a professional manner	Li Rong et al. [3]
	A2	The AI digital human anchor is familiar with the products she recommends
	A3	This AI digital human anchor delivers accurate and objective information about the product
Visibility	B1	This AI digital human anchor has high visibility and influence in the relevant field	Meng Fei [11]
	B2	I regularly watch the live streams of this AI digital human anchor
	B3	I follow this AI digital human anchor live because of its popularity
Responsiveness	C1	This AI digital human anchor can communicate with the audience at any time	Feng Runliu et al. [30]
	C2	This AI digital human anchor can respond to audience questions as quickly as possible
	C3	The AI digital human anchor’s responses are closely related to audience questions
Personalization	D1	This AI digital human anchor provides content that meets my personal needs and preferences	Sun Zimei et al. [4]
	D2	This AI digital human anchor will offer discounts or promotions that match my interests!
	D3	Recommendations from this AI digital human anchor improved my shopping experience
Arousal	E1	There is a lack of this product in your life and you want to know about buying it	Koo Ju [31]
	E2	Infected by the explanation of this AI digital human anchor, want to know more about the product
	E3	Want to know about products through user pop-up discussions
Pleasure	F1	Watching the live bandwagon process of this AI digital human anchor made me feel at ease	Donovan Rossiter [32]
	F2	Learning from this AI Digital Human Anchor brings me joy!
	F3	This AI digital human anchor was able to respond to and give feedback on questions in a timely manner, which made it enjoyable for me
Trust	G1	I chose to trust the product recommended by the AI digital human anchor because I liked it	Ridings [33]
	G2	I consider the AI digital human anchor to be representative of a particular field and choose to trust its recommendations
	G3	I chose to trust this AI digital human anchor because of the detailed, precise and vivid introduction of the product’s functionality.
Purchase intention	H1	I’m likely to buy products recommended by this AI digital human anchor	Wang Cuicui et al. [34]
	H2	I am willing to buy the product recommended by this AI digital human anchor if needed
	H3	I would like to purchase the product recommended by this AI Digital Human Anchor

. The performance of AI digital human anchors’ characteristics (professionalism, visibility, responsiveness, and personalization) refers to the studies of Li Rong [3], Meng Fei [11], Feng Runliu [30], and Sun Zimei [4], respectively. The emotional component’s arousal and pleasurable response design mainly refers to the studies of Koo and Ju [31], Donovan and Rossiter [32]. The trust measurement scale is from the study of Ridings [33]. The purchase intention scale refers to the study of Wang Cuicui et al. [34]. Part two of the questionnaire collects respondents’ basic information, including gender, age, and location, as shown in

Table 4. Descriptive statistics of demographic variables.

Variable	Option	Frequency	Percentage	Variable	Option	Frequency	Percentage
Gender	Male	203	50.12	Monthly income	≤1000 RMB	86	21.23
	Female	202	49.88		1000–3000 RMB	34	8.4
Age	≤18	85	20.99		3000–5000 RMB	126	31.11
	18–30	228	56.3		5000–10,000 RMB	111	27.41
	31–50	61	15.06		≥10,000 RMB	48	11.85
	≥50	31	7.65	Educational level	Below high school	71	17.53
Region	First-tier cities	160	39.51		High school	175	43.21
	Second-tier cities	128	31.6		Undergraduate/Junior college	129	31.85
	Other cities	83	20.49		Postgraduate and above	30	7.41
	Towns and villages	34	8.4	Careers	students	87	21.48
AI digital human live online shopping age	≤1 month	63	15.56		Professional staff	96	23.7
	1–6 months	160	39.51		Research worker	35	8.64
	6–12 months	104	25.68		Individuals/Freelancers	110	27.16
	≥1 year	78	19.26		Farmers	48	11.85
					Other	29	7.16

. The designed questionnaire was sent to e-commerce experts for review. After revising it according to their opinions, we formed the initial questionnaire. We then conducted a pre-survey with 50 respondents and revised it again to form the final questionnaire.

5.1.2. Questionnaire Data Collection

To ensure data relevance and accuracy, the questionnaire was distributed to members of the AI Digital People Host private-domain group, who are loyal users that have made purchases. During distribution, 450 questionnaires were distributed and all successfully recovered. To ensure data quality for the final analysis, questionnaires completed too quickly were excluded. Specifically, questionnaires completed in under one minute were deemed invalid due to potential impacts on data accuracy. After screening, 405 valid questionnaires were obtained, yielding a 90% validity rate. As shown in Table 4, the sample data exhibits youthfulness and urbanization, aligning with the characteristics of short-video users. According to AiMedia Consulting, among webcasting shoppers, 58% are men and 42% are women. Live e-commerce users are mainly in first- and second- tier cities, with the latter accounting for 42%. Live e-commerce also positively promotes the economic development of second- and third- tier cities. This study’s demographic data aligns with these characteristics, indicating sufficient reliability. To enhance study transparency and reproducibility, respondents’ demographic characteristics were recorded in detail, including age, gender, region, education level, occupation, and online-shopping age. This information aids in understanding behavioral differences among user groups in AI digital–human live shopping and offers valuable references for future research.

We chose the LDA model for its ability to reveal underlying thematic structures in large-scale text data, which is crucial for understanding user behavior and emotions during live broadcasts. SEM (Structural Equation Modeling) is a powerful statistical technique for validating complex variable relationships. SEM is a variance-based statistical analysis method that constructs models of relationships between latent and observed variables, enabling the simultaneous estimation of multiple equations to verify complex variable relationships. In this study, SEM is suitable for analyzing the causal paths between AI digital human anchor characteristics and variables such as consumer emotions and purchase intention because it can handle the relationships between latent variables (such as trust and pleasure, which are psychological states that cannot be directly observed) and observed variables (such as evaluations of the anchor’s professionalism, which can be measured through questionnaires), while also considering measurement errors to enhance the accuracy and reliability of the model. Combining these methods creates a comprehensive and in-depth analytical framework for the study.

5.2. Theoretical Hypotheses

Based on Study 1’s results, Study 2 innovatively combined the SOR model and PAD emotion theory to construct a comprehensive framework for analyzing how AI digital human anchor characteristics affect users’ purchasing behavior. Specifically, the PAD emotion theory suggests that pleasure, arousal, and trust indirectly affect users’ purchase intentions by influencing their internal mental states (the “organism” in the SOR model). To visualize this conceptual mapping, we created a detailed schematic diagram (Figure 4) that clearly shows the intrinsic connection between AI digital human anchor characteristics, emotional responses, and purchase behaviors. As external stimuli (the “stimulus” in the SOR model), these features effectively influence users’ emotional states. Emotional responses, including pleasure, arousal, and trust, align with the PAD emotion theory’s three core dimensions. They form the key part of the SOR model’s “organism,” reflecting the internal psychological changes triggered by AI digital human anchors’ features. Purchase behavior is reflected in users’ purchase intentions, including impulse and purposeful purchases. Both are unified as “reaction” elements in the SOR model. Through this mapping relationship, this study clearly reveals how AI digital human anchors’ features influence users’ emotional responses and indirectly shape their purchasing behaviors. This provides a comprehensive and systematic theoretical explanation for understanding consumer behaviors in AI digital human live streaming e-commerce.

5.2.1. Hypothesized Influence of AI Digital Human Information Source Characteristics on Consumption Intention (S-R)

As artificial intelligence becomes pervasive in live-commerce, AI-driven digital human anchors are emerging as pivotal determinants of consumer decision-making. A converging body of research indicates that four attributes—professionalism, visibility, responsiveness, and personalized recommendation—jointly exert a multidimensional influence on purchase intention [30,35].

First, professionalism reflects the anchor’s capacity to convey accurate, authoritative product information [3,30]. When digital avatars demonstrate expert knowledge and reliable information-transfer capabilities, consumers perceive heightened decision-making value, which fosters trust and increases willingness to interact. This utilitarian fulfillment, in turn, translates into a higher probability of purchase [36,37].

Second, visibility functions as a digital extension of social identity that reshapes trust formation. Empirical evidence shows that highly visible AI anchors attract disproportionate consumer attention, making endorsed products appear more credible and desirable [38]. For instance, JD.com’s “Liu Qiangdong” and Baidu’s “Luo Yonghao” avatars leveraged existing celebrity IPs to generate exponential surges in viewer attention and conversion rates [39]. Such visibility operates via two mechanisms: (a) the familiarity heuristic, where consumers exhibit an innate preference for recognizable figures, and (b) the industry-benchmark effect, whereby high-profile endorsements signal quality and reduce perceived risk [39].

Third, responsiveness plays a critical role in shaping real-time user experience [3,35]. Instantaneous feedback to consumer queries compresses perceived social distance in virtual space and heightens telepresence through anthropomorphic immediacy [3]. This rapid-response mechanism not only elevates perceived service responsiveness but also enhances both the utilitarian and emotional value of the transaction process [30,34,35].

Fourth, personalized recommendations, powered by deep-learning algorithms, are reconfiguring live-commerce logic. By mining behavioral trajectories, AI avatars construct precise demand profiles, shifting the paradigm from mass marketing to hyper-personalized suggestions. Such customization prolongs dwell time, cultivates brand loyalty, and boosts shopping satisfaction by catering to idiosyncratic preferences [39]. Illustratively, Liu Qiangdong’s digital avatar leveraged personalized recommendation engines to drive breakthrough improvements in user stickiness and conversion efficiency [39]. Collectively, these observations provide a solid theoretical and empirical foundation for the following hypotheses:

H1a: 

AI digital human professionalism significantly positively impacts user purchase intention.

H1b: 

AI digital human visibility significantly positively affects users’ purchase intention.

H1c: 

AI digital human responsiveness significantly positively affects users’ purchase intention.

H1d: 

AI digital human personalization significantly positively affects users’ purchase intention.

5.2.2. Hypothesized Influence of AI Digital Human on Purchase Intentions (S-O)

(1)

Influence of AI Digital Human Professionalism on Purchase Intentions

In the context of AI-driven live commerce, the professionalism of digital human anchors orchestrates a multi-layered emotional mechanism that channels cognitive appraisal into affective and conative outcomes. Positioned as the core dimension of consumers’ cognitive evaluation, professionalism operates through two sequential routes. First, the depth and authority of information stimulate cognitive attraction. When virtual anchors articulate accurate product knowledge and deliver rigorous explanations, viewers experience heightened arousal driven by curiosity and focused attention, thereby entering a state of intense cognitive engagement [3]. Although professionalism is primarily utilitarian, its indirect effect on pleasure emerges via perceived control: detailed and accurate content improves information-filtering efficiency and decision quality, generating intrinsic satisfaction that translates into subtle pleasure through a value-perception pathway [40]. This pleasure is not hedonic amusement but the quiet gratification derived from effectively meeting utilitarian goals [41]. Second, professionalism exerts a direct impact on trust formation. Recognition of professional competence naturally extends to the credibility of the recommendations themselves, encompassing both belief in information authenticity and authoritative endorsement of the anchor’s expertise [38]. When AI anchors consistently present logically coherent, data-supported narratives, their professional persona becomes a salient credibility cue within consumers’ decision-making schemata. Integrating these pathways, we posit that professionalism enhances arousal, indirectly fosters pleasure via perceived control, and directly cultivates trust. Accordingly, the following hypotheses are proposed:

H2a: 

AI digital human professionalism significantly positively affects arousal perception.

H3a: 

AI digital human professionalism significantly positively affects pleasure perception.

H4a: 

AI digital human professionalism significantly positively affects trust perception.

(2)

Influence of AI Digital Human visibility on Purchase Intentions

Within live-commerce environments, the visibility of AI-driven digital human functions as a multidimensional conduit that forges an emotional nexus between avatar and consumer. Acting as a digital proxy of social identity, visibility exerts a dual-path influence on psychological processing. First, on the cognitive route, the authoritative aura of well-known IPs or industry benchmarks captures selective attention and stimulates exploratory interest, thereby elevating arousal [8]. Consumers anticipate high-quality, scarce information from these high-profile sources, driving heightened engagement. Second, on the affective route, familiarity with the underlying persona elicits positive affective matching, enhancing pleasure [31]. When consumers’ existing knowledge structures align with the celebrated virtual anchor, emotional resonance emerges, amplifying perceived enjoyment. Crucially, visibility operates as a trust amplifier. Relative to ordinary anchors, high-visibility avatars are perceived as embodying industry consensus. During decision-making, consumers implicitly fuse the anchor’s social identity with credibility cues, ensuring both information authenticity and symbolic social endorsement [38]. Empirical cases—such as JD.com’s “Liu Qiangdong” and Baidu’s “Luo Yonghao” avatars—demonstrate how replicated real-person influence translates into traffic surges and elevated trust [38]. Integrating these cognitive–affective and credibility mechanisms, we advance the following hypotheses:

H2b: 

AI digital human visibility significantly positively affects arousal perception.

H3b: 

AI digital human visibility significantly positively affects pleasure perception.

H4b: 

AI digital human visibility significantly positively affects trust perception.

(3)

Influence of AI Digital Human responsiveness on Purchase Intentions

In live-streaming e-commerce, the responsiveness of AI digital humans constitutes a pivotal bridge between virtual service delivery and consumer affect. As a principal indicator of human–computer interaction quality, responsiveness establishes a distinctive emotional conduit through instantaneous feedback. When virtual hosts demonstrate rapid query resolution and maintain an efficient conversational cadence, they compress perceived social distance in the digital space and cultivate a heightened sense of telepresence, thereby immersing consumers within the consumption context [3]. This influence unfolds across three sequential stages. First, at the cognitive level, immediate feedback sustains conversational continuity, sustains attentional focus, and elevates arousal, thus keeping consumers in a state of readiness for decision-making [30]. Second, at the affective level, swift responses generate a perception of being valued; when consumers sense that their needs are promptly acknowledged, their emotional appraisal of the interaction becomes markedly positive [42]. Third, in terms of trust formation, response latency is inversely related to perceived decision certainty—prompt answers reduce information asymmetry and bolster a sense of control, which in turn enhances the reliability attributed to the recommender system [4]. Notably, emerging evidence suggests boundary conditions for these effects. While most studies confirm the beneficial emotional impact of responsiveness, scholars such as Li Feng caution that purely technical speed may not directly translate into purchase behavior unless it is coupled with interaction depth and content quality [3]. This implies that AI digital humans must transcend mere velocity by integrating emotional resonance and nuanced demand insight to achieve true interaction value sublimation. Drawing upon this theoretical foundation and empirical evidence, we propose the following hypotheses:

H2c: 

AI digital human responsiveness significantly positively affects arousal perception.

H3c: 

AI digital human responsiveness significantly positively affects pleasure perception.

H4c: 

AI digital human responsiveness significantly positively affects trust perception.

(4)

Influence of AI Digital Human personalization on Purchase Intentions

Within precision-marketing live commerce, the personalization capability of AI digital humans has emerged as a pivotal lever for reconfiguring consumer experiences. Powered by algorithmic profiling, this feature constructs a distinctive emotional-value chain by aligning content with individual preference maps. When virtual anchors deliver recommendations tailored to behavioral data, the mechanism simultaneously awakens latent demand and fosters an exclusive emotional bond. The process unfolds along three complementary pathways. Cognitively, accurate recommendations alleviate information overload, thereby heightening arousal; when consumers perceive that system-generated content mirrors their needs, the delight of being understood translates into sustained attention [43]. Affectively, customized interactions evoke a sense of exclusivity that amplifies pleasure; personalized attention signals self-worth confirmation and triggers positive emotional feedback [44]. Trust-wise, the degree of need–content congruence is tightly coupled with perceived system credibility; as recommendation accuracy repeatedly validates preference models, trust in the platform’s intelligence is incrementally reinforced [4]. Importantly, effective personalization transcends superficial attribute matching. As Cai et al. [12] emphasize, genuine personalization entails dynamic demand capture and creative experience design. When AI digital humans embed consumption scenarios and exceed expected customization, their influence shifts from instrumental utility to emotionally sticky relationships, cultivating a “being understood” psychological state that simultaneously satisfies functional requirements and builds affective trust [12]. Building on these theoretical premises, we propose the following hypotheses:

H2d: 

AI digital human personalization significantly positively affects arousal perception.

H3d: 

AI digital human personalization significantly positively affects pleasure perception.

H4d: 

AI digital human personalization significantly positively affects trust perception.

5.2.3. Buying Emotions: The Influential Relationship Between Arousal, Pleasure, and Trust (O)

Emotions—intensive affective states arising from cognitive appraisals of external stimuli—have received sustained scholarly attention as key antecedents of attitude and behavior [19]. Once activated, emotions permeate the individual’s mental state and prompt behavioral adjustment [45]. Jacob et al. demonstrate that external triggers can elicit emotional or cognitive reactions that steer individuals toward either approach or avoidance strategies [45]. Within consumer research, two dimensions—pleasure and arousal—are repeatedly identified as the primary affective responses to external stimuli [19]. Arousal denotes the level of physiological excitement elicited by a stimulus, whereas pleasure reflects the degree of happiness or satisfaction experienced [19]. Importantly, the two dimensions are positively correlated: elevated arousal typically amplifies perceived pleasure [43], and Koo et al. show that stronger arousal magnifies subsequent pleasure enhancement [31]. In online shopping contexts, Rose further documents that positive affect and satisfaction bolster perceived trust [46]. In AI-powered live-streaming commerce, the transmission and transformation of these emotions become even more salient. Li et al. reveal that AI anchors cultivate a warm atmosphere through humor and engaging activities, thereby significantly enhancing consumers’ pleasure [3]. This affective gain stems not only from interactive behaviors but also from the novelty and excitement inherent to the AI-mediated experience. Simultaneously, AI anchors leverage visibility, detailed demonstrations, and knowledge-rich narratives to strengthen perceived trust. A paradigmatic example occurred on 18 June 2025, when Luo Yonghao’s dual-digital-human broadcast combined distinctive presentation styles with expertise-based storytelling, deepening both pleasure and trust by forging knowledge-based connections between viewers and products [3]. Such elevated trust, in turn, propels purchase behavior. Drawing on these theoretical and empirical insights, we advance the following hypotheses:

H5a: 

Arousal perception significantly positively affects consumer pleasure perception.

H5b: 

Pleasure perception significantly positively affects consumer trust perception.

5.2.4. Hypothesized Influence of Buying Sentiment on Purchase Intention (O-R)

In AI-driven live commerce, purchase-oriented emotions operate as the critical psychological bridge that translates virtual interactions into concrete buying decisions. Extant research indicates that, within live-shopping contexts, emotions exert a dual-amplified influence on purchase intention through a sequential chain linking arousal, pleasure, and trust [31]. When AI anchors mediate this chain, the effect is further intensified by advanced technological affordances [47]. Specifically, technological augmentation manifests along three interrelated dimensions. First, arousal: AI anchors deploy dynamic visual stimuli and instantaneous feedback loops during activities such as virtual try-ons and situational demonstrations, significantly elevating consumers’ physiological arousal relative to traditional broadcasts [48]. This heightened arousal narrows attentional focus and, via dopaminergic pathways, increases impulsive purchase propensity [48]. Second, pleasure: algorithm-curated personalized content extends the duration of positive effect, thereby amplifying favorable product evaluations [31]. Third, trust: blockchain-enabled transparency and emotion-aware computing foster empathetic interactions, substantially compressing the trust-building cycle compared with conventional e-commerce [49]. When AI anchors maintain near-zero error rates and exhibit flawless professionalism, they surpass consumers’ trust thresholds and directly facilitate purchase decisions. Moreover, the technologically mediated transfer of emotion is uniquely salient in live-streaming environments. Drawing on Chen’s symbolic interactionism, the combined fidelity of 3-D avatar modeling, semantic accuracy of bullet-screen processing, and immersive scene design forms a complex system of digital symbols [50]. Within Hsu et al.’s website-quality framework, these symbols exert immediate effects on emotional arousal and perceived value [51]. Guided by these conduction mechanisms and empowerment features, we advance the following hypotheses:

H6a: 

Arousal perception significantly positively affects purchase intention.

H6b: 

Pleasure perception significantly positively affects purchase intention.

H6c: 

Trust perception significantly positively affects purchase intention.

5.3. Analysis of Empirical Results

5.3.1. Reliability and Validity Tests

In the reliability and validity assessment phase, the scale was rigorously evaluated with SPSS 23.0 and AMOS 26.0 to establish the credibility and precision of the collected data. First, internal-consistency reliability was quantified through Cronbach’s α, whose values range from 0 to 1, with higher figures denoting stronger item coherence [52]. All indicators exhibited standardized loadings greater than 0.70, indicating that the observed variables adequately captured their corresponding latent constructs (

Table 5. Measurement model reliability validity related indicators test.

Construct	Items	Estimate	Cronbach’s α	AVE	CR
Professionalism	A1	0.826	0.814	0.604	0.820
	A2	0.735
	A3	0.768
Visibility	B1	0.831	0.888	0.734	0.892
	B2	0.835
	B3	0.903
Responsiveness	C1	0.742	0.798	0.573	0.801
	C2	0.767
	C3	0.762
Personalization	D1	0.626	0.765	0.536	0.773
	D2	0.824
	D3	0.732
Arousal	E1	0.805	0.898	0.751	0.900
	E2	0.905
	E3	0.887
Pleasure	F1	0.815	0.861	0.680	0.865
	F2	0.795
	F3	0.864
Trust	G1	0.789	0.864	0.685	0.867
	G2	0.865
	G3	0.828
Purchase intention	H1	0.830	0.875	0.708	0.879
	H2	0.841
	H3	0.854

). Next, composite reliability (CR) and average variance extracted (AVE) were employed to evaluate scale reliability and convergent validity, respectively. In accordance with Fornell and Larcker (1981), the thresholds of CR > 0.60 and AVE > 0.50 were applied [53]. All constructs surpassed these benchmarks, confirming both dependable internal consistency and sufficient convergence at the construct level. Discriminant validity was then systematically examined to ensure that each latent variable represents a unique conceptual domain. Following Fornell and Larcker’s (1981) criterion, the square root of each construct’s AVE was compared with the correlations among constructs [53].

Table 6. Discriminant validity analysis.

Construct	Professionalism	Visibility	Responsiveness	Personalization	Arousal	Pleasure	Trust	Purchase Intention
Professionalism	0.777
Visibility	0.236	0.857
Responsiveness	0.231	0.265	0.757
Personalization	0.28	0.244	0.207	0.732
Arousal	0.404	0.471	0.433	0.376	0.867
Pleasure	0.316	0.434	0.532	0.308	0.656	0.825
Trust	0.365	0.434	0.331	0.252	0.419	0.564	0.828
Purchase intention	0.353	0.506	0.399	0.333	0.547	0.691	0.705	0.841

presents the results: diagonal elements display the square root of AVE, whereas off-diagonal cells report inter-construct correlation coefficients. In every instance, the square root of AVE exceeded the shared variance between constructs, thereby providing robust evidence of discriminant validity. Collectively, the empirical evidence satisfies all recommended reliability and validity thresholds, demonstrating that the measurement instrument is both theoretically and statistically sound. These findings substantiate the appropriateness of the scale for subsequent hypothesis testing and enhance confidence in the accuracy of the research conclusions.

5.3.2. Model Fit Testing

To evaluate the hypothesized relationships, a structural equation model (SEM) was estimated via maximum-likelihood estimation in AMOS 26.0. The adequacy of model fit was assessed using a comprehensive set of absolute and incremental fit indices. As reported in

Table 7. Model fitness test results.

Fitness Index	χ2/df	GFI	AGFI	CFI	NFI	RMSEA
Suggested	<3	>0.90	>0.80	>0.90	>0.90	<0.08
Actual	1.924	0.914	0.889	0.961	0.922	0.048

, the χ²/df ratio reached 1.924, which lies within the recommended range of 1.00–3.00. The absolute fit indices also indicated satisfactory correspondence between the hypothesized structure and the observed data: the Goodness-of-Fit Index (GFI) was 0.914 and the Adjusted Goodness-of-Fit Index (AGFI) was 0.889, both surpassing the conventional threshold of 0.90. Incremental fit indices corroborated these results, with the Comparative Fit Index (CFI) at 0.961 and the Normed Fit Index (NFI) at 0.922, exceeding the 0.90 benchmark. Finally, the Root Mean Square Error of Approximation (RMSEA) was 0.048, well below the 0.08 cut-off that signifies acceptable approximation error. Collectively, these diagnostics demonstrate that the proposed model achieves an excellent fit to the data, thereby providing a statistically sound foundation for testing the hypothesized effects of AI digital human anchors’ characteristics on users’ purchase behavior.

(1)

Direct effect test

Results summarized in

Table 8. Results of the fitted path test.

Hypothetical Path	Standardized Coefficient	S.E.	C.R.	p	Results
H1a: Professionalism→Purchase intention	0.032	0.038	0.693	0.489	rejection
H1b: Visibility→Purchase intention	0.145	0.037	2.975	0.003	support
H1c: Responsiveness→Purchase intention	0.012	0.054	0.227	0.82	rejection
H1d: Personalization→Purchase intention	0.073	0.048	1.593	0.111	rejection
H2a: Professionalism→Arousal	0.256	0.044	4.921	***	support
H3a: Professionalism→Pleasure	0.053	0.048	1.048	0.295	rejection
H4a: Professionalism→Trust	0.185	0.044	3.473	***	support
H2b: Visibility→Arousal	0.353	0.04	6.892	***	support
H3b: Visibility→Pleasure	0.16	0.046	3.101	0.002	support
H4b: Visibility→Trust	0.226	0.042	4.115	***	support
H2c: Responsiveness→Arousal	0.296	0.058	5.475	***	support
H3c: Responsiveness→Pleasure	0.304	0.067	5.466	***	support
H4c: Responsiveness→Trust	0.035	0.064	0.568	0.57	rejection
H2d: Personalization→Arousal	0.223	0.057	4.126	***	support
H3d: Personalization→Pleasure	0.057	0.061	1.114	0.265	rejection
H4d: Personalization→Trust	0.041	0.055	0.764	0.445	rejection
H5a: Arousal→Pleasure	0.43	0.072	6.776	***	support
H5b: Pleasure→Trust	0.387	0.06	5.573	***	support
H6a: Arousal→Purchase intention	0.052	0.058	0.876	0.381	rejection
H6b: Pleasure→Purchase intention	0.336	0.06	4.866	***	support
H6c: Trust→Purchase intention	0.415	0.058	7.145	***	support

Note: *** indicates p significant at 0.001.

reveal a nuanced nomological network in which the remaining hypotheses are supported, whereas the postulated direct paths from professionalism, responsiveness and personalization to purchase intention are non-significant. Furthermore, professionalism and personalization exert no direct influence on either pleasure or trust perceptions, contrary to initial expectations. Specifically, visibility retains a positive and significant direct effect on purchase intention (H1b, β = 0.145, p = 0.003), corroborating its role as a salient visual cue in consumer decision-making. Responsiveness exerts an indirect influence on purchase intention exclusively via pleasure perception (H3c, β = 0.304, p < 0.001), underscoring the importance of instantaneous interactivity in cultivating affective engagement. In turn, pleasure perception significantly reinforces trust perception (H5b, β = 0.387, p < 0.001), thereby positioning emotional satisfaction as a pivotal antecedent of trust formation. Trust perception subsequently emerges as the strongest predictor of purchase intention (H6c, β = 0.415, p < 0.001), highlighting its centrality in the decision-making process. Although arousal exerts a pronounced positive effect on pleasure perception (H5a, β = 0.430, p < 0.001), its direct influence on purchase intention is negligible (H6a, β = 0.052, p = 0.381), suggesting that heightened emotional activation must be channeled through intervening affective and cognitive variables to translate into behavioral outcomes. These insights refine theoretical accounts of AI digital human anchors by delineating the mediating mechanisms that link technological characteristics to consumer behavior, and offer actionable guidance for optimizing interactive strategies and conversion efficiency in AI-mediated live streaming contexts.

(2)

Mediation effects test

To clarify the psychological mechanisms by which AI digital human anchors’ characteristics translate into purchase intention, this study employed bias-corrected bootstrap mediation analyses (5000 resamples) to construct 95% confidence intervals (CIs) for all indirect effects [54]. Four anchor characteristics—professionalism, visibility, responsiveness, and personalization—were integrated into a multiple-mediator model with pleasure and trust as parallel mediators. As reported in

Table 9. Results of the mediation effect test.

Model	Paths	Effect Value	SE	Bias-Corrected 95%CI
				Lower	Upper
Mediation model 1 (Professionalism)	DE1: Professionalism→Purchase intention	0.039	0.039	−0.038	0.116
	IE1: Professionalism→Arousal→Purchase intention	0.017	0.014	−0.006	0.050
	IE2: Professionalism→Pleasure→Purchase intention	0.014	0.015	−0.012	0.048
	IE3: Professionalism→Trust→Purchase intention	0.059	0.019	0.022	0.095
	IE4: Professionalism→Arousal→Pleasure→Purchase intention	0.026	0.008	0.012	0.044
	IE5: Professionalism→Pleasure→Trust→Purchase intention	0.006	0.006	−0.005	0.02
	IE6: Professionalism→Arousal→Pleasure→Trust→Purchase intention	0.011	0.004	0.005	0.019
Mediation model 2 (Visibility)	DE2: Visibility→Purchase intention	0.13	0.037	0.058	0.203
	IE7: Visibility→Arousal→Purchase intention	0.022	0.018	−0.009	0.06
	IE8: Visibility→Pleasure→Purchase intention	0.037	0.017	0.011	0.077
	IE9: Visibility→Trust→Purchase intention	0.066	0.024	0.027	0.122
	IE10: Visibility→Arousal→Pleasure→Purchase intention	0.033	0.011	0.018	0.06
	IE11: Visibility→Pleasure→Trust→Purchase intention	0.015	0.007	0.004	0.031
	IE12: Visibility→Arousal→Pleasure→Trust→Purchase intention	0.014	0.005	0.007	0.025
Mediation model 3 (Responsiveness)	DE3: Responsiveness→Purchase intention	0.031	0.042	−0.052	0.113
	IE13: Responsiveness→Arousal→Purchase intention	0.02	0.015	−0.007	0.052
	IE14: Responsiveness→Pleasure→Purchase intention	0.079	0.018	0.04	0.112
	IE15: Responsiveness→Trust→Purchase intention	0.016	0.018	−0.019	0.052
	IE16: Responsiveness→Arousal→Pleasure→purchase intention	0.031	0.01	0.013	0.05
	IE17: Responsiveness→Pleasure→Trust→Purchase intention	0.032	0.009	0.014	0.049
	IE18: Responsiveness→Arousal→Pleasure→Trust→Purchase intention	0.013	0.004	0.005	0.02
Mediation model 4 (Personalization)	DE4: Personalization→Purchase intention	0.006	0.039	−0.012	0.139
	IE19: Personalization→Arousal→Purchase intention	0.014	0.011	−0.005	0.038
	IE20: Personalization→Pleasure→Purchase intention	0.013	0.013	−0.012	0.04
	IE21: Personalization→Trust→Purchase intention	0.019	0.018	−0.014	0.056
	IE22: Personalization→Arousal→Pleasure→Purchase intention	0.021	0.007	0.009	0.035
	IE23: Personalization→Pleasure→Trust→Purchase intention	0.005	0.005	−0.006	0.016
	IE24: Personalization→Arousal→Pleasure→Trust→Purchase intention	0.009	0.003	0.003	0.015

, visibility exerts both a direct effect on purchase intention (β = 0.13, 95% CI [0.058, 0.203]) and significant indirect effects via pleasure (β = 0.037, 95% CI [0.011, 0.077]) and trust (β = 0.066, 95% CI [0.027, 0.122]), indicating partial mediation. By contrast, professionalism, responsiveness, and personalization display no significant direct paths to purchase intention, yet their indirect effects remain substantial—through trust (β = 0.059, 95% CI [0.022, 0.095]) and pleasure (β = 0.079, 95% CI [0.040, 0.112])—approaching full mediation. These divergent patterns imply that visibility, as a salient epiphenomenal cue, can stimulate purchase decisions immediately, whereas professionalism, responsiveness, and personalization must first cultivate emotional resonance and trust before influencing behavior. Practically, firms should leverage high-visibility anchors to attract immediate attention, while simultaneously refining content professionalism, real-time responsiveness, and personalized services to systematically shepherd viewers through the sequential stages of interest arousal, pleasure creation, and trust consolidation, thereby maximizing live-stream marketing conversion.

6. Study 3: fsQCA-Based Purchase Intention Configuration Analysis

The SEM analysis results show that AI digital human-assisted selling is influenced by multiple factors. However, as Table 8 shows, the standardized coefficients between factors, except for H5a and H6c, are below 0.4, indicating a weak linear relationship between variables. Considering the causal complexity between independent and dependent variables, traditional linear regression cannot measure how multiple independent variables interact in complex combinations to affect the dependent variable. Therefore, this study used the fsQCA method to further analyze the data and explore the combination effects among variables.

6.1. Variable Selection and Data Calibration

In fuzzy-set qualitative comparative analysis (fsQCA), raw data must be calibrated into set membership scores that range from 0 to 1 [55]. Guided by the topic dimensions extracted via LDA in Study 1, we selected professionalism, visibility, responsiveness, personalization, arousal, pleasure, and trust as antecedent conditions and applied a uniform calibration strategy to the dependent variable (purchase intention) and all independent variables. For each construct, Likert-type items were first summed to create composite raw scores. Following Ragin’s procedure [56], the 95th, 50th, and 5th percentiles of the pooled distribution were specified as the anchors for full membership (1), the crossover point (0.5), and full non-membership (0), respectively. These anchors were then entered into the Calibrate function within the fsQCA 3.0 software to transform the raw scores into continuous fuzzy-set memberships. The resulting calibration thresholds are reported in

Table 10. Calibration anchors for variable data.

Variable		Calibration Anchors
		Full Membership	Cross-Over	Full Non- Membership
Independent variable	Professionalism	15	12	4
	Visibility	15	13	4
	Responsiveness	14	12	4
	Personalization	15	13	4
	Arousal	15	13	4
	Pleasure	15	13	4.3
	Trust	15	12	4
Implicit variable	Purchase intention	15	12	4

[56]. Although some scholars adjust the crossover point by ±0.001 to avoid simultaneous subset relations, such manipulation can introduce bias in large samples. Therefore, the original crossover value of 0.500 was retained without modification.

6.2. Necessary Conditions Analysis

Prior to conducting configurational analyses, the necessity of each individual condition must be ascertained [57]. A condition is deemed necessary if it is consistently present whenever the outcome is present, i.e., it represents a superset of the outcome set. Using the calibrated data, we examined the necessity of professionalism, visibility, responsiveness, personalization, arousal, pleasure, and trust for both high and low levels of purchase intention in fsQCA 3.0. Consistency scores serve as the primary diagnostic; values ≥ 0.90 are typically interpreted as evidence of necessity [57,58]. As shown in

Table 11. Necessary Conditions Analysis.

Variable	High Purchase Intention		Low Purchase Intention
	Consistency	Coverage	Consistency	Coverage
Professionalism	0.7549	0.7507	0.6616	0.5364
~Professionalism	0.5338	0.6593	0.6925	0.6972
Visibility	0.7394	0.7772	0.6201	0.5314
~Visibility	0.5542	0.6415	0.7400	0.6983
Responsiveness	0.7078	0.7497	0.6223	0.5373
~Responsiveness	0.5632	0.6465	0.7101	0.6646
Personalization	0.7003	0.7365	0.6525	0.5594
~Personalization	0.5810	0.6723	0.6927	0.6533
Arousal	0.7774	0.7787	0.6167	0.5035
~Arousal	0.5044	0.6175	0.7290	0.7275
Pleasure	0.8220	0.8072	0.5930	0.4747
~Pleasure	0.4650	0.5836	0.7591	0.7766
Trust	0.8190	0.8523	0.5870	0.4979
~Trust	0.5175	0.6058	0.8258	0.7881

Note: ~ indicates that the condition does not exist.

, all consistency scores for the individual conditions—whether for high or low purchase intention—remain below the 0.90 threshold. Consequently, no single antecedent constitutes a necessary condition for the outcome, indicating limited standalone explanatory power. This finding necessitates an analysis of configurational combinations to uncover how the joint presence (or absence) of these conditions collectively influences purchase intention.

6.3. Conditional Configuration Analysis

In contrast to the preceding necessity assessment, the sufficiency analysis aims to uncover alternative configurations of conditions that consistently produce the outcome—i.e., “equifinal pathways” from a set-theoretic perspective. Sufficiency requires that the configuration set be a consistent subset of the outcome set. Consistency remains the primary diagnostic, yet its minimum acceptable value and computational formula differ from those used in necessity tests. Recent scholarship adopts thresholds of 0.75 [58,59] and 0.80 [60,61,62], while the Proportional Reduction in Inconsistency (PRI) should exceed 0.70 [63,64]. Frequency cut-offs are calibrated to sample size; for small-to-medium samples, a threshold of 1 suffices. With 405 original observations, we retained fsQCA 3.0 defaults: consistency ≥ 0.80, PRI ≥ 0.70, and frequency ≥ 2. The algorithm generates three solutions—parsimonious, complex, and intermediate. The intermediate solution balances parsimony with theoretical plausibility and is therefore reported; the parsimonious solution is used solely to identify core conditions. Following Fiss (2011), this paper uses ● to denote a core condition present in both the parsimonious and intermediate solutions, and ⊗ to denote a core absence. ○ indicates an edge condition’s presence, × its absence, and “space” means the condition can be present or absent [59].

Table 12. Configuration Analysis of Consumer Purchase Intention.

Path Configuration	High Purchase Intention							Low Purchase Intention
	Interaction-Driven		Multidimensional–Synergistic			Core-Focused		Experience-Broken
	H1a	H1b	H2a	H2b	H2c	H3a	H3b	NH1	NH2	NH3
Professionalism	○		○		●	×	×		×	×
Visibility		×		○	●	×	×	⊗	⊗	⊗
Responsiveness	●	●		●	●	×	×	⊗	⊗	⊗
Personalization		×	●	●	●	×	●			×
Arousal	●	●	●	●	×	×	●	⊗		⊗
Pleasure	●	●	●	●	×	●	●	⊗	⊗
Trust	●	●	●	●	●	●	×	⊗	⊗	⊗
Consistency	0.9588	0.9590	0.9621	0.9618	0.9597	0.9580	0.9629	0.9369	0.9505	0.9545
Raw consist	0.4614	0.2787	0.4590	0.3983	0.1941	0.1808	0.2036	0.3968	0.3629	0.3101
Unique coverage	0.0359	0.0058	0.0464	0.0247	0.0176	0.0117	0.0121	0.0664	0.0325	0.0266
Solution consistency	0.9359							0.9238
Solution coverage	0.6132							0.4559

Note: ● indicate core condition present, ○ indicates an edge condition’s presence, ⊗ indicate core absence, × indicate absence, “space” means the condition can be present or absent.

summarizes the results. Overall solution consistency reaches 0.9359 for high purchase intention and 0.9238 for low purchase intention, surpassing conventional benchmarks and attesting to the reliability of the configurations. Solution coverage—an indicator of empirical relevance—amounts to 61.32% for high purchase intention and 45.59% for low purchase intention, demonstrating strong explanatory power [57]. Consequently, seven configurations account for instances of high purchase intention, whereas three configurations explain low purchase intention. Guided by established typologies, these configurations were further categorized into four patterns: interaction-driven, multidimensional-synergistic, core-focused, and experience-broken, based on the distributional characteristics of their core conditions.

6.4. Analysis of Configuration Results

Three empirically dominant patterns emerge. First, the interaction-driven pattern leverages responsiveness as the sole core condition to achieve high purchase intention through single-dimension reinforcement. Second, the omnidimensional-synergistic pattern integrates responsiveness, visibility, pleasure, and trust in simultaneous resonance, creating a full-chain experiential trajectory. Third, the core-focused pattern relies on the minimal configuration of responsiveness and trust, demonstrating that a concise emotional-trust nexus can suffice when peripheral conditions are absent. These patterns jointly confirm the asymmetric causal nature of consumer behavior: high purchase intention does not demand the presence of all design elements but requires the synergistic alignment of a parsimonious set of core conditions. Consequently, the findings provide a configurational foundation for differentiated AI digital-marketing strategies and deepen theoretical understanding of group-level stimulation mechanisms within the S-O-R framework.

6.4.1. Interaction-Driven: Immediate Response Activates Purchase Intention

Configuration H1a–H1b identifies two functionally equivalent pathways in which responsiveness, arousal, pleasure, and trust emerge as core conditions, underscoring the pivotal role of real-time interaction. Within these pathways, AI digital human anchors provide instantaneous feedback that stimulates arousal and pleasure, which, in turn, foster cognitive trust and ultimately enhance purchase intention. This causal sequence corroborates the SEM findings (H1c), where responsiveness was shown to exert an indirect effect on purchase intention through arousal → pleasure → trust. Notably, professionalism, visibility, and personalization appear as peripheral conditions, implying that sustained dyadic interaction can compensate for their absence. Specifically, repeated immediate responses cultivate a psychological chain of familiarity → security → trust, thereby securing long-term consumer engagement. Empirical exemplars include (1) private-community and daily service contexts, where AI avatars maintain rapport via routine greetings and responsive Q&A, and (2) promotional contexts, where limited-time offers or stock alerts trigger transient yet potent pleasure spikes that activate immediate purchase impulses. Consequently, this configuration is best suited to high-frequency, low-involvement product categories, for which continuous interaction and affective engagement outweigh attribute-based differentiation.

6.4.2. Multidimensional–Synergistic: Full-Element Resonance and Emotion-Driven Trust

The multidimensional-synergistic configurations comprise three high-purchase-intention solutions whose consistencies exceed 0.95. Each configuration exhibits an asymmetric structure in which a parsimonious set of core conditions is complemented by peripheral ones, thereby confirming the non-linear complexity underlying AI digital human anchors’ influence on consumer behavior.

Configuration H2a identifies personalization, arousal, pleasure, and trust as core conditions, with professionalism relegated to a peripheral role. This pattern demonstrates that affect-laden, tailor-made experiences can compensate for perceived deficiencies in professional competence, yielding a “sense-driven” decision path in which emotional resonance overrides rational evaluation. Consequently, this configuration represents an emotion-dominant route to conversion.

Configuration H2b retains responsiveness, personalization, arousal, pleasure, and trust as core conditions while treating visibility as peripheral. By foregrounding real-time interaction and bespoke service, this pathway decouples purchase intention from the traditional celebrity effect, illustrating how small and medium-sized anchors can leverage algorithmic recommendations to achieve competitive parity.

Configuration H2c positions professionalism, visibility, responsiveness, personalization, and trust as core conditions, with arousal and pleasure serving as peripheral facilitators. This arrangement delineates a rational decision path wherein institutional authority and brand endorsement cultivate trust and secure conversion even in the absence of pronounced emotional stimulation, thereby extending the boundary conditions of the S-O-R framework.

Collectively, these configurations delineate three distinct yet complementary routes—emotional resonance, professional authority, and service differentiation—that jointly explain how AI digital anchors shape purchase intention. Emotional resonance exploits personality cues to activate consumers’ limbic systems, establishing an “emotion-supplements-cognition” compensatory mechanism. Professional authority synergizes institutional trust and brand credibility to stabilize product turnover. Service differentiation integrates real-time interaction and intelligent recommendation, offering small and medium-sized anchors a viable strategy for user retention. Across all three routes, responsiveness and personalization emerge as foundational capabilities, confirming their status as necessary but not sufficient conditions. The multi-factor resonance observed here not only supports the “affect → trust → behavior” chain posited by perceptual decision models but also constructs a “cognitive appraisal → direct conversion” channel characteristic of rational decision processes. These findings signal a paradigm shift in live e-commerce toward a “decentralization-recentralization” dynamic driven by technology-enabled anchors.

6.4.3. Core-Focused: Emotionally Precipitated Lightweight Triggering

Core-focused configurations are characterized by a parsimonious combination of antecedents. In configuration H3a, only pleasure and trust operate as core conditions; all remaining variables are peripheral. This configuration indicates that affective resonance alone can be sufficient to trigger purchase conversion, provided that AI digital human anchors reliably elicit positive emotional responses. Empirically, healing-oriented visual aesthetics have been shown to heighten pleasure perception, which, in turn, drives conversion. A prominent illustration is the virtual idol “Xiaobing”, whose nightly radio program delivers emotional companionship, fostering user stickiness and elevating click-through rates for associated merchandise. Configuration H3b extends H3a by incorporating light personalization as an additional peripheral condition while preserving pleasure and trust as cores. This arrangement establishes a “light interaction–strong memory” decision chain: minimal but consistent personalized cues—such as a humorous personality script—generate memorable micro-interactions that accumulate into sustained engagement. The virtual IP “Deer”, for example, integrates algorithmic transparency, social-proof gamification, and incremental micro-commitments to regulate dynamic arousal. This approach enhances long-term user value even in the absence of explicit trust signals, thereby offering an asset-light operational template for scalable AI commercialization [65,66].

6.4.4. Experience-Broken: Core Characteristics Missing, Purchase Intention Inhibited

The experience-broken configurations (NH1–NH3) are characterized by the systematic absence of key causal conditions. Overall solution consistency reaches 0.9238 and coverage 0.4559, indicating that more than 92% of low-purchase-intention cases are captured by these patterns. Across all three configurations, AI digital human anchors simultaneously lack visibility, responsiveness, and trust, while arousal and pleasure are likewise absent. This constellation disrupts the “cognition → interaction → trust” chain that is essential for consumer conversion. Specifically, the absence of visibility prevents initial attention capture, whereas the absence of real-time responsiveness generates consumer anxiety and dissatisfaction; together, these deficits erode trust formation. Consequently, even when the product itself is objectively superior, the failure to establish a viable human–commodity connection precipitates rejection. These findings underscore the strategic primacy of basic experience management: trust is a multi-dimensional asset that accrues over time, yet its collapse can be triggered by the failure of a single core condition.

7. Discussion

7.1. Interpretation of Results and Theoretical Connections

7.1.1. Discussion of SEM Results

(1)

AI Digital Human Anchors’ Characteristics and Their Impact on Purchase Intention (S-R)

Empirical Results: Visibility of AI digital human anchors significantly positively affects purchase intention (H1b; β = 0.145, p = 0.003); Professionalism (H1a; β = 0.032, p = 0.489), responsiveness (H1c; β = 0.012, p = 0.820), and personalization (H1d; β = 0.073, p = 0.111) do not exert significant direct effects on purchase intention.

Interpretation and Theoretical Implications: Empirical evidence indicates that the visibility of AI digital human anchors exerts a significant, positive effect on purchase intention (H1b), corroborating Huang et al.’s celebrity-effect transmission mechanism [67]. Within the SOR-PAD framework, visibility operates as a salient stimulus (S) that influences behavioral outcomes via two distinct routes. First, visibility leverages social influence to cultivate trust; for example, JD’s “Liu Qiangdong” avatar capitalized on the entrepreneur’s established reputation, simultaneously increasing viewership, dwell time, and order volume while activating consumers’ trust perception. Second, visibility can elicit impulsive arousal, thereby accelerating purchase behavior. Consistent with this dual-path logic, fsQCA results reveal that visibility functions as either a core or a peripheral condition in multidimensional configurations that lead to high purchase intention.

In contrast, professionalism (H1a), responsiveness (H1c), and personalization (H1d) failed to exert significant direct effects on purchase intention. We attribute this null finding to current technological constraints that limit the expressiveness of digital humans’ vocal tone and facial micro-expressions, thereby attenuating the emotional responses required to meet PAD (Pleasure–Arousal–Dominance) thresholds. Supporting this interpretation, user-retention during product explanations was 37% lower for AI anchors than for human streamers despite comparable traffic volumes [68]. fsQCA further indicates that professionalism and personalization can influence purchase intention indirectly when combined with other conditions, underscoring their role as contingent rather than dominant stimuli.

The “depersonalized” nature of AI anchors introduces an efficiency–experience paradox. On the one hand, depersonalization enables 24 h broadcasting and reduces labor costs, functioning as a continuous stimulus input in the SOR model. On the other hand, overly standardized content accelerates habituation; viewers’ average attention span drops to 15 min versus 42 min for human streamers [68]. Consequently, under the SOR-PAD lens, AI anchors’ current stimulus attributes appear better suited for brand-exposure objectives (i.e., awareness generation through sustained stimulation) than for deep conversion tasks that hinge on rich emotional engagement. Illustratively, JD Supermarket’s AI anchors elevated brand awareness by 230% without a commensurate lift in purchase conversion [39]. Future research should therefore investigate how advances in affective computing and personality modeling shift consumer responses along the PAD continuum, and the technological tipping point at which AI anchors transition from awareness catalysts to conversion engines within the SOR-PAD framework.

(2)

AI Digital Human Anchors’ Characteristics and Their Impact on Pleasure, Arousal, and Trust (S-O)

Empirical evidence demonstrates that professionalism (H2a), visibility (H2b), responsiveness (H2c), and personalization (H2d) all exert significant positive effects on arousal, which in turn enhances purchase intention. Correspondingly, fsQCA identifies responsiveness and personalization as recurrent core conditions that—together with arousal, pleasure, and trust—form sufficient configurations for high purchase intention. Specifically, AI anchors capture consumer attention by delivering accurate product knowledge and visually intuitive demonstrations, thereby elevating arousal and downstream intention [41]. Leveraging natural-language processing and affective computing, these avatars provide real-time, tailored recommendations that increase satisfaction and extend dwell time [12]. Their anthropomorphic design further satisfies curiosity and sustains engagement, reinforcing the arousal-intention linkage. Visibility (H3b) and responsiveness (H3c) also significantly enhance pleasure, corroborating Sun and Feng’s flow-theoretic argument that interactivity deepens immersion and purchase likelihood [4,30]. AI avatars sustain flow through uninterrupted service and prompt Q&A; in fsQCA, responsiveness and personalization again emerge as core drivers of pleasurable experiences. Professionalism (H4a) and visibility (H4b) positively influence trust perception (O), which, via PAD theory, elevates purchase intention (R). Consumers rely on professional avatars for credible information, while celebrity visibility confers additional trustworthiness [3,38]. Within multidimensional fsQCA solutions, professionalism and visibility function as core conditions that anchor trust-building pathways. Collectively, these findings underscore how professionalism, visibility, responsiveness, and personalization operate as stimuli that reliably activate arousal, pleasure, and trust, ultimately shaping purchase intention within the SOR-PAD framework.

(3)

Purchase Emotions: The Influence Among Arousal, Pleasure, and Trust (O)

Empirical Results: Arousal significantly positively influences pleasure (H5a; β = 0.430, p < 0.001); Pleasure significantly enhances trust (H5b; β = 0.387, p < 0.001).

Interpretation and Theoretical Implications: Empirical results demonstrate that arousal exerts a significant positive influence on pleasure (H5a), and pleasure, in turn, significantly enhances trust (H5b), corroborating the findings of Koo and Liu Yang [22,23]. Within AI digital human live-shopping contexts, avatars simultaneously transmit utilitarian product information and hedonic emotional value through real-time interaction. Tactical stimuli such as red envelopes and coupons first heighten arousal, subsequently generate pleasure, and ultimately foster trust and customer stickiness [8]. Although Wen et al. note that the magnitude of these effects may vary across consumption scenarios, the directional relationships among arousal, pleasure, and trust remain consistently positive, underscoring their reciprocal reinforcement [8]. This tripartite emotional sequence is further validated by the fsQCA configurations: both interaction-driven and multidimensional-synergistic solutions identify arousal, pleasure, and trust as core conditions that jointly drive high purchase intention. Illustratively, in AI-assisted agricultural live streams, the avatar positions the product origin as a backdrop and narrates farmers’ stories and cultivation processes, crafting an authentic and rustic atmosphere. Such emotionally resonant content simultaneously delivers pleasure and elevates consumer trust in the anchor and the product, thereby boosting purchase intention. These findings reaffirm the intricate interplay among the three PAD-model affective dimensions and highlight their collective role in shaping consumer psychology and behavior.

(4)

The Effects of Pleasure, Arousal, and Trust on Purchase Intention (O-R)

Grounded in SOR–PAD theory, the present study demonstrates that both pleasure (H6b) and trust (H6c) exert significant positive effects on purchase intention, corroborating the findings of Wen and Liu [8,23]. Within AI digital human live-streaming contexts, emotionally charged experiences induce a state of heightened focus, pleasure, and arousal, which in turn amplifies consumers’ online purchase propensity. Trust is particularly pivotal given the virtual and algorithmic nature of AI anchors; as trust increases, perceived transaction risk—stemming from information asymmetry, exaggerated claims, and uncertain product quality—diminishes, thereby elevating willingness to buy. Empirically, AI avatars cultivate this trust and pleasure through gamified interactions (e.g., countdown discounts, lucky draws, songs, and dances), which simultaneously entertain and immerse viewers, shifting decision-making from rational evaluation to affective impulse.

Mediation analyses further illuminate these mechanisms. The visibility model (Model 2) exhibits partial mediation, indicating that high-visibility avatars can directly trigger purchase behavior while also exerting indirect effects via arousal→ pleasure→trust. Such avatars leverage existing brand equity or IP recognition to generate swift social proof. Conversely, the professionalism (Model 1), responsiveness (Model 3), and personalization (Model 4) models display full mediation: their influence on purchase intention is entirely transmitted through emotional variables. Professionalism relies on authoritative elaboration to foster trust; responsiveness depends on real-time feedback to enhance pleasure and trust; personalization leverages precise recommendations to satisfy idiosyncratic needs, yet each mechanism ultimately requires emotional engagement.

fsQCA corroborates these nuanced pathways. Visibility appears as either a core or peripheral condition across interaction-driven and multidimensional-synergistic configurations, confirming its contextual flexibility. Professionalism and personalization, although not always central, indirectly shape purchase intention when combined with responsiveness, arousal, pleasure, or trust. This divergence can be attributed to (1) differential psychological mechanisms—high-visibility anchors capitalize on celebrity effects to bypass emotional mediation, whereas professionalism, responsiveness, and personalization must cultivate affective resonance—and (2) technological disparities: visibility is primarily contingent on branding and promotional strategies, while professionalism, responsiveness, and personalization demand sophisticated NLP, affective computing, and recommender-system support; when such technologies are immature, affective variables compensate for functional deficits [12,68].

7.1.2. Discussion of fsQCA Results

Configuration analysis of Study 3 unpacks how distinct constellations of AI digital human anchor features differentially influence consumers’ purchase intention. In the high-purchase intention segment, three qualitatively different solution paths emerge. First, the interaction-driven configuration demonstrates that high responsiveness, arousal, pleasure, and trust jointly constitute a sufficient condition for elevated purchase intention. This finding corroborates the view that real-time interactivity coupled with effective emotional elicitation is pivotal. Illustratively, JD Supermarket deployed the AI digital human anchor “Liu Qiangdong” to deliver 24 h interactive services, which markedly enhanced user stickiness and conversion rates. Second, the multi-dimensional synergistic configuration reveals that personalization, arousal, pleasure, and trust operate as an interconnected bundle to optimize the consumption experience. By generating tailored recommendations based on browsing history and embedding emotionally appealing design elements, AI digital human anchors can significantly elevate consumers’ willingness to purchase [4]. Third, the core-focused configuration indicates that when pleasure and trust are present, purchase intention can remain high even if other peripheral features are deficient. The AI digital human anchor “Xiaobing” exemplifies this pathway by forging deep emotional resonance through therapeutic conversational design and affective linguistic cues [66]. Conversely, within the low-purchase intention segment, the experience-disconnection configuration shows that the absence of visibility, responsiveness, and trust jointly constitutes a necessary condition for diminished purchase intention, underscoring the foundational role of these attributes. Taken together, the configurational findings underscore that AI digital human anchor design should prioritize the synergistic orchestration of feature bundles rather than isolated feature optimization. Practically, marketers must dynamically calibrate the feature composition of AI digital human anchors to align with target audience characteristics and contextual contingencies, thereby maximizing conversion potential. Future research is encouraged to further elucidate how alternative feature combinations perform across heterogeneous marketing contexts, thereby advancing the theoretical and practical frontier of AI-driven digital human marketing strategies.

7.1.3. Comparative Discussion of SEM and fsQCA Results

The present study integrates Structural Equation Modeling and fuzzy-set Qualitative Comparative Analysis to examine how artificial intelligence digital human anchor characteristics shape consumers’ purchase intention. The two methods offer distinct yet complementary insights that jointly illuminate the complexity of this phenomenon.

(1)

Analytical perspectives

Structural Equation Modeling tests linear causal linkages and quantifies the magnitude of direct and indirect effects within the predefined theoretical framework. It therefore delineates individual causal paths linking anchor characteristics to purchase intention. Conversely, fuzzy-set Qualitative Comparative Analysis evaluates the sufficiency and necessity of configuration patterns; it does not presuppose a single causal path and is explicitly designed to capture non-linear relationships and complex interactions among multiple conditions. For instance, Structural Equation Modeling results indicate that visibility exerts a significant positive direct effect on purchase intention, whereas the direct effects of professionalism, responsiveness, and personalization are non-significant. Fuzzy-set Qualitative Comparative Analysis group analysis further demonstrates that professionalism and personalization can indirectly influence purchase intention when combined with other antecedent conditions, even if they are not core conditions within a given configuration.

(2)

Variable relationships

Structural Equation Modeling reveals that emotional variables—namely arousal, pleasure, and trust—mediate the influence of certain anchor characteristics on purchase intention, with trust occupying a central position in the decision chain. Fuzzy-set Qualitative Comparative Analysis extends this insight by identifying multiple equifinal pathways. In high-purchase-intention configurations, responsiveness, arousal, pleasure, and trust consistently emerge as core conditions, whereas visibility and personalization may serve as either core or peripheral conditions depending on the specific configuration. In low-purchase-intention configurations, the absence of visibility, responsiveness, and trust is jointly sufficient to suppress purchase intention, underscoring their foundational roles.

(3)

Result complementarity

Structural Equation Modeling provides an in-depth understanding of individual causal pathways and is therefore well-suited for testing theoretically derived hypotheses. Fuzzy-set Qualitative Comparative Analysis, by contrast, demonstrates the existence of multiple configurational routes and emphasizes the synergistic effects of condition combinations. Integrating the two methodologies yields a more comprehensive account of how anchor characteristics influence purchase intention. Specifically, Structural Equation Modeling may underestimate the non-linear effects of professionalism, responsiveness, and personalization; fuzzy-set Qualitative Comparative Analysis compensates for this limitation by revealing their contingent relevance within specific configurations. This finding holds significant practical implications for complex digital environments such as live-streaming commerce. In actual operations, companies cannot rely solely on a single method to assess and optimize marketing strategies. Combining SEM and fsQCA methods allows for a more comprehensive understanding of the impact of AI digital human anchor characteristics on consumer purchase intention, thereby enabling the development of more targeted and effective marketing plans. For instance, by using SEM to identify key direct impact factors and then employing fsQCA to explore the combinatorial effects of these factors under different circumstances, businesses can achieve optimal marketing outcomes.

(4)

Practical implications

Structural Equation Modeling results suggest that firms should prioritize improving anchor visibility through celebrity endorsements and brand co-branding, while simultaneously refining professionalism and responsiveness to strengthen trust. Fuzzy-set Qualitative Comparative Analysis findings indicate that enterprises must dynamically tailor characteristic combinations to align with distinct marketing objectives and situational contexts. For scenarios emphasizing rapid trust building and emotional resonance, the combination of responsiveness, arousal, pleasure, and trust should be accentuated. In contexts oriented toward personalized service and differentiated competition, the combination of personalization, responsiveness, and pleasure is more appropriate.

7.2. Management Insights

7.2.1. Optimizing the Design of Artificial Intelligence Digital Human Anchors

The present study clarifies the critical design attributes of artificial intelligence digital human anchors and their configurational effects on consumers’ purchase intention. These insights provide actionable guidance for merchants and interface designers to refine avatar appearance, interaction modalities, and functional affordances. Specifically, firms should systematically enhance professionalism, visibility, responsiveness, and personalization. Visibility can be amplified by leveraging celebrity endorsement and brand co-branding effects, thereby elevating perceived credibility and purchase willingness. Concurrently, professionalism and responsiveness serve as pivotal trust-building mechanisms that attenuate consumer uncertainty in live-streaming contexts.

7.2.2. Crafting Differentiated Marketing Strategies

Leveraging the fuzzy-set qualitative comparative analysis-derived subgroup analyses, firms can dynamically tailor attribute bundles to match heterogeneous marketing objectives and situational contingencies. In trust-building and emotional-resonance scenarios, the optimal configuration comprises responsiveness, arousal, pleasure, and trust. Conversely, in contexts prioritizing personalized service and differentiated competition, the emphasis should shift to personalization, responsiveness, and pleasure. Such configurational flexibility empowers organizations to adapt to evolving market environments and divergent consumer needs.

7.2.3. Enhancing Live-Streaming Efficacy and Competitive Advantage

By integrating structural equation modeling and fuzzy-set qualitative comparative analysis findings, enterprises obtain a holistic understanding of the multiple equifinal pathways through which artificial intelligence digital human anchor attributes influence purchase intention. This knowledge enables the design of precise, context-sensitive marketing strategies that bolster live-streaming performance.

8. Research Limitations

This study has several limitations that need to be acknowledged. First, the empirical data were collected exclusively from live pop-ups and online questionnaires deployed on three Chinese e-commerce platforms—Jingdong, Meituan, and Baidu. This limits the generalizability of the findings to broader user populations. Additionally, the majority of questionnaire respondents were recruited from the private-domain communities of artificial intelligence digital human anchors. This sampling strategy may introduce self-selection bias and overestimate positive attitudes. Second, the measurement instruments exhibit inherent constraints. The scales employed to capture artificial intelligence digital human anchor characteristics and consumer emotions rely on self-reported Likert items. These are susceptible to subjective interpretation and inter-individual variability in evaluation standards. Consequently, these instruments may not fully capture the nuanced and dynamic nature of consumers’ inner feelings and emotional fluctuations. Third, the sample data were sourced from three specific Chinese e-commerce platforms. This may introduce geographical and cultural biases, thus restricting the universality of the study results. Fourth, the study primarily focused on a few key characteristics of AI digital human anchors. However, there may be other influencing factors in the actual live-streaming environment, such as the layout of the live-streaming scene and the display methods of products. The interactive effects of these factors merit further investigation. Lastly, the cross-sectional data used in this study cannot capture the dynamic changes in consumer behavior. Future research could consider employing longitudinal data to more deeply analyze the formation and change mechanisms of consumer purchase intention.

9. Conclusions

This study offers a comprehensive understanding of how AI digital human anchor characteristics influence consumer purchase intention using a combination of LDA, SEM, and fsQCA methods. The findings reveal the significant roles of various anchor attributes and their interplay in shaping consumer behavior. Specifically, while certain anchor characteristics like professionalism, responsiveness, and personalization have limited direct effects on purchase intention, they significantly influence consumer emotions such as pleasure and trust, which mediate their impact on purchase intention. For instance, the study shows that arousal, a key emotional state, significantly influences pleasure, which subsequently enhances trust. This trust formation is crucial for elevating consumers’ willingness to purchase. The mediation analysis further elucidates that arousal mediates the relationship between anchor characteristics like professionalism and personalization and purchase intention, while pleasure mediates the relationship between responsiveness and visibility and purchase intention. This detailed examination of the mediation pathways provides a more nuanced understanding of the underlying mechanisms.

Theoretically, this research extends the SOR-PAD framework and offers new insights into consumer behavior in live-streaming e-commerce. By integrating LDA topic modeling, SEM, and fsQCA, we have constructed a robust methodological framework that captures the thematic structure of consumer feedback and elucidates the complex relationships between anchor characteristics, emotional states, and purchase intention. This integrative approach enriches the existing literature on consumer behavior in the context of AI-driven live commerce.

Practically, the results offer actionable guidelines for optimizing AI digital human anchor strategies to enhance consumer engagement and purchase conversion. For instance, the findings suggest that while visibility is a critical factor in directly influencing purchase intention.

Future research should further explore the dynamic interactions between different factors and their long-term impact on consumer loyalty and brand reputation. Additionally, given the rapid advancements in AI technologies, future studies could investigate how emerging technologies can further enhance the effectiveness of AI digital human anchors in driving consumer behavior.