Measuring the Intelligence of Virtual Anchors in E-Commerce: Scale Development and Validation from a Human–Computer Interaction Perspective

Abstract

E-commerce virtual anchors’ live-streaming sales represent a critical intersection of virtual human technology and artificial intelligence (AI), paving the way for new developments in digital marketing. Recent advancements in AI have significantly enhanced the intelligence of virtual anchors, driving increased consumer acceptance. As intelligence is a key determinant of user adoption, this study employs a mixed-methods approach combining grounded theory and quantitative analysis to conceptualize, structure, and measure virtual anchor intelligence. The grounded theory results reveal that virtual anchor intelligence encompasses multiple capabilities—manifested in guidance, recognition, analysis, and feedback during human–computer interaction, which enable effective hosting and sales performance. Four core dimensions of intelligence are identified: guidance intelligence, recognition intelligence, analysis intelligence, and feedback intelligence. Following established protocols for scale development, we constructed and validated an 18-item measurement scale, demonstrating strong reliability and validity. Empirical findings indicate that while guidance intelligence exerts an adverse effect on consumer participation, the remaining dimensions have positive effects, mediated by perceived shopping value. This study provides a comprehensive framework for understanding and measuring virtual anchor intelligence, elucidating its underlying mechanisms. The findings lay a theoretical foundation for future research on e-commerce virtual anchors but also offer practical implications for optimizing live-streaming strategies and advancing the design of virtual humans.

1. Introduction

1.1. Practical Background

In live-streaming commerce, the three core factors shaping consumer behavior are people (anchors), products, and settings, with anchors being the most influential element [1]. The integration of AI technology with e-commerce can enhance consumer experience [2]. E-commerce virtual anchors in live-streaming represent a successful application of AI technology, marking the future of the industry [3], and are a key strategic focus for e-commerce platform operators. Compared to human anchors, virtual anchors offer greater stability, lower costs, and 24/7 availability, while also avoiding human resource challenges such as turnover and performance variability. Early virtual anchors, despite their highly human-like appearances, faced resistance from consumers and operators due to limited AI capabilities. This resistance highlighted a core issue: as AI entities, intelligence is essential for virtual anchors to excel in e-commerce. Fortunately, recent advances in AI have significantly enhanced the intelligence of virtual anchors. As a result, they can now not only mimic human appearances but also simulate human language, gestures, and expressions—achieving the status of true digital twins. On 16 April 2024, JD.com launched an AI-powered digital avatar of its founder, Richard Liu (also known as Qiangdong Liu), referred to as “Sales Manager Brother Dong.” This virtual anchor made its debut live on both the JD Home Appliance and Furniture and JD Supermarket live-streaming channels, showcasing sophisticated product presentation skills and natural, interactive communication with viewers. Intelligence, as a defining attribute of AI entities [4], refers to their capacity to adapt to and influence the environment through various behavioral competencies [5]. The degree of intelligence development serves as a benchmark for evaluating the advancement of AI systems, with intelligence regarded as their core attribute [6]. E-commerce virtual anchors inherently embody this form of intelligence, and the more effectively they simulate human-like abilities, the more capable they are of fostering consumer value co-creation [7]. A deeper investigation into the intelligence of e-commerce virtual anchors not only enriches the theoretical foundation of this emerging field but also offers valuable insights for industry practitioners seeking to optimize virtual anchor design and deployment strategies.

1.2. Theoretical Background

The level of intelligence determines the extent to which e-commerce virtual anchors can serve consumers effectively [8]. Academic research on the intelligence of e-commerce virtual anchors remains in its early stages. Prior studies on AI have employed three primary measurement approaches.

The first approach evaluates intelligence-related attributes to determine whether users perceive AI entities as intelligent. For instance, Ran (2022) [9] assessed government chatbots using three dimensions: cognitive capacity, consciousness, and emotional awareness.

The second approach examines specific intelligent capabilities. Abraham (2018) [10] evaluated AI system intelligence by measuring decision-making competence, operational adaptability, and knowledge generation.

The third approach provides a comprehensive assessment of intelligent performance. Huang (2018) [11] classified service AI into four dimensions: mechanical intelligence (learning and adaptability), analytical intelligence (logical reasoning and decision-making), intuitive intelligence (understanding), and empathetic intelligence (emotion recognition and communication).

These existing measurement tools; however, are insufficient for assessing the intelligence of e-commerce virtual anchors. First, assessing intelligence based solely on isolated attributes cannot capture the complexity of live-streaming scenarios involving virtual anchors. The objective of this study is not to evaluate intelligence, wisdom, or consciousness per se, but to identify the key dimensions of intelligence, develop a validated measurement scale, and provide actionable insights for industry application. Second, significant variations in intelligence are observed across different AI application contexts [12]. Although AI entities share certain foundational attributes, their intelligence implementations, performance outputs, and technical designs vary considerably according to the context and role-specific requirements. Finally, as AI technologies evolve—especially with the integration of large language models—it becomes essential to develop updated intelligence measurement scales tailored to the emerging use case of e-commerce virtual anchors.

While existing intelligence frameworks are not directly applicable to virtual anchors in e-commerce, they offer valuable foundational insights. This study synthesizes prior research with real-world practices in live-streaming e-commerce to construct a comprehensive framework that defines the conceptualization, dimensions, and measurement scale of virtual anchor intelligence. The aim is to systematically interpret and quantify the intelligent capabilities of AI-enabled virtual anchors, thereby laying the groundwork for future research and providing practical guidance for e-commerce and virtual human applications.

2. Literature Review

2.1. Intelligence

Gottfredson (1997) [13] defined intelligence as the capability to reason, plan, solve problems, think abstractly, comprehend complex ideas, learn quickly, and acquire knowledge through experience. In the context of AI agents, intelligence refers to their capacity to acquire and apply knowledge and skills using AI technologies [14].

Prior research has examined the intelligence dimensions of virtual humans through their task performance and human–computer interaction capabilities. For instance, Lee (2007) [15] emphasized that the intelligence of virtual humans primarily manifests in human–computer interaction, encompassing cognitive, communicative, and emotional competencies that facilitate effective user engagement. Conversely, Swartout (2010) [16] argued that virtual humans display intelligence mainly through task-driven capabilities, including autonomy, communication, complex reasoning, and behavioral execution. Other scholars have explored cognitive dimensions in specific applications, such as the information retrieval and media management capabilities [17] of corporate virtual employees and the consultation abilities of human–machine virtual customer service systems [18].

Three key insights emerge from the existing literature on the intelligence dimensions of virtual humans. First, as emerging labor resources, virtual humans are widely implemented in various practical contexts, with intelligence traits closely tied to their role identities. Second, virtual humans display several common intelligence dimensions, including interaction, communication, comprehension, recognition, and expression. Third, the intelligence of virtual humans integrates multiple cognitive and behavioral competencies. However, the existing research still lacks a comprehensive examination of intelligence dimensions tailored to specific application settings. For example, Huang (2010) [19] classified the intelligence of entertainment-focused virtual humans into three dimensions—entertainment ability, emotional recognition, and autonomy—but omitted essential attributes such as communication and interaction. This exclusion is problematic, as entertainment-oriented virtual humans must interact dynamically with audiences. Therefore, this study not only incorporates established intelligence frameworks but also considers the role-specific and scenario-based characteristics of virtual humans in e-commerce live-streaming contexts.

2.2. Cognitive Theory

Cognitive theory seeks to construct models of human cognition and ultimately replicate human intelligence using AI. The core aim of AI research is to translate human cognitive behaviors into machine-replicable intelligent actions [20], thereby making intelligence the fundamental attribute of AI systems [4]. The core of AI research lies in transforming human cognitive behaviors into machine-replicable intelligent actions [20]. Newell’s (1976) [21] proposed that simulating human cognition is essential for machines to acquire intelligence. By modeling human mental processes, AI systems can enhance the efficiency of human–machine interaction [22], a process that relies on the computational processing of information [23]. Wilson (2002) [24] further argued that cognition is context-dependent, emerging through interactions with real-world environments and evolving in response to stimuli and circumstances [25]. As AI technology evolves and its applications diversify, the intelligence exhibited through human–machine interaction similarly transforms, reflecting contextual and temporal variations [26].

In summary, cognitive theory enables AI systems to simulate human cognitive processes and capabilities [27] while emphasizing the contextual nature of cognition [24]. E-commerce virtual anchors, as a distinct application of AI in live-streaming commerce, exhibit intelligence shaped by specific interactive scenarios. To comprehensively evaluate this intelligence, it is necessary to examine the human–machine interaction process between virtual anchors and consumers and to identify the intelligent capabilities demonstrated at each stage.

2.3. Information Interaction Theory

Initially introduced by Bill Moggridge of IDEO in 1984, interaction design was used to analyze information exchange phenomena in human–computer interaction. In the field of informatics, this theory describes the full cycle of messaging from production and circulation to reception, revision, and feedback [28]. These interrelated stages constitute a complete information flow, enabling effective communication and interpretation. A key feature of information interaction is its bidirectionality [29]. In live-streaming e-commerce, consumers and virtual anchors alternate between roles as information senders and receivers through multiple rounds of interactive exchanges.

AI products inherently possess the ability to collect, process, and generate information, a capability often described as “intelligence” [30]. Information serves as the basic unit of cognitive and mental activity [27], with cognition reflected in the mental processes of information input, encoding, storage, and retrieval during task completion. Drawing from cognitive and information interaction theories, this study posits that examining the intelligence of e-commerce virtual anchors requires an analysis of information interaction in product promotion, with a focus on the intelligent capabilities exhibited throughout the process.

2.4. Consumer Participation Behavior

Cermak (1994) [31] believes that consumers participate in the quality and satisfaction of business activities, and will respond to behavior, which is a feedback of the consumption process. Value co-creation behavior is a kind of high-level consumer behavior [32] that occurs between enterprises and consumers to create and enhance value [33]. The role of consumers in value co-creation influences their behavior and contribution [34].

From the perspective of consumers, Yi (2013) [35] further conceptualized value co-creation behavior as consisting of consumer participation behavior and consumer citizenship behavior. Consumer participation involves intrinsic behaviors such as information seeking, information sharing, responsible conduct, and interpersonal interaction. In e-commerce live streams, consumers seek information by asking questions, share information through comments, engage responsibly by adhering to stream norms, and express interpersonal behavior through actions such as praising and following virtual anchors.

2.5. Shopping Value

Hirschman (1982) [36] conceptualized shopping value as the experiential benefits derived from acquiring a product or service. Some scholars emphasize that shopping value should consider both consumer income and the costs incurred during the shopping process [37]. Baker (2002) [38] classified consumer costs into two dimensions: non-emotional (financial, time, and energy) and emotional (psychological and affective). To enhance utilitarian value, consumers aim to reduce time and effort through goal-oriented behavior [39]. In contrast, hedonic value arises from emotionally satisfying and enjoyable shopping experiences [40].

Babin (1994) [41] developed a scale that classified shopping value into utilitarian and hedonic dimensions. Utilitarian value focuses on the functional benefits derived from efficient and economical product acquisition [42], whereas hedonic value pertains to the emotional gratification experienced during the shopping process itself [43].

2.6. Stimulus-Organism-Response (S-O-R) Model

Woodworth (1918) [44] introduced the stimulus-response (S-R) model, which posits that external stimuli elicit specific behavioral responses. Mehrabian proposed (1974) [45] expanded this framework into the S-O-R model by incorporating the organism (i.e., the individual’s cognitive and emotional processes) as a mediating variable. In this model, stimulus variables represent changes in the external environment that influence decision-making [46], organism variables denote the individual’s internal responses, and response variables capture the resulting behavior. The model’s core premise is that external stimuli influence behavior through psychological processes. The S-O-R framework has been widely applied in contemporary marketing research. As e-commerce live streaming continues to develop in China, it has gained significant attention from scholars studying its various scenarios. For example, Lee (2021) [47] applied the S-O-R model to examine how anchor characteristics in e-commerce live streaming affect consumers’ impulsive purchase behaviors.

3. Methodology

This study developed a measurement scale for assessing the intelligence of e-commerce virtual anchors by adapting Churchill’s (1979) [48] scale development methodology and incorporating insights from grounded theory. Through empirical research, the study explores how virtual anchor intelligence influences consumers’ participation behavior. Figure 1 illustrates the research roadmap.

4. Analysis and Results

4.1. Step 1: Specify Conceptual Scope

Glaser and Strauss (1967) [49] first introduced the grounded theory research method, a theory-generation approach [50] that is particularly suited for developing new concepts, such as e-commerce virtual anchor intelligence. Scholars employing this methodology typically use in-depth interviews as the primary data collection method, although it also supports the analysis of textual data, such as newspapers and biographical texts [51]. Figure 2 illustrates the grounded theory research implementation process used in this study.

First, researchers need to collect data and systematically organize the interview process. They should then employ the three-level coding method to conduct grounded theory analysis. After completing the coding, if theoretical saturation is reached, researchers can finalize the theoretical framework. However, saturation has not yet been achieved. Researchers must continue collecting data and conducting additional coding until theoretical saturation is confirmed, ultimately enabling the construction of a new theoretical framework.

4.1.1. Data Collection and Collation

This study examines the conceptual framework and dimensions of virtual anchor intelligence in e-commerce live streaming. Given the novelty of this research area, a grounded theory approach was employed, using primary data (in-depth interview transcripts) as the primary source and secondary data (online textual materials) as supplementary sources for content analysis.

The study conducted in-depth interviews with 28 participants, including consumers, e-commerce live-streaming operators, and providers of virtual human technology, with each interview lasting a minimum of 30 min. Interviews were concluded once theoretical saturation was reached. The face-to-face interviews were conducted by two researchers: one as the interviewer and the other as the recorder and observer. The interview transcripts were thoroughly cleaned and proofread by both researchers. The final set of in-depth interview records totaled 126,000 characters. A statistical breakdown of the interview samples is presented in

Table 1. Interview sample statistics on the intelligence of e-commerce virtual anchors.

Basic Information	Category		Count	Percentage
Gender	Male		12	42.86%
	Female		16	57.14%
Age	50 years old and above		2	7.14%
	40–49 years old		6	21.43%
	30–39 years old		8	28.57%
	18–29 years old		11	39.29%
	Under 18 years old		1	3.57%
Participant Roles	Consumer	Private Enterprise	6	21.43%
		State-Owned Enterprise	5	17.86%
		Government/Public Institution	1	3.57%
		Freelancer	2	7.14%
		University Student	5	17.86%
	e-Commerce Live Stream Operation Company		4	14.29%
	Virtual Human Technology Supplier		5	17.86%
Education	Doctor’s degree		3	10.71%
	Master’s degree		7	25.00%
	Bachelor’s degree		9	32.14%
	Associate Degree		8	28.57%
	High School/Vocational		1	3.57%

.

In addition, this study explored the intelligence of e-commerce virtual anchors through a “multi-platform coverage and precision keyword search” strategy. Five major Chinese websites were selected: Zhihu, Weibo, Baidu Tieba, Douban, and Xiaohongshu. Using core terms such as “e-commerce virtual anchors,” “digital human live-streaming sales,” and “virtual human e-commerce live-streaming,” supplemented by scenario keywords like “intelligence” and “interaction,” relevant texts from January 2021 to June 2024 were retrieved. Through platform searches and compliant web crawlers, 327 texts (203,000 words) were obtained. After dual deduplication—removing 48 fully duplicated entries and merging 77 fragmented texts—202 unique texts remained. Of these, 62% were short comments (50–300 words), and 38% were in-depth analyses (over 300 words). Finally, based on three-tier criteria—“intelligent expressions,” “focus on capability descriptions,” and “clear viewpoints”—48 disputed texts were removed, resulting in 179 intelligence-related texts (87,000 words) that were used for grounded theory coding.

4.1.2. Data Coding

This study employed NVivo12.0 software to conduct three rounds of grounded theory coding [51], aiming to establish a comprehensive theoretical framework and conceptual system. The first phase, open coding, involved in-depth analysis of the raw data to develop initial concepts and categories, resulting in 41 preliminary concepts (coding results detailed in

Table A1. Conceptualization analysis of part of the original data.

Concept	Original Data for Open Coding	Information Provided
a1 Proactive welcoming	As soon as I entered the live stream room, the e-commerce virtual host warmly said ‘Welcome’.	Consumer X3
a2 Proactive self-introduction	The e-commerce virtual host would proactively introduce itself: ‘Hello, I am Xiao Zhi.’	Consumer X11
a3 Proactive service introduction	The e-commerce virtual host would proactively say it can introduce products to me, and redeem coupons, among other services.	Consumer X14
a4 Proactive product information sharing	The e-commerce virtual host would proactively introduce the traceability information of the orange’s origin.	Consumer X5
a5 Proactive event/promotion sharing	We provide all the product information data to it, and the e-commerce virtual host can introduce the product’s performance parameters and other information.	Consumer X1
a6 Product knowledge introduction	It would introduce the brand advantages, history, and spokesperson of the product.	Operation operator Y1
a7 Brand knowledge introduction	Consumers need input the discount information following the template, and upon triggering the keyword, it will proactively present the store’s discounts.	Consumer X1
a8 Store discount/promotion introduction	The e-commerce virtual host repeatedly emphasizes that after receiving the platform’s subsidies, the price is very reasonable.	Technical staff J3
a9 Platform discount/promotion introduction	It can obtain the consumer ID and address consumers by their ID to enhance the relationship.	Consumer X18
a10 Consumer ID information collection	The e-commerce virtual host knows I am male and never uses feminine terms.	Operation operator Y3
a11 Consumer gender recognition	The first time I entered the e-commerce virtual host’s live stream room, it knew I was a new customer.	Consumer X7
a12 New customer recognition	The e-commerce virtual host can greet returning customers directly as ‘old friends’.	Consumer X10
a13 Returning customer recognition	I said I wanted to learn more about the product, and after chatting for a while, the e-commerce virtual host realized I was quite interested in buying it.	Operation operator Y2
a14 Shopping intent perception	Mother’s Day is coming soon, and when introducing the product, it specifically mentioned that it is especially suitable for giving to mothers.	Consumer X12
a15 Shopping motivation perception	I said I was not very satisfied with the product I bought before, asked what to do, and the e-commerce virtual host guided me to the after-sales service channel.	Consumer X4
a16 After-sales intent perception	I am not very familiar with this product, and the e-commerce virtual host promises dissatisfaction, return postage paid.	Consumer X13
a17 After-sales concern perception	The e-commerce virtual host can filter out invalid commands and understand precise commands.	Consumer X15
a18 Command comprehension	As long as consumers express their questions, it can understand them.	Technical staff J2
a19 Question comprehension	I left a message with a bad tone in the live stream room, and it quickly said ‘Sorry’.	Operation staff Y4
a20 Tone comprehension	I sent several ‘!!!!!!’, and it replied with ‘Please don’t worry.’	Consumer X7
a21 Punctuation comprehension	I sent a smiley face emoticon, and the e-commerce virtual host quickly understood that I was greeting in a friendly manner, replying with ‘Hello’.	Consumer X13
a22 Emoji/sticker comprehension	As long as you ask a question, the e-commerce virtual host immediately understands what you are asking.	Consumer X16
a23 Quick reception	When I asked if there were any discounts, it quickly gave me feedback with an exclusive coupon.	Online resources
a24 Agile decision-making	If the algorithm is precise enough and the prediction library is large enough, when answering questions, it is more accurate than humans.	Consumer X4
a25 Accurate response	The e-commerce virtual host does not know fatigue and can work 24 h a day, maintaining a good working state all day long.	Online resources
a26 Stable performance	I think the e-commerce virtual anchor is very patient.	Operation operator Y2
a27 Emotional stability	The e-commerce virtual anchor’s emotions are very full, very energetic, and very enthusiastic.	Consumer X17
a28 Emotional expression	The virtual anchor speaks very fluently, feeling like the rhythm of speaking is almost the same as a real person, and even has an accent and mannerisms.	Consumer X11
a29 Language expression	The e-commerce virtual anchor’s body movements are coordinated with the language.	Consumer X19
a30 Body language expression	From a technical perspective, it has already achieved complete human voice simulation, matching the corresponding script, just like a real person speaking.	Technical staff J4
a31 Human-like voice	The intonation is very similar to a real person, and even when speaking, it has a bit of an accent, with the speech intonation winding and twisting, sounding absolutely not like a robot.	Technical staff J3
a32 Human-like tone	The e-commerce virtual anchor can learn anyone’s voice, the timbre is exactly the same as a real person, there is no difference.	Consumer X19
a33 Human-like timbre	The e-commerce virtual anchor needs to be compatible with the product, its entire character design must be carefully designed.	Technical staff J2
a34 Product compatibility	The e-commerce virtual anchor can choose different language styles for conversation.	Technical staff J1
a35 Language style	As long as the basic activity content is entered, the e-commerce virtual anchor can automatically generate multiple activity scripts.	Technical staff J3
a36 Promotional script	The e-commerce virtual anchor can generate sales scripts based on different consumer statuses.	Technical staff J1
a37 Sales script	The e-commerce virtual anchor can adjust the live room atmosphere at the appropriate time.	Operation operator Y2
a38 Atmosphere building script	It will combine relevant knowledge, such as how many times a baby should be fed milk a day, what temperature it should be, etc.	Technical staff J2
a39 Knowledge based topics	We will regularly update topics, so that the e-commerce virtual anchor can keep up with the trends when replying to consumers.	Consumer X9
a40 Trendy topics	As soon as I entered the live stream room, the e-commerce virtual host warmly said ‘Welcome’.	Technical staff J1

). Next, the axial coding phase involved systematic data analysis, structural organization, and synthesis of each category, identifying core issues through data compression, synthesis, and interpretation. By examining the operational realities of e-commerce virtual live-streaming platforms, this study ultimately identified eight competency dimensions as subcategories of intelligent capabilities, with detailed coding results presented in

Table 2. Analysis results of axis coding.

Category	Sub-Category	Initial Category
C1 Guidance Intelligence	B1 Proactive ability	A1 Proactive greeting
		A2 Proactive sharing
	B2 Knowledge acquisition ability	A3 Product and brand knowledge
		A4 Promotional event knowledge
C2 Recognition Intelligence	B3 Information acquisition ability	A5 Personal information collection
		A6 Shopping information collection
	B4 Scenario awareness ability	A7 Shopping needs perception
		A8 After-sales needs perception
C3 Analysis Intelligence	B5 Understanding ability	A9 Language understanding
		A10 Emotional understanding
	B6 Reaction ability	A11 Quick response
		A12 Precision
		A13 Stability
C4 Feedback Intelligence	B7 Personification ability	A14 Natural expression
		A15 Human-like pronunciation
		A16 Persona and personality
	B8 Communication ability	A17 Conversation script
		A18 Topic

.

The third phase, selective coding, then aimed to identify the main categories, systematically link them, and verify their interconnections. At this stage, the study sought to establish a comprehensive theoretical framework and conceptual system to explain and address the identified issues.

Building on these coding results, the study incorporated concepts from information interaction theory. According to this theory, common stages in information exchange include generation, circulation, reception, correction, and feedback [28], with specific scenarios influencing the interaction process. Drawing on this framework, the study analyzed in-depth interview materials to map the actual information exchange flow between e-commerce virtual anchors and consumers. In the original transcripts, interviewees explicitly described service processes related to virtual anchors’ capabilities, such as “entering the live-streaming room,” “asking questions,” “during conversations,” and “when receiving responses.” By integrating information interaction theory with interview data, we identified key phases in human–computer interaction between virtual anchors and consumers, as illustrated in Figure 3.

1.

Guidance

When consumers enter a live-streaming room to watch e-commerce virtual anchors selling products, they may not immediately engage in communication. At this stage, the virtual anchor should guide them through proactive dialog and professional knowledge sharing to facilitate interaction.

2.

Recognition

Before speaking with a consumer, the virtual anchor collects information through different channels. This helps assess if the consumer wants to shop or needs after-sales service. Consumers requiring after-sales assistance are directed to the customer service. Consumers who want to buy continue to receive help from the anchor.

3.

Analysis

The virtual anchor analyzes the collected data to accurately interpret and understand the consumer’s purchasing intentions.

4.

Feedback

The virtual anchor provides feedback based on the analysis of consumer needs.

Furthermore, cognitive theory suggests that e-commerce virtual anchors must demonstrate distinct capabilities at every stage of human–computer interaction to effectively serve consumers. Based on this theoretical background and the previous stage analysis, this study categorizes the competencies exhibited during consumer-virtual anchor interactions into four core domains: guidance intelligence, recognition intelligence, analysis intelligence, and feedback intelligence. Figure 4 illustrates the results of the selective coding.

In summary, this study adheres closely to the semantic content of the original materials, distilling definitions and explanations of the four primary categories based on the actual practices of e-commerce virtual anchors in live selling.

1.

C1 Guidance Intelligence

Guidance intelligence is the proactive capability and knowledge that e-commerce virtual anchors use to help consumers interact. Even if consumers do not interact with virtual anchors during live streams, anchors can use their knowledge to initiate conversations and help consumers express their needs.

2.

C2 Recognition Intelligence

Recognition intelligence is the virtual anchor’s capability to gather information and understand the situation. During interactions, virtual anchors collect data such as consumer IDs, gender, and purchase history. They also study each scenario in real-time to direct consumers to after-sales service or keep them in the shopping process.

3.

C3 Analysis Intelligence

Analysis intelligence is the virtual anchor’s ability to understand and reply to consumer dialog. To do this well, virtual anchors must first grasp what consumers say. Then, they use their knowledge to match conversation content and emotions [11] quickly.

4.

C4 Feedback Intelligence

Feedback intelligence is the way virtual anchors respond and communicate in a manner that mimics human behavior. During live streaming, anchors tailor their responses to align with their online persona, using replies and tones that match their brand. Virtual anchors are designed to behave in a human-like manner [52]. They can also use pre-set dialog and topics when talking with consumers [53].

4.1.3. Theoretical Saturation Analysis

Francis (2010) [54] suggested that when no new concepts emerge during coding, three additional interviews should be conducted and analyzed. If no further concepts arise, theoretical saturation is considered to have been achieved. To ensure comprehensive conceptual development of “e-commerce virtual anchors’ intelligence” and to establish a robust foundation for analyzing interdimensional relationships, this study conducted real-time coding of consumer interview data and online textual materials. By the 16th consumer interview, no new concepts were detected. Three additional interviews were conducted, but no new concepts emerged, resulting in a 100% repetition rate in the main category structure and factor composition. This outcome aligns with Glaser and Strauss’s (1967) [49] criterion for saturation: “no new data extends the conceptual scope.” This indicates that consumer perspectives achieved theoretical saturation.

To further refine the conceptual boundaries of e-commerce virtual anchors’ intelligence, interviews were expanded to include industry practitioners, providing additional insights from both demand and technical perspectives. After six interviews with practitioners, no new contributing concepts were identified. Three supplementary interviews also yielded no new concepts, and the main category structure and factor composition again exhibited a 100% repetition rate, indicating that practitioner perspectives had also reached theoretical saturation.

In conclusion, the consistency of the main category structures and constituent factors confirms that the core concept of “e-commerce virtual anchors’ intelligence” has been thoroughly articulated, achieving a stable state of theoretical saturation.

4.2. Step 2: Generate a Sample of Items

This study generated initial measurement items for assessing the intelligence of e-commerce virtual anchors by integrating findings from grounded theory analysis and existing academic literature. These items were further refined using the Delphi method. The Delphi method is a predictive decision-making technique that gathers expert opinions through multiple rounds of anonymous feedback [55]. We invited five experts specializing in management, e-commerce, virtual anchor technology, and multilingual communication to evaluate and enhance the preliminary questionnaire items through iterative discussions. After completing the Delphi refinement process, a 20-item initial scale was formulated, as presented in

Table 3. Initial item pool of intelligence for e-commerce virtual anchors.

Dimensions	Item	Item Description	Item Source
Guidance Intelligence	GIN1	When I enter the e-commerce virtual anchor’s live-streaming room, I find it proactively greets me with a warm welcome to capture my attention.	Grounded theory
	GIN2	When I enter the e-commerce virtual anchor’s live-streaming room, I find it actively shares information to guide my shopping experience.	Grounded theory
	GIN3	When I enter the e-commerce virtual anchor’s live-streaming room, I find that its extensive product knowledge helps me understand featured items.	Grounded theory
	GIN4	When I enter the e-commerce virtual anchor’s live-streaming room, I find its comprehensive marketing knowledge introduces me to promotions.	Grounded theory
Recognition Intelligence	RIN1	While in the e-commerce virtual anchor’s live-streaming room, I find that it recognizes my identity.	Grounded theory
	RIN2	When in the e-commerce virtual anchor’s live-streaming room, I find that it can recognize whether I’ve shopped there before.	Grounded theory
	RIN3	During conversations with the e-commerce virtual anchor, I find that it detects my purchase intent.	Grounded theory
	RIN4	During conversations with the e-commerce virtual anchor, I find that it recognizes when I need after-sales support.	Grounded theory
	RIN5	During conversations with the e-commerce virtual anchor, I find that it understands my preferences.	Mühlhäuser (2008) [56]
Analysis Intelligence	AIN1	When I share my needs with the e-commerce virtual anchor, it analyzes our conversation content.	Grounded theory
	AIN2	When I tell the e-commerce virtual anchor my needs, I find that it can analyze my emotions during our conversation.	Grounded theory
	AIN3	When I share my needs with the e-commerce virtual anchor, it performs analysis much faster than a real person.	Grounded theory
	AIN4	When I share my needs with the e-commerce virtual anchor, it responds more accurately with fewer errors than a real person.	Moussawi (2020) [57]
	AIN5	When I tell the e-commerce virtual anchor about my needs, its analysis remains more stable than a real person’s—both in processing state and emotional response.	Grounded theory
Feedback Intelligence	FIN1	When the e-commerce virtual anchor responds to me, I find its emotional expressions, verbal communication, and body language all appear natural and smooth.	Grounded theory
	FIN2	When the e-commerce virtual anchor gives feedback, its voice sounds exactly like a real human’s, in terms of speech, tone, and vocal quality.	Grounded theory
	FIN3	When e-commerce virtual anchor gives feedback to me, its style aligns with its character and the products it sells.	Grounded theory
	FIN4	When the e-commerce virtual anchor communicates, I find it effectively uses sales techniques to achieve its goals.	Grounded theory
	FIN5	When the e-commerce virtual anchor responds, I find it skillfully incorporates relevant topics into our conversation.	Grounded theory
	FIN6	When I receive feedback from the e-commerce virtual anchor, I feel like I’m interacting with a real person.	McLean et al. (2021) [58]

.

4.3. Step 3: Purify Measures

4.3.1. Data Collection for Scale Purification

This study used a 7-point Likert scale to create the initial e-commerce virtual anchor intelligence questionnaire. Refined demographic items and screening criteria were included. Respondents rated each item from 1 (strongly disagree) to 7 (strongly agree). The questionnaire was distributed online. Out of 150 responses, 122 were valid after screening. This produced an effective response rate of 81.33%.

Table 4. Demographic data analysis results of scale purification (N = 122).

Basic Information	Category	Count	Percentage
Gender	Male	49	40.16%
	Female	73	59.84%
Age	18–30 years old	31	25.41%
	31–40 years old	44	36.07%
	41–50 years old	27	22.13%
	51–60 years old	17	13.93%
	Over 60 years old	3	2.46%
Education	Junior High or Below	3	2.46%
	High School/Vocational	26	21.31%
	Associate Degree	44	36.07%
	Bachelor’s Degree	37	30.33%
	Master’s or Above	12	9.84%
Occupation	State-Owned Enterprise	21	17.21%
	Government/Public Institution	17	13.93%
	Private Enterprise	45	36.89%
	Foreign Enterprise	22	18.03%
	University Student	12	9.84%
	Freelancer	5	4.10%
Monthly Income (CNY)	Below 3000 CNY	23	18.85%
	3000–6500 CNY	49	40.16%
	6500–10,000 CNY	32	26.23%
	Above 10,000 CNY	18	14.75%

presents participant demographics for this scale purification phase. The sample closely matches the usual audience profile of e-commerce virtual anchor live-streaming sessions.

4.3.2. Scale Purification

To purify the scale, we tested the reliability of the initial items using internal consistency analysis.

Table 5. Reliability test results of the purification stage (N = 122).

Variable	Variable Dimensions	Item	CITC	Cronbach’s α After Deletion	Cronbach’s α
Intelligence of e-commerce virtual anchors	Guidance Intelligence	GIN1	0.849	0.928	0.942
		GIN2	0.862	0.925
		GIN3	0.871	0.921
		GIN4	0.867	0.923
	Recognition Intelligence	RIN1	0.143	0.899	0.788
		RIN2	0.672	0.713
		RIN3	0.664	0.715
		RIN4	0.794	0.676
		RIN5	0.752	0.699
	Analysis Intelligence	AIN1	0.68	0.867	0.883
		AIN2	0.674	0.868
		AIN3	0.72	0.857
		AIN4	0.745	0.852
		AIN5	0.776	0.844
	Feedback Intelligence	FIN1	0.769	0.79	0.845
		FIN2	0.808	0.787
		FIN3	0.074	0.931
		FIN4	0.731	0.8
		FIN5	0.811	0.785
		FIN6	0.78	0.789

shows the reliability test results. Using SPSS, we found that items relating to guidance and analysis intelligence passed reliability testing. However, item RIN1 (Recognition Intelligence) had a Corrected Item-Total Correlation (CITC) value of 0.143, which is below the 0.4 threshold. After removing RIN1, the Cronbach’s α for recognition intelligence improved from 0.788 to 0.899. This supported its exclusion. Similarly, item FIN3 (Feedback Intelligence) had a CITC of 0.074, also below 0.4. Removing FIN3 increased Cronbach’s α from 0.845 to 0.931. This justified its deletion as well.

After removing RIN1 and FIN3, the remaining items in each dimension were renumbered sequentially.

Table 6. Reliability test results for the revised purification (N = 122).

Variable	Variable Dimensions	Item	CITC	Cronbach’s α After Deletion	Cronbach’s α
Intelligence of e-commerce virtual anchors	Guidance Intelligence	GIN1	0.849	0.928	0.942
		GIN2	0.862	0.925
		GIN3	0.871	0.921
		GIN4	0.867	0.923
	Recognition Intelligence	RIN1	0.724	0.889	0.899
		RIN2	0.747	0.881
		RIN3	0.862	0.837
		RIN4	0.78	0.87
	Analysis Intelligence	AIN1	0.68	0.867	0.883
		AIN2	0.674	0.868
		AIN3	0.72	0.857
		AIN4	0.745	0.852
		AIN5	0.776	0.844
	Feedback Intelligence	FIN1	0.778	0.924	0.931
		FIN2	0.868	0.907
		FIN3	0.782	0.922
		FIN4	0.85	0.909
		FIN5	0.821	0.915

shows the reliability results for the refined scale.

The final questionnaire, as summarized in

Table 7. Purification results of the initial intelligence scale of e-commerce virtual anchors.

Variable	Item	Item Description
Guidance Intelligence	GIN1	When I enter the e-commerce virtual anchor’s live-streaming room, I find it proactively greets me with a warm welcome to capture my attention.
	GIN2	When I enter the e-commerce virtual anchor’s live-streaming room, I find it actively shares information to guide my shopping experience.
	GIN3	When I enter the e-commerce virtual anchor’s live-streaming room, I find that its extensive product knowledge helps me understand featured items.
	GIN4	When I enter the e-commerce virtual anchor’s live-streaming room, I find its comprehensive marketing knowledge introduces me to promotions.
Recognition Intelligence	RIN1	When in the e-commerce virtual anchor’s live-streaming room, I find that it can recognize whether I’ve shopped there before.
	RIN2	During conversations with the e-commerce virtual anchor, I find that it detects my purchase intent.
	RIN3	During conversations with the e-commerce virtual anchor, I find that it recognizes when I need after-sales support.
	RIN4	During conversations with the e-commerce virtual anchor, I find that it understands my preferences.
Analysis Intelligence	AIN1	When I share my needs with the e-commerce virtual anchor, it analyzes our conversation content.
	AIN2	When I tell the e-commerce virtual anchor my needs, I find that it can analyze my emotions during our conversation.
	AIN3	When I share my needs with the e-commerce virtual anchor, it performs analysis much faster than a real person.
	AIN4	When I share my needs with the e-commerce virtual anchor, it responds more accurately with fewer errors than a real person.
	AIN5	When I tell the e-commerce virtual anchor about my needs, its analysis remains more stable than a real person’s—both in processing state and emotional response.
Feedback Intelligence	FIN1	When the e-commerce virtual anchor responds to me, I find its emotional expressions, verbal communication, and body language all appear natural and smooth.
	FIN2	When the e-commerce virtual anchor gives feedback, its voice sounds exactly like real humans’, in terms of speech, tone, and vocal quality.
	FIN3	When e-commerce virtual anchor gives feedback to me, its style aligns with its character and the products it sells.
	FIN4	When the e-commerce virtual anchor communicates, I find it effectively uses sales techniques to achieve its goals.
	FIN5	When the e-commerce virtual anchor responds, I find it skillfully incorporates relevant topics into our conversation.

, contains 18 items across four intelligence dimensions.

4.4. Step 4: Assess Reliability

4.4.1. Data Collection for Scale Assessment

This study employed a 7-point Likert scale to develop a questionnaire comprising 18 items on e-commerce virtual anchor intelligence, along with demographic and screening questions. The survey was distributed to 600 participants both online and offline. After screening, 536 valid responses remained, resulting in an effective response rate of 89.33%.

4.4.2. Demographic Data Analysis

The reliability evaluation was based on the 536 valid responses.

Table 8. Demographic data analysis results for reliability test (N = 536).

Basic Information	Category	Count	Percentage
Gender	Male	214	39.93%
	Female	322	60.07%
Age	18–30 years old	188	35.07%
	31–40 years old	168	31.34%
	41–50 years old	103	19.22%
	51–60 years old	62	11.57%
	Over 60 years old	15	2.80%
Education	Junior High or Below	26	4.85%
	High School/Vocational	77	14.37%
	Associate Degree	191	35.63%
	Bachelor’s Degree	156	29.10%
	Master’s or Above	86	16.04%
Occupation	State-Owned Enterprise	106	19.78%
	Government/Public Institution	78	14.55%
	Private Enterprise	164	30.60%
	Foreign Enterprise	95	17.72%
	University Student	78	14.55%
	Freelancer	15	2.80%
Monthly Income (CNY)	Below 3000 CNY	90	16.79%
	3000–6500 CNY	173	32.28%
	6500–10,000 CNY	188	35.07%
	Above 10,000 CNY	85	15.86%

summarizes the demographic characteristics of the sample during this phase, demonstrating that the participant profile aligns with typical audience composition in e-commerce virtual anchor live-streaming settings.

4.4.3. Sample Analysis and Results

Building on the demographic results, we evaluated the reliability of the refined measurement scale using Cronbach’s α and CITC as key indicators. We used thresholds of α > 0.7 and CITC > 0.4. As shown in

Table 9. Reliability test results of the scale evaluation stage (N = 536).

Variable	Variable Dimensions	Item	CITC	Cronbach’s α After Deletion	Cronbach’s α
Intelligence of e-commerce virtual anchors	Guidance Intelligence	GIN1	0.844	0.924	0.939
		GIN2	0.867	0.917
		GIN3	0.853	0.921
		GIN4	0.859	0.92
	Recognition Intelligence	RIN1	0.731	0.905	0.909
		RIN2	0.785	0.886
		RIN3	0.866	0.856
		RIN4	0.803	0.881
	Analysis Intelligence	AIN1	0.699	0.863	0.883
		AIN2	0.696	0.864
		AIN3	0.739	0.854
		AIN4	0.731	0.856
		AIN5	0.731	0.855
	Feedback Intelligence	FIN1	0.79	0.927	0.935
		FIN2	0.844	0.916
		FIN3	0.815	0.922
		FIN4	0.848	0.916
		FIN5	0.836	0.918

, the lowest CITC value across items in all four dimensions was 0.696, exceeding the minimum threshold. Furthermore, the Cronbach’s α values did not improve upon the removal of any item, indicating strong internal consistency. These results confirm that the purified scale meets reliability standards. Further validation analyses are warranted to establish broader measurement validity.

4.5. Step 5: Assess Validity

Table 10. KMO and Bartlett’s sphericity test results for the scale evaluation (N = 536).

KMO Value		0.934
Bartlett’s Test of Sphericity	Approximate Chi-Square	3842.884
	df	153
	Sig.	0

presents the results of the Kaiser-Meyer-Olkin (KMO) and Bartlett’s test of sphericity for the scale evaluation stage.

As shown in Table 10, the KMO value for the pre-research questionnaire on the intelligence of e-commerce virtual anchors in the scale evaluation stage is 0.934, which exceeds the general minimum standard of 0.7. The approximate chi-square value for Bartlett’s test of sphericity is 3842.884, with 153 degrees of freedom and a significance level of sig < 0.05, indicating that the questionnaire results are very suitable.

The 536 valid samples were randomly divided into two groups of 268 each for Exploratory Factor Analysis (EFA) and Confirmatory Factor Analysis (CFA). The EFA employed principal component analysis to extract common factors, setting an eigenvalue greater than one as the criterion for extraction. We selected the maximum variation method for orthogonal rotation. Additionally, all factor loadings were required to be greater than 0.5. As shown in

Table 11. EFA results of the scale assessment (N = 268).

Item	Component
	Factor 1	Factor 2	Factor 3	Factor 4
GIN1	0.819
GIN2	0.843
GIN3	0.776
GIN4	0.815
RIN1		0.719
RIN2		0.784
RIN3		0.826
RIN4		0.771
AIN1			0.775
AIN2			0.771
AIN3			0.821
AIN4			0.81
AIN5			0.841
FIN1				0.812
FIN2				0.797
FIN3				0.763
FIN4				0.847
FIN5				0.804

, the test items in this study align with the theoretical hypotheses, indicating that the scale has good validity and that the EFA results are acceptable.

CFA was conducted using AMOS 24.0 on the second subsample (N = 268), with fit indices reported in

Table 12. Test results for the fit index in the scale evaluation.

${\mathbf{X}}^2/\mathbf{d}\mathbf{f}$	RMSEA	GFI	NFI	CFI	IFI	TLI
1.289	0.033	0.94	0.958	0.99	0.99	0.988

.

As shown in Table 12, the ${\mathrm{X}}^2/\mathrm{d}\mathrm{f}$ ratio is 1.289 (below the recommended maximum of 3), RMSEA is 0.033 (below 0.08), and GFI, NFI, CFI, IFI, and TLI values all exceed 0.90. These indices collectively indicate an excellent model fit. Therefore, no model modifications are necessary [59].

Figure 5 and

Table 13. Convergence validity test results for the scale evaluation.

Variable	Path	Loading Coefficient	AVE	CR
Guidance Intelligence	GIN1 ← guidance intelligence	0.874	0.7806	0.9344
	GIN2 ← guidance intelligence	0.895
	GIN3 ← guidance intelligence	0.881
	GIN4 ← guidance intelligence	0.884
Recognition Intelligence	RIN1 ← recognition intelligence	0.786	0.7248	0.9131
	RIN2 ← recognition intelligence	0.839
	RIN3 ← recognition intelligence	0.92
	RIN4 ← recognition intelligence	0.855
Analysis Intelligence	AIN1 ← analysis intelligence	0.764	0.5998	0.8822
	AIN2 ← analysis intelligence	0.742
	AIN3 ← analysis intelligence	0.801
	AIN4 ← analysis intelligence	0.782
	AIN5 ← analysis intelligence	0.782
Feedback Intelligence	FIN1 ← feedback intelligence	0.848	0.7478	0.9368
	FIN2 ← feedback intelligence	0.869
	FIN3 ← feedback intelligence	0.846
	FIN4 ← feedback intelligence	0.872
	FIN5 ← feedback intelligence	0.888

present the results of further CFA and convergent validity testing. The minimum Average Variance Extracted (AVE) is 0.5998, exceeding the 0.50 benchmark, and the minimum Composite Reliability (CR) is 0.8822, surpassing the 0.80 standard. These results confirm satisfactory convergent validity and construct reliability.

To assess discriminant validity, we examined the correlations among the four dimensions of virtual anchor intelligence, as shown in

Table 14. Discriminant validity test results of the subscale evaluation stage.

	Guidance Intelligence	Recognition Intelligence	Analysis Intelligence	Feedback Intelligence
Guidance Intelligence	0.7806
Recognition Intelligence	0.695 ***	0.7248
Analysis Intelligence	0.379 ***	0.349 ***	0.5998
Feedback Intelligence	0.672 ***	0.721 ***	0.385 ***	0.7478
AVE square root	0.8835	0.8514	0.7745	0.8819

*** Represents p value less than 0.001; Diagonal is AVE.

. These dimensions—guidance intelligence, recognition intelligence, analysis intelligence, and feedback intelligence—exhibit significant pairwise correlations (p < 0.001). However, all correlations are less than the square roots of the corresponding AVEs. For instance, the correlation between recognition intelligence and analytical intelligence is 0.349. This is distinctly lower than the square root of AVE for both constructs (0.8514 and 0.7745, respectively). This confirms that the dimensions are related but remain empirically distinct, thereby demonstrating acceptable discriminant validity.

In summary, the results of both EFA and CFA indicate that the refined scale has robust validity. The selected items adequately capture their intended constructs and are appropriate for further research in this domain.

4.6. Step 6: Scale Verification

This study employs the S-O-R theoretical framework to establish the logical pathway: “intelligence → shopping value → consumer participation behavior.” We administered questionnaire surveys to assess both the construct validity and theoretical soundness of the E-commerce Virtual Host Intelligence Scale. Specifically, we examined the relationships between intelligence and external variables, including shopping value and consumer participation behavior [60]. Additionally, we evaluated the scale’s predictive validity in relation to actual consumption behavior. Predictive validity, a key component of construct validity, reflects the scale’s ability to forecast consumption behavior through consistent and significant correlations with external criterion variables—such as perceived shopping value and participation behavior. This assessment provides a crucial foundation for verifying the structural robustness and practical applicability of the scale.

4.6.1. Research Hypothesis

Prior studies have demonstrated that AI product guidance fosters positive human behavior and influences human–computer interaction [61]. AI products can detect user emotions by interpreting contextual cues [62]. E-commerce platforms that utilize AI for personalized product recommendations have significantly enhanced consumer experiences and click-through rates [63]. Song (2020) [64] found that in judicial environments, the analysis intelligence of AI voice systems helps judges process cases more effectively. Additionally, the more e-commerce virtual anchors emulate human behavior, the better they stimulate consumer participation in value co-creation [7].

In summary, AI attributes have a significant influence on consumer behavior. Building on this, the study proposes the following hypotheses:

H1a. 

Guidance intelligence has a positive effect on consumer participation behavior.

H1b. 

Recognition intelligence has a positive effect on consumer participation behavior.

H1c. 

Analysis intelligence has a positive effect on consumer participation behavior.

H1d. 

Feedback intelligence has a positive effect on consumer participation behavior.

Miners (2013) [65] found that interactive AI characters enhance player enjoyment in gaming environments. Holstein (2019) [66] suggested that the extensive knowledge base of AI educators improves learning outcomes. Yang (2021) [67] empirically showed that self-learning and associative capabilities of AI products enhance consumers’ overall experience. Liu (2022) [68] highlighted that higher anthropomorphism in AI instructors leads to greater student acceptance, reduced burnout, and enhanced enjoyment. Rijsdijk et al. (2009) [30] asserted that AI is based on learning capability, which can enhance user experience. Qin (2025) [69] found that consumers’ perceptions of e-commerce virtual anchors’ competence foster psychological contract formation.

In summary, inherent AI capabilities offer users enhanced experiential value. Shopping value reflects consumers’ perceived value during the purchasing process. Based on this, the following hypotheses are proposed:

H2a. 

Guidance intelligence positively influences consumers’ utilitarian shopping value.

H2b. 

Guidance intelligence has a positive influence on consumers’ hedonic shopping value.

H2c. 

Recognition intelligence positively influences consumers’ utilitarian shopping value.

H2d. 

Recognition intelligence has a positive influence on consumers’ hedonic shopping value.

H2e. 

Analysis intelligence positively influences consumers’ utilitarian shopping value.

H2f. 

Analysis intelligence positively influences consumers’ hedonic shopping value.

H2g. 

Feedback intelligence positively influences consumers’ utilitarian shopping value.

H2h. 

Feedback intelligence positively influences consumers’ hedonic shopping value.

Moreover, Sun (2023) [70] found that perceived shopping value within virtual communities promotes the co-creation of brand value. Similarly, Liu (2022) [71] demonstrated a positive relationship between shopping value and value co-creation behavior on e-commerce platforms.

Based on these findings, the following hypotheses are put forward:

H3a. 

Consumers’ utilitarian shopping value positively affects their participation behavior.

H3b. 

Consumers’ hedonic shopping value positively affects their participation behavior.

This study selects “shopping value (utilitarian + hedonic)” and “consumer participation behavior” as key criterion variables to validate the predictive validity of the intelligence scale for the following reasons. (1) According to S-O-R theory, the sequence “stimulus (intelligence) → organism (shopping value) → response (participation behavior)” positions these variables as mediators and outcomes. This satisfies the requirement that criterion variables must be theoretically correlated with the construct. (2) Shopping value is a fundamental psychological driver of purchasing behavior. Participation behavior is a primary performance metric in e-commerce live streaming. These contextually relevant variables ensure both theoretical rigor and practical relevance.

4.6.2. Research Model

Drawing upon S-O-R and S-R theoretical frameworks and based on the proposed hypotheses, this study constructs a model illustrating the influence of e-commerce virtual anchors’ intelligence on consumer participation behavior in live-streaming scenarios. The conceptual model is presented in Figure 6.

4.6.3. Questionnaire Design and Distribution

This study adopted previously validated scales to assess the intelligence of e-commerce virtual anchors and other relevant variables, as presented in

Table 15. Sources of Core Variable Items.

Variable	Code	Source
Guidance Intelligence	GIN1	Previous Research
	GIN2
	GIN3
	GIN1
Recognition Intelligence	RIN1
	RIN2
	RIN3
	RIN4
Analysis Intelligence	AIN1
	AIN2
	AIN3
	AIN4
	AIN5
Feedback Intelligence	FIN1
	FIN2
	FIN3
	FIN4
	FIN5
Utilitarian Shopping Value	USV1	Babin (1994) [41] Bridges (2007) [72]
	USVI2
	USV3
	USV4
Hedonic Shopping Value	HSV1
	HSV2
	HSV3
	HSV4
Participation Behavior	PAR1	Yi (2013) [35]
	PAR2
	PAR3

. To begin, during the pre-survey phase, 200 questionnaires were distributed via an online research platform, yielding 181 valid responses—a response rate of 90.5%. Reliability testing showed that Cronbach’s α exceeded 0.7 and the CITC was above 0.35, confirming the questionnaire met the required standards for formal deployment. Building on these results, for the main survey, 700 questionnaires were distributed equally across online and offline channels (350 each). After data screening, 659 valid responses were retained, resulting in an effective response rate of 94.14%.

4.6.4. Data Measurement and Analysis

1.

Demographic analysis

The participants in the formal survey were closely aligned with the target audience profiles for e-commerce virtual anchor services. The detailed demographic characteristics are presented in

Table 16. Demographic data analysis results.

Basic Information	Category	Count	Percentage
Gender	Male	292	44.31%
	Female	367	55.69%
Age	18–30 years old	187	28.38%
	31–40 years old	154	23.37%
	41–50 years old	146	22.15%
	51–60 years old	106	16.08%
	Over 60 years old	66	10.02%
Education	Junior High or Below	43	6.53%
	High School/Vocational	148	22.46%
	Associate Degree	201	30.50%
	Bachelor’s Degree	183	27.77%
	Master’s or Above	84	12.75%
Occupation	State-Owned Enterprise	117	17.75%
	Government/Public Institution	125	18.97%
	Private Enterprise	142	21.55%
	Foreign Enterprise	128	19.42%
	University Student	93	14.11%
	Freelancer	54	8.19%
Monthly Income (CNY)	Below 3000 CNY	197	29.89%
	3000–6500 CNY	206	31.26%
	6500–10,000 CNY	172	26.10%
	Above 10,000 CNY	84	12.75%

. This overview provides a contextual foundation for the subsequent analyses of reliability and validity.

2.

Reliability and validity test

Table 17. Measurement model evaluation results.

Item	Mean	Standard Dev.	CITC	Cronbach’s α	CR	AVE
GIN1	4.39	1.609	0.882	0.924	0.9044	0.7029
GIN2	4.47	1.584	0.817
GIN3	4.43	1.582	0.832
GIN4	4.41	1.605	0.821
RIN1	4.9	1.315	0.757	0.895	0.8513	0.5888
RIN2	4.83	1.296	0.773
RIN3	4.95	1.368	0.76
RIN4	4.94	1.391	0.779
AIN1	4.87	1.405	0.644	0.881	0.842	0.5177
AIN2	4.73	1.443	0.74
AIN3	4.73	1.419	0.821
AIN4	4.61	1.549	0.69
AIN5	4.69	1.526	0.69
FIN1	4.82	1.504	0.8	0.9	0.8682	0.5699
FIN2	4.75	1.476	0.683
FIN3	4.82	1.477	0.677
FIN4	4.85	1.459	0.798
FIN5	4.88	1.491	0.805
USV1	4.63	1.441	0.686	0.873	0.8211	0.5356
USV2	4.63	1.383	0.678
USV3	4.92	1.29	0.807
USV4	4.88	1.316	0.749
HSV1	5.03	1.384	0.8	0.907	0.8949	0.63
HSV2	5.04	1.351	0.791
HSV3	5.17	1.403	0.818
HSV4	4.96	1.359	0.784
HSV5	4.97	1.405	0.775
PAR1	5.17	1.543	0.751	0.907	0.8705	0.628
PAR2	5.09	1.483	0.878
PAR3	5.22	1.496	0.782
PAR4	5.22	1.452	0.752

presents the structural reliability of the measurement model. All the Cronbach’s α values exceeded the recommended threshold of 0.7, indicating strong internal consistency. Convergent validity was assessed using AVE and standardized factor loadings. All factor loadings were above 0.5, and AVE values met acceptable criteria, confirming the reliability and validity of the constructs.

Discriminant validity was subsequently evaluated, as shown in

Table 18. Discrimination validity test result.

	Guidance Intelligence	Recognition Intelligence	Analysis Intelligence	Feedback Intelligence
Guidance Intelligence	0.7029
Recognition Intelligence	0.409 **	0.5888
Analysis Intelligence	0.685 ***	0.482 ***	0.5177
Feedback Intelligence	0.7 ***	0.480 ***	0.729 ***	0.5699
AVE square root	0.8384	0.7673	0.7195	0.7549

** Represents p value less than 0.01; *** Represents p value less than 0.001; Diagonal is AVE.

. All AVE square roots exceeded the corresponding inter-construct correlations, supporting adequate discriminant validity [73].

Next, model fit and predictive validity were assessed, with the results summarized in

Table 19. Model fit and prediction validity.

Fitting Index	CMIN/DF	RMSEA	NFI	CFI	IFI	TLI	TLI	Utilitarian Shopping Value R2	Hedonic Shopping Value R2	Consumer Participation Behavior R2
Model value	2.295	0.044	0.942	0.967	0.967	0.962	0.988	0.38	0.42	0.45

. All fit indices adhered to the recommended standards (CMIN/DF < 3, RMSEA < 0.08, CFI > 0.90). Additionally, the R² value (coefficient of determination) reflects the scale’s predictive explanatory power for external criterion variables, further supporting the validity of the predictions.

As shown in the table, the intelligence dimensions explain 38% of the variance in utilitarian shopping value (R² = 0.38), while intelligence accounts for 42% of the variance in hedonic shopping value (R² = 0.42), reflecting its significant influence on emotional experience. Furthermore, when combined with shopping value, intelligence explains 45% of the variance in participation behavior (R² = 0.45), indicating its substantial role in shaping consumer behavior.

According to Hair et al. (2010) [73], the criterion that “R² > 0.2 indicates predictive power” is met, as all R² values in this study exceed the threshold of 0.2. This confirms the scale’s high predictive accuracy for external criterion variables, supporting the validity of the predictions.

3.

Hypothesis testing

Hypotheses were tested using SPSS 20.0, and the key influence pathways were visualized in Figure 7.

The path analysis results are summarized in

Table 20. Path test results.

Path	Non-Standardized Coefficient α	Standardized Coefficient β	S.E.	C.R.	p
GIN → PAR	−0.28	−0.312	0.043	−2.983	***
RIN → PAR	−0.169	−0.147	0.057	3.526	0.003
AIN → PAR	0.257	0.207	0.073	4.129	***
FIN → PAR	0.247	0.248	0.06	−6.528	***
GIN → USV	0.09	0.119	0.032	2.84	0.005
RIN → USV	0.447	0.461	0.036	12.556	***
AIN → USV	0.207	0.199	0.049	4.204	***
FIN → USV	0.231	0.275	0.04	5.747	***
GIN → HSV	0.09	0.107	0.034	2.609	0.009
RIN → HSV	0.116	0.109	0.033	3.5	***
AIN → HSV	0.44	0.382	0.056	7.838	***
FIN → HSV	0.355	0.383	0.044	8.042	***
USV → PAR	0.45	0.379	0.086	5.235	***
HSV → PAR	0.427	0.397	0.069	6.184	***

*** Represents p value less than 0.001.

. The findings indicate:

Hypothesis H1a is unsupported: guidance intelligence does not significantly enhance consumer participation behavior.

Hypothesis H1b is also unsupported: recognition intelligence negatively affects consumer participation behavior.

Hypotheses H1c, H1d, H2a–H2h, H3a, and H3b are all supported.

4.

Test of mediating effect and direct effect

The mediating and direct effects were tested using bias-corrected 95% confidence intervals and percentile 95% confidence intervals. None of the intervals included zero, indicating significant direct effects of e-commerce virtual anchor intelligence on consumer participation behavior. Moreover, both utilitarian and hedonic shopping values exhibited significant mediating effects on the relationship between virtual anchor intelligence and consumer participation. These results are detailed in

Table 21. Non-standardized Bootstrap mediating effect and direct effect test.

Parameter	Bias-Corrected 95%CI			Percentile 95%CI
	Lower	Upper	p	Lower	Upper	p
RIN → PAR	−0.277	−0.068	0.002	−0.279	−0.07	0.002
AIN → PAR	0.11	0.411	0.002	0.108	0.408	0.002
FIN → PAR	0.129	0.366	0	0.131	0.369	0
GIN → USV → PAR	0.014	0.078	0.002	0.012	0.076	0.003
RIN → USV → PAR	0.128	0.308	0	0.125	0.3	0
AIN → USV → PAR	0.048	0.158	0	0.046	0.154	0
FIN → USV → PAR	0.058	0.17	0	0.054	0.163	0
GIN → HSV → PAR	0.01	0.076	0.008	0.008	0.073	0.012
RIN → HSV → PAR	0.02	0.09	0.001	0.018	0.088	0.001
AIN → HSV → PAR	0.119	0.277	0	0.116	0.272	0
FIN → HSV → PAR	0.097	0.224	0	0.095	0.221	0

.

5. Discussion

The hypothesis H1a, that guidance intelligence has a positive impact on consumer participation behavior, was not supported, as the two variables exhibited a negative correlation. This outcome may be attributed to the fact that guidance intelligence in e-commerce virtual anchors encompasses their proactive abilities and knowledge, which are typically crucial for facilitating consumer participation. Broniarczyk (2014) [74] found that proactive capabilities reduce consumers’ decision-making demands, thereby altering their behavioral patterns. Consumers tend to value their autonomy in decision-making and may resist being entirely guided or directed [75]. Xie (2025) [76] found that when audiences feel AI has stripped them of control, their satisfaction drops. In real-world scenarios, when virtual anchors demonstrate strong proactive capabilities—such as actively providing extensive information to preemptively resolve consumer doubts—this may paradoxically reduce consumers’ willingness to engage. This highlights a contradiction between the proactive nature of information intelligence and consumers’ desire for autonomous decision-making. The increased proactive capabilities of virtual anchors may inadvertently diminish consumers’ sense of control. Additionally, the preemptive provision of information by virtual anchors to address consumer concerns may reduce engagement behaviors to some extent.

Similarly, the hypothesis H1b, that recognition intelligence has a positive impact on consumer participation behavior, was not supported, as the two variables exhibited a negative correlation. This could be due to the fact that recognition intelligence refers to the e-commerce virtual anchors’ ability to acquire relevant information and contextualize it. While e-commerce operators can obtain partial consumer data, virtual anchors may leverage this information to identify consumer identities and preferences. However, consumers retain complete control over their personal data during transactions [77]. The loss of control over personal data can evoke profound anxiety [78]. Zhou (2010) [79] noted that privacy concerns often deter consumers from participation in transactions. The information-gathering and contextualization capabilities of virtual anchors may, therefore, heighten concerns about data privacy, thereby diminishing consumers’ willingness to engage in live-streaming interactions.

5.1. Theoretical Significance

Yang (2018) [80] defined the intelligence of non-living entities as the various behavioral capabilities that enable them to adapt to, alter, and select environments. Based on grounded theory, this study explores the concept and dimensions of intelligence in e-commerce virtual anchors from the perspective of human–computer interaction services. Following the scale development process, we constructed a measurement tool that not only builds upon previous research on AI agent intelligence but also advances related studies in this area. Additionally, we established an influence mechanism model based on shopping value, elucidating the pathways through which the intelligence of e-commerce virtual anchors affects consumer participation behavior.

Firstly, this study pioneers a new theoretical framework for analyzing intelligence in e-commerce live-streaming contexts, addressing the limitations of AI research that often prioritize technical specifications over contextual adaptation. Previous research on intelligence has predominantly focused on general AI systems (e.g., chatbots, educational platforms) with conceptual dimensions such as mechanical intelligence and analysis intelligence [11]. However, the unique characteristics of live-streaming environments—characterized by real-time interaction and a shopping orientation—have not been sufficiently accounted for in such studies. Through grounded theory analysis, this study established a four-dimensional framework of intelligence, comprising “guidance, recognition, analysis, and feedback.” This scenario-specific categorization not only fills the conceptual gap in defining “e-commerce virtual anchor intelligence” but also shifts AI research from “general theory” to “contextual application,” offering methodological guidance for future intelligence studies in specialized domains.

Secondly, this study provides the first empirical evidence of the “asymmetric effects of intelligence on consumer behavior,” thereby expanding the applicability of the S-O-R theory in AI marketing. While existing S-O-R theory-based research generally assumes the “positive driving effects of external stimuli (intelligence) on organismal responses (psychological perception)” [47], the findings reveal that guidance intelligence and recognition intelligence exhibit negative correlations with consumer participation behaviors, whereas only analysis intelligence and feedback intelligence demonstrate positive influences. This discovery challenges the linear assumption that “higher intelligence correlates with more proactive consumer behavior.” By incorporating consumer autonomy and privacy concerns, we extend the S-O-R model’s mediating mechanisms: when stimuli (intelligence) breach psychological boundaries (autonomy, privacy), organismal responses shift from “positive value perception” to “negative psychological resistance,” ultimately suppressing behavioral responses. This nonlinear “stimulus-psychology-behavior” transmission logic offers new theoretical insights for applying S-O-R theory in AI contexts and provides empirical evidence for future research focusing on “stimulus intensity thresholds” and “psychological response boundaries”.

Finally, this study developed an 18-item intelligence scale for e-commerce virtual anchors, providing a reliable measurement tool for future research and addressing the gap in the “operationalization of constructs” within this field. Existing studies predominantly rely on single-item or generic scales to measure intelligence, lacking specialized tools tailored for e-commerce virtual anchors. Following Churchill’s rigorous scale development process [48], the scale underwent grounded theory item generation, Delphi method optimization, scale refinement, reliability assessment, and validity assessment, ultimately meeting the standards of academic rigor. This scale not only facilitates the assessment of intelligence in e-commerce virtual anchors but also extends the “guidance, recognition, analysis, and feedback” dimensional framework to measure intelligence in other service-oriented virtual agents (e.g., virtual sales assistants, customer service bots), laying the foundation for cross-scenario intelligence comparison studies.

5.2. Practical Significance

With the advancement of digital technologies, virtual live-streaming anchors have become a central tool in e-commerce marketing, exhibiting intelligence comparable to that of human counterparts. Their adoption by industry leaders continues its upward trajectory [81]. This research provides actionable insights for two key stakeholders—live-streaming platforms and virtual human technology providers—facilitating the industry’s transition toward a dual-driven model of user experience and commercial value.

For e-commerce live-streaming operators, the intelligence scale developed in this study directly addresses the challenge of “selecting and evaluating virtual anchors.” Currently, due to significant price variations across virtual anchor products, companies often make purchasing decisions without adhering to scientific assessment criteria. The proposed scale, encompassing four dimensions—guidance, recognition, analysis, and feedback—enables operators to conduct targeted evaluations based on product category needs, align anchor capabilities with business objectives, and mitigate cost inefficiencies. Additionally, the scale supports dynamic performance monitoring through a monthly data-collection system, helping operators identify underperforming virtual anchors and optimize resource allocation by replacing them. Furthermore, the study confirms the mediating roles of utilitarian and hedonic shopping value. When constructing virtual anchor script databases, companies should balance “efficiency” and “engagement”: utilitarian content should focus on “discount explanations and after-sales responses” to enhance decision-making, while hedonic content should incorporate “contextual storytelling and emotional interaction” to increase user enjoyment. A dual-dimensional design maximizes consumer participation behavior.

Virtual human technology providers can leverage these research findings to inform R&D differentiation strategies, thereby preventing inefficient investments. To mitigate the negative effects of guidance and recognition intelligence, engineers should prioritize “boundary control technologies”. This involves designing guidance systems with “scenario-specific triggers” that activate after 30 s of consumer inactivity and provide manual override capabilities. Recognition intelligence should adopt “privacy-protective data collection,” limiting the information gathered to basic customer details (e.g., identity) and implementing explicit mechanisms for withdrawing consent to safeguard consumer autonomy and privacy. As positive drivers, analysis intelligence and feedback intelligence require enhanced core capabilities. Analysis intelligence should integrate large language models for “multi-dimensional dynamic analysis,” including real-time discount combinations and cross-scenario problem-solving. Feedback intelligence must overcome the limitations of “stiff voices and mechanical actions” by developing “emotionally responsive voice synthesis” and “body-language coordination technologies”. These innovations would enable virtual anchors to adjust their tone to match the situation, synchronize movements with explanations, and even retain customer preferences, thereby enhancing user participation through anthropomorphic interactions.

Finally, the findings of this study offer valuable insights for industry regulation. As virtual anchors become more prevalent, privacy concerns arising from the “misuse of intelligence” have gained increasing attention. Regulatory authorities can use these conclusions to establish a “negative list for intelligent applications,” prohibiting the unauthorized collection of sensitive information and forced consumption guidance. Additionally, incorporating the evaluation framework into service quality assessment systems will facilitate regular compliance checks. This approach aims to achieve a triple-win scenario by balancing technological innovation, commercial value, and consumer rights.

5.3. Limitations and Research Prospects

Although this study provides an empirical examination of the dimensions of e-commerce virtual anchors’ intelligence, scale development, and influence mechanisms, several limitations exist that warrant further exploration in future research.

First, the scope of the sample and background is limited, which may affect the generalizability of the study’s conclusions. Due to the rapid growth of e-commerce virtual anchors in China’s live-streaming sales market over the past two years, Chinese consumers possess a relatively advanced understanding of these virtual anchors. This study specifically focused on online texts and survey responses from users of dominant Chinese e-commerce platforms (e.g., Taobao and JD.com), excluding participants from diverse geographical and cultural contexts. This sampling strategy may limit the applicability of the results to cross-border e-commerce or more culturally diverse contexts.

Second, the study faces potential biases and measurement limitations. Researchers’ subjective perspectives influenced the identification of dimensions during the grounded theory phase. Although bias-reduction methods such as “double researcher coding” and “theoretical saturation testing” were employed, the initial operationalization of “intelligence” from a human–computer interaction perspective may not have fully captured other relevant dimensions, such as “ethical intelligence.” Additionally, the scale did not account for certain “contextual moderating variables,” such as product involvement, which could impact the intensity of intelligence’s effect on participation behavior. This limitation may lead to conclusions that do not fully account for variations in mechanisms across different contexts.

Finally, the applicability of the research conclusions requires further clarification. The study’s findings are currently limited to “service-oriented virtual anchors in live-streaming sales scenarios,” excluding other categories such as “brand promotion virtual anchors” or “customer service virtual anchors.” Additionally, moderating variables such as “virtual anchor persona types” and “live-streaming duration” were not considered in the study, which could have influenced the outcomes.

To address these limitations, future studies could focus on three key areas: first, expanding the sample to include users from cross-border e-commerce platforms and different regional contexts, while incorporating cultural dimension theory to examine how contextual factors influence the identified mechanisms. Second, including moderating variables, such as product involvement and virtual anchor persona types, to explore their impact on the intelligence–participation behavior relationship, thereby refining theoretical models. Third, extending the research scope to consider intelligence from an “ethical perspective,” exploring privacy protection, and analyzing differences in intelligence demands between virtual and live-streaming anchors through a comparative perspective. This would provide a more comprehensive theoretical framework for the industry.