How real-time interaction drives purchase intention: the mediating pathway through perceived trust in live streaming

Live-streaming e-commerce has reshaped how consumers acquire information through real-time interaction, emerging as a key growth engine in digital retail. By 2025, China's live-streaming users had exceeded 833 million, with transaction volumes surpassing RMB 5 trillion, accounting for one-third of total online retail sales (China Internet Network Information Center, 2025; State Administration for Market Regulation, 2026). However, rapid growth has not eased consumers' trust concerns; trust remains a bottleneck constraining further development. Thus, understanding and cultivating consumer trust in the live-streaming context has become a core issue for both academia and industry.

Real-time interaction is the core competitive advantage of live-streaming (Xue et al., 2020) and can be distinguished into two forms: consumer-streamer interaction (CSI) and consumer-consumer interaction (CCI). The former reduces information asymmetry through professional demonstrations and timely responses, enabling consumers to directly assess the streamer's expertise and responsiveness (Joo and Yang, 2023). The latter involves experienced consumers sharing product experiences and recommendations via danmaku thereby reinforcing social bonds and enhancing consumers' perceived personalization and entertainment value (Fan et al., 2024). Together, these two interaction types form the basis for consumer trust in the live-streaming.

Yet, literature leaves three important questions unresolved. First, CSI and CCI have been examined in isolation (Gu et al., 2023), leaving their comparative mechanisms undertheorized. Second, expertise has predominantly been treated as a static characteristic, with limited consideration of how it may function as a dynamic signal that consumers actively assess and validate through real-time interaction (Li et al., 2025). Third, prior research has not examined whether these interaction types foster trust through distinct cognitive routes, a distinction that carries theoretical weight given the Elaboration Likelihood Model's (ELM) prediction that central and peripheral-route trust differ in durability (Petty and Cacioppo, 1986). Accordingly, this study addresses the following research questions:

• Q1: How do CSI and CCI differentially influence consumer trust in live-streaming e-commerce?

• Q2: How does perceived trust affect purchase intention, and how does price sensitivity moderate this effect?

To answer these questions, this study integrates Uses and Gratifications Theory (UGT) with ELM. UGT identifies consumers' core motivations for engaging in live streaming (Fu et al., 2024), whereas ELM explains how the fulfillment of different needs translates into persuasion through its central and peripheral routes (Petty and Cacioppo, 1986). Drawing on this framework, this study proposes that CSI activates the central route, building trust through careful evaluation of information quality. Specifically, expertise is reconceptualized as a dynamically validated signal: credibility is established when knowledge is mobilized in real time and publicly tested, with responsiveness reflecting the streamer's capacity for timely and substantive replies. In contrast, CCI activates the peripheral route, building trust through heuristic cues: personalization (trust in peers transfers to the streamer) and entertainment (positive affect spills over to purchase decisions). This study further tests the boundary conditions of peripheral-route persuasion using price sensitivity as a proxy for processing motivation.

This study makes three contributions. First, it advances understanding of consumer persuasion by demonstrating statistical equivalence does not imply mechanistic equivalence. Although the effects of different interaction types on trust and purchase intention were statistically comparable, price sensitivity as a moderator reveals a mechanistic difference between the peripheral and central routes.

Second, departing from the traditional view of expertise as a static attribute, this study reconceptualizes it as a dynamic signal that must be validated through real-time interaction. Drawing on signaling theory (Spence, 1973), expertise and responsiveness are shown to independently contribute to trust formation.

Third, it extends ELM to live-streaming (Luo et al., 2024a) and supports its core prediction that the peripheral route loses effectiveness when processing motivation is high.

Existing research has examined the antecedents of purchase intention in live-streaming from two perspectives: streamer characteristics and live-streaming content features. Streamer characteristics include streamer-product fit (Park and Lin, 2020), linguistic features and speech acts (Chen et al., 2023), attractiveness and communication styles (Guo et al., 2022), and professional competence (Li et al., 2025). Live-streaming content feature research has focused on how time pressure and immersive experiences drive impulse purchase behavior (Hao and Huang, 2025). However, these two perspectives have largely overlooked live-streaming's defining feature: real-time interaction.

A growing body of research has begun to acknowledge real-time interaction as a distinct construct, differentiating between CSI and CCI. CSI centers on information quality, driving behavior through the satisfaction of cognitive needs, the cultivation of trust, and the formation of parasocial relationships (Gu et al., 2023; Zhang et al., 2022; Ma et al., 2024). By contrast, CCI emphasizes emotional contagion, fostering a sense of community, enhancing social presence, and amplifying herd behavior (He et al., 2026; Chen et al., 2025). Scholars have extended research to explore micro-level phenomena such as danmaku consistency and emotional contagion (Zhang and Ruan, 2024; Kim et al., 2026), corroborating the distinction between informational and emotional communication strategies (Luo et al., 2024b).

Despite these contributions, three substantive gaps remain. First, no study has directly contrasted the mechanisms by which CSI and CCI steer consumer behavior. Second, prior research has largely characterized expertise as a static attribute (Li et al., 2025), with its dynamic and interaction-dependent nature receiving little empirical scrutiny. Third, the boundary conditions governing ELM's peripheral route remain empirically unexplored in live-streaming, leaving a key theoretical prediction unverified.

Thus, this study employs dual-process theory, distinguishing between central and peripheral routes, to juxtapose CSI and CCI pathways, reconceptualizes expertise as a dynamic, interaction-validated signal, and introduces price sensitivity as a moderating variable to probe the boundary conditions of peripheral-route persuasion.

Existing research has approached trust formation from three perspectives: streamer characteristics, encompassing professional competence and attractiveness (Alnoor et al., 2024); interactive process, wherein real-time interaction bolsters social presence (Fu et al., 2024); and social influence, operationalized through peer interactions and social validation (Zhang et al., 2022). Recent studies have further extended this line to trust transfer mechanisms (Li et al., 2025) and their boundary conditions (Li et al., 2024).

However, these studies have not distinguished the pathways through which trust is constructed. According to ELM, trust generated through the central and peripheral routes may structurally diverge in both its cognitive architecture and longevity (Petty and Cacioppo, 1986). Therefore, this study introduces price sensitivity as a moderator to identify when peripheral-route trust formation occurs, providing empirical evidence for the model's boundary conditions.

This study integrates UGT and ELM to distinguish the cognitive routes activated by different interaction types. UGT captures consumers' informational, social, and hedonic motivations for live-streaming shopping, while ELM explains how these gratifications translate into trust via central or peripheral routes (Table 1).

In live-streaming, informational gratification triggers the central route, prompting systematic evaluation of streamer expertise and responsiveness, whereas social and hedonic gratifications activate the peripheral route, relying on heuristic cues such as social proof and affect transfer. Grounded in signaling theory (Spence, 1973), expertise is reconceived as a dynamic signal whose credibility is established and verified through real-time interaction. Both pathways jointly shape consumer trust and purchase intention, with meaningful implications for the persistence of trust formed though distinct cognitive mechanisms.

Figure 1 presents the proposed theoretical model, illustrating how real-time interaction leads to purchase intention through perceived trust and how price sensitivity moderates this process.

Under the UGT-ELM framework, CSI activates the central route by fulfilling informational needs: expertise provides the cognitive basis of the interactive signal, while responsiveness validates it in real time.

Expertise refers to a streamer's ability to recommend products based on professional knowledge and experience, a key dimension of interaction quality (Lu et al., 2024). Through the central route, expertise fosters consumer trust via rational evaluation in two ways. First, it reduces information asymmetry by clarifying product attributes, thereby lowering perceived purchase risk (Wongkitrungrueng and Assarut, 2020). Second, systematic demonstration of professional knowledge encourages parasocial trust, increasing consumer confidence in the streamer and their recommendations (Zhang et al., 2023). This trust deepens consumer immersion and engagement, strengthening CSI and reinforcing perceived expertise in a positive feedback loop (Liao et al., 2023). Moreover, consumers are more receptive to recommendations from competent streamers (Cheah et al., 2024). Based on this, we hypothesize:

Additionally, expertise may directly influence purchase intentions by simplifying complex product evaluations and lowering decision-making barriers. Professional demonstrations simplify complex product attributes, enabling more efficient informed decisions, especially for experience and credence goods that are difficult to evaluate prior to purchase (Dang-Van et al., 2023). Thus, we propose:

The real-time question-and-answer nature of live-streaming gives responsiveness its dynamic character (Zhai and Chen, 2023), with instantaneous feedback conveying product value, fostering immersion, and encouraging participation (Gao et al., 2025).

As the real-time interactive manifestation of underlying expertise, responsiveness directly addresses consumers' informational needs. Timely and substantive responses transform static expertise claims into dynamic signals subject to real-time verification, feeding evaluable information into the central route. The credibility of expertise thus hinges on interactive validation.

Responsiveness may build trust through three interrelated pathways. First, it provides immediate product clarification, enhancing information adequacy and consumer understanding, while signaling empathy and a customer-oriented stance (Zhang et al., 2023). Second, minimizing response latency fosters psychological closeness, reducing perceived social distance between consumers and streamers (He et al., 2022). Third, this closeness strengthens perceptions of authenticity and sincerity, reinforcing relational trust (Zhang et al., 2022). Based on this, we hypothesize:

Responsiveness may also directly influence purchase intention. By capturing attention peaks in real-time interaction, it can trigger impulse mechanisms that convert transient interest into purchase behavior (Gu et al., 2023). Streamers' immediate responses may evoke scarcity or fear of missing out, prompting consumers to bypass deliberative processing and accelerate purchase decisions (Sun et al., 2024). Thus, we propose:

Under the UGT-ELM framework, CCI activates the peripheral route by fulfilling personalization and entertainment needs, building trust via heuristic cues: social proof and affect transfer.

In live-streaming, informed consumers share product-related experiential insights, creating a diversified information flow from which other participants selectively extract content that they perceive as personally relevant (Zhang and Zhang, 2024; Zhang et al., 2024; Zhang and Ruan, 2024).

When consumers observe similar others endorsing a product, they use this social cue to form positive judgment of the product's reliability (Fu et al., 2024). Such a favorable product assessment may then transfer to the streamer, elevating perceptions of the streamer's own trustworthiness (Tripp et al., 2022). Thus, we hypothesize:

Personalization may also directly influence purchase intention by enhancing information efficiency. It enables real-time access to customized information (Zhang et al., 2022), reducing cognitive load and decision barriers, shortening decision time (Lv et al., 2022), and increasing receptivity to recommendations (Wang et al., 2024). Personalized filtering may also help consumers evaluate options more effectively and arrive at more informed purchase decisions (Fu et al., 2024). Thus, we propose:

Entertainment in live-streaming arises from multidirectional consumer interactions, such as danmaku (Gu et al., 2023). These interactions enable consumers to co-construct socioemotional experiences and foster collective emotional resonance (Mo and Yang, 2026).

Shared entertainment experiences may create a sense of parasocial community, fostering emotional connections with the streamer and perceptions of warmth and authenticity, thereby facilitating trust formation (Kim et al., 2026). Positive emotions may also strengthen consumer bonds and indirectly enhance trust (Hsu and Hu, 2024). Thus, we hypothesize:

Entertainment may also directly influence purchase intention through three pathways. First, via affect transfer, positive emotions generated through interaction may transfer to recommended products, stimulating purchase intention (Yu et al., 2025). Second, via flow experience, interactive entertainment may induce temporal distortion, enhancing purchase intention by deepening information processing and filtering distractions (Wang et al., 2026). Third, via hedonic extension, increased viewing duration and interaction frequency may lead consumers to view products as pleasurable extensions, stimulating desire for possession (Liu et al., 2025). Thus, we propose:

In live-streaming, trust reduces information asymmetry and facilitates behavioral conversion, serving as a primary driver of purchase intention (Li et al., 2025). This dual functionality operates via two pathways. Cognitively, it may reduce decision uncertainty (Hameed et al., 2025); affectively, it may elicit emotional commitment and psychological security, thereby facilitating purchase decisions (Alnoor et al., 2024). Thus, we propose:

In live-streaming, where promotions are frequent, price is a core decision-making incentive (Gu et al., 2024). Price sensitivity, defined as consumers' attention to price changes and the weight assigned to price in decisions (Li et al., 2024), has generated two competing views on its moderating role in the trust–purchase intention relationship.

The trust signal attenuation hypothesis posits that high price sensitivity activates systematic processing, prompting attribute-based comparisons and price calculations, thereby weakening trust's diagnostic value and its positive effect on purchase intention (Pavlou et al., 2007). Conversely, the risk-mitigation hypothesis suggests that highly price-sensitive consumers, concerned about potential losses, may rely more on trust as a risk-mitigating mechanism (Mayer et al., 1995).

Given the abundant real-time information available through live-streaming interaction (Huang et al., 2024), price-sensitive consumers tend to adopt attribute-based evaluation strategies. Accordingly, we support the trust signal attenuation hypothesis and propose:

The measurement scales were adapted from validated scales in the literature to suit the live-streaming context. The English items were translated into Chinese using a back-translation procedure. Further, two bilingual researchers verified the translation to ensure item consistency and clarity (Lu et al., 2024).

Specifically, expertise (Li et al., 2025), personalization (Kang et al., 2021), responsiveness (Ma et al., 2024), entertainment (Joo and Yang, 2023), perceived trust (Cheah et al., 2024), price sensitivity (Gu et al., 2024), and purchase intention (Fu et al., 2024) were each measured using a five-point Likert scale (1 = strongly disagree; 5 = strongly agree). Demographic control variables included gender, age, education level, and monthly consumption expenditure (Table 2).

The survey targeted consumers who had watched and purchased through live-streaming. A pilot test with 52 respondents led to the revisions of several measurement items. The main survey was administered via the WJX platform (Link to the website), a national panel of over 2.6 million registered users with demographic quotas. A screening question (“Do you frequently watch live-streaming?”) identified eligible respondents; only those who answered “Yes” proceeded. From March 3 to 31, 2025, 532 responses were collected. After excluding incomplete, straight-lined, and overly fast responses (<90 seconds), 409 valid responses remained (valid response rate = 76.9%). The sample size exceeded the minimum requirement for Partial Least Squares Structural Equation Modeling (PLS-SEM), providing sufficient statistical power (Kock and Hadaya, 2018).

The sample was roughly balanced by gender (50.3% male, 49.7% female). The age distribution was dominated by the 21–30 group (49.3%), with 89.6% of respondents aged 41 or under, closely matching the core post-90s consumer demographic. Bachelor's degree holders comprised 46.4%, and enterprise employees represented the largest occupational group (45.7%), both consistent with the live-streaming shopping user profile. Monthly consumption was most concentrated in the ¥1,001–2,000 range (30.0%), with a sample mean of ¥1,846—approximating the national average of ¥1,720 (iResearch, 2025). Relative to industry benchmarks, the sample shows strong representativeness (Table 3).

To control for CMB, this study incorporated a common method factor into the PLS-SEM model (Podsakoff et al., 2024); results are reported in Table 4. The analysis indicates that method variance accounted for only 0.2% of total variance, yielding a substantive-to-method variance ratio of 341:1. Furthermore, none of the 25 method factor loadings reached statistical significance, suggesting that common method bias does not materially affect the study's findings.

Smart PLS 4 was used for PLS-SEM analysis. Variance inflation factor (VIF) diagnostics (Table 5) shows that all VIF values fell below the threshold of 5 (range: 1.434–2.241), indicating no serious multicollinearity concerns (Liang et al., 2007).

Reliability analysis (Table 5) shows that Cronbach's alpha and composite reliability values for all variables exceed 0.700 (Cheung et al., 2023), confirming satisfactory internal consistency reliability.

Convergent validity analysis (Table 5) shows that all indicator loadings exceeded 0.700 and were statistically significant (t > 1.96, p < 0.05) (Cheung et al., 2023), and the average variance extracted (AVE) for each construct exceeded 0.50 (Fu et al., 2024), demonstrating good convergent validity. Content validity was established by consulting experts and practitioners during the scale-development phase.

Discriminant validity analysis (Table 6) shows that the square root of each construct's AVE exceeded its highest correlation with any other construct, thereby demonstrating satisfactory discriminant validity (Fu et al., 2024).

The theoretical framework and hypotheses were assessed using a PLS-SEM model built in Smart PLS 4 (Figure 1). Model fit was acceptable (SRMR = 0.05, below 0.08; Liang et al., 2007).

The direct effects of the four interaction dimensions on perceived trust and purchase intention (H1-H8) were tested using 5,000 bootstrap resamples (Table 7).

Ex positively affected perceived trust (H1: β = 0.158, p < 0.01) and PI (H2: β = 0.246, p < 0.001). Re positively affected PT (H3: β = 0.208, p < 0.001) and PI (H4: β = 0.162, p < 0.01). Pe positively affected PT (H5: β = 0.247, p < 0.001) and PI (H6: β = 0.111, p < 0.05). En positively affected PT (H7: β = 0.128, p < 0.05) and PI (H8: β = 0.252, p < 0.001). In addition, PT positively affected PI (H9: β = 0.15, p < 0.01). All hypotheses received support.

In the base mediation model, Pe had a significant direct effect on PI. After introducing PS and its interaction term with PT, however, the direct effect of Pe became non-significant (β = 0.060, p = 0.268, 95% CI [–0.046, 0.166]), whereas the PT × PS interaction was significant (H10: β = −0.088, p < 0.05, 95% CI [–0.168, −0.008]). This finding supports the trust signal attenuation hypothesis: when price-sensitive consumers have access to abundant real-time information, they tend to adopt attribute-based evaluation rather than rely on trust. The results are illustrated in Figure 2.

The mediating role of PT was tested using SPSS PROCESS (Model 4) with 5,000 bootstrap resamples and 95% bias-corrected confidence intervals.

As shown in Table 8, all four dimensions exerted significant partial mediating effects on PI through PT. Partial mediation was confirmed for Ex (indirect effect = 0.091, 95% CI [0.051, 0.138]; direct effect = 0.352, 95% CI [0.281, 0.423]), Pe (indirect effect = 0.111, 95% CI [0.070, 0.160]; direct effect = 0.256, 95% CI [0.180, 0.333]), En (indirect effect = 0.090, 95% CI [0.053, 0.132]; direct effect = 0.370, 95% CI [0.298, 0.441]), and Re (indirect effect = 0.105, 95% CI [0.065, 0.152]; direct effect = 0.236, 95% CI [0.160, 0.312]).

Conditional indirect effect analyses (Table 8) further show that the indirect effects varied across PS levels. At low PS, the conditional indirect effects of Re (0.042, 95% CI [0.008, 0.079]) and Pe (0.050, 95% CI [0.010, 0.093]) were significant. At high PS, neither effect was significant. In contrast, Ex (low PS: 0.032, 95% CI [−0.001, 0.068]; high PS: 0.004, 95% CI [−0.019, 0.028]) and En (low PS: 0.026, 95% CI [−0.006, 0.060]; high PS: 0.003, 95% CI [−0.019, 0.027]) did not exhibit significant conditional indirect effects at either price sensitivity level.

These results indicate that PS moderates the indirect effects of Re and Pe which are significant only for consumers with low PS. By contrast, the indirect effects of Ex and En on purchase intention are independent of PS.

Significant differences in the path coefficients of the interaction dimensions were tested using Smart PLS's model constraint function (Sarstedt et al., 2020); results are reported in Table 9. The effects of the dimensions were not statistically distinguishable. As presented in Table 9, none of the twelve pairwise comparisons for the paths to trust and purchase intention reached statistical significance. The 95% confidence intervals for all pairwise differences included zero, indicating that the effects of the four interaction dimensions were not statistically distinguishable.

This study supports a dual-process framework in which real-time interaction influences purchase intention through perceived trust, moderated by price sensitivity (Table 10).

First, this study demonstrates that comparable effect sizes do not inherently imply equivalent mechanisms. All four dimensions were similarly associated with perceived trust and purchase intention. However, under conditions of price sensitivity moderation, only personalization's direct association became non-significant, suggesting that statistically indistinguishable effects may operate through different mechanisms.

Second, perceived trust acts as a partial mediator, exhibiting conditional patterns highly consistent with ELM. Central-route variables differ in trust mediation (responsiveness conditional, expertise stable), whereas peripheral-route variables diverge: personalization operates via trust only under low motivation, while entertainment primarily through direct pathways.

Third, price sensitivity negatively moderates the trust–purchase intention relationship. Heightened price sensitivity appears to trigger more systematic cognitive processing, shifting consumers to favor attribute-based comparisons and reducing their reliance on trust as a peripheral heuristic cue. Consequently, the positive association between trust and purchase intention weakens as price sensitivity rises.

First, this study contributes to marketing theory by challenging the implicit assumption in prior research that similar effect sizes reflect comparable underlying mechanisms (Luo et al., 2024b). While conventional literature treats matching path coefficients as evidence of a uniform persuasion process, the conditional evidence here suggests that statistically indistinguishable effects can operate through entirely different mechanisms. Specifically, surface-level statistical symmetry does not guarantee a shared psychological route to purchase intention when environmental boundaries like price sensitivity are introduced.

Second, extending the static view of streamer expertise (Li et al., 2025) and drawing on signaling theory (Spence, 1973), this study reconceptualizes expertise as a dynamic signal that must be validated through live interaction. In doing so, the findings demonstrate that real-time responsiveness functions as an independent, parallel pillar in trust formation, alongside established expertise.

Third, this study empirically tests the boundary conditions of the peripheral-route persuasion within live-streaming. Personalization's indirect effect via trust is exclusively significant under low-price-sensitivity consumers, and price sensitivity attenuates the direct trust–purchase intention relationship. These findings provide robust empirical backing for ELM's foundational premise that peripheral-route persuasion requires low processing motivation (Petty and Cacioppo, 1986), thereby extending understanding of the peripheral route's boundary conditions in live-streaming.

First, achieving operational success in live-streaming depends on systematic selection of interaction dimensions, which platform managers must tailor to both product classification and strategic goals. Expertise yields are most effective for technically complex products that require cognitive evaluation. Real-time responsiveness functions as a robust driver of consumer trust and purchase intention, meriting priority investment. Entertainment is conversion-focused, with most of its effect bypassing the trust, and is best deployed in mid-to-late live-streaming stages to reduce decision fatigue and encourage impulse purchases.

Second, practitioners should deploy price sensitivity as a basis for customer segmentation. Among high price-sensitive segments, value-oriented signals such as competitive pricing and quality assurance should be prioritized over personalized social proof, since personalization's direct effect on purchase intention diminishes under conditions of heightened price consciousness. In contrast, for low-sensitivity segments, personalization should be harnessed to build trust, danmaku interactions encouraged to foster community, and entertainment value mobilized as a conversion mechanism.

Third, expertise should be validated through dynamic real-time interaction. Platforms should implement streamer training protocols that place particular emphasis on delivering thorough and substantive responses to consumer questions. Encouraging viewers to challenge streamers with rigorous questions allow abstract claims of professional knowledge to be publicly verified, effectively converting them into observable and credible trust signals. Finally, platforms should strengthen responsiveness infrastructure, including adequate peak-hour staffing and resource allocation, so that timely and comprehensive replies can be consistently delivered, treating responsiveness as a trust-building force rather than a secondary complement to expertise.

First, a tension exists between the process-oriented framework and the cross-sectional data. The model posits that different interaction types build trust via distinct pathways, which may yield trust with different durability and resistance to counter-persuasion. However, cross-sectional data cannot capture trust's dynamic evolution or test such pathway differences (He et al., 2026). Future research could address this using longitudinal designs or trust-violation experiments.

Second, concerns about single-source bias and common method variance persist. Despite procedural and statistical controls, shared variance may still influence the results. Additionally, measuring purchase intention rather than actual behavior leaves an intention–behavior gap. Future research could enhance robustness by incorporating platform-generated behavioral data (Mo and Yang, 2026), multiple secondary sources, and multi-method triangulation.

Third, other boundary conditions beyond price sensitivity merit exploration. Streamer type is a promising candidate: influencer streamers likely provide peripheral cues, while expert streamers may elicit central-route processing (Zheng et al., 2026). Future research should test the generalizability of these pathway effects across different streamer types.

The authors are grateful to the editor and anonymous reviewers for their insightful and constructive comments, which have significantly improved the quality of this paper.