Impact of live chat on purchase in electronics markets

Impact of Live Chat on Purchase in Electronic Markets: The Moderating Role of Information Cues Xue (Jane) Tan Department of Operations and Decision Technologies, Kelley School of Business, Indiana University, Bloomington, Indiana 47405, janetan@iu.edu Youwei Wang Department of Information Management and Information Systems, School of Management, Fudan University, Yangpu District, Shanghai 200433, China, ywwang@fudan.edu.cn Yong Tan Michael G Foster School of Business, University of Washington, Seattle, Washington 98195, ytan@uw.edu Abstract Live chat tools have emerged as a channel for fostering synchronous communication between sellers and buyers The role of live chat in the e-commerce environment, however, is largely underexplored Using granular data from Alibaba, we examine the effect of live chat on consumers’ purchase decisions After controlling for the selection process that customers with high purchase intentions are more likely to initiate live chat in the first place, we find that live chat can increase the purchase probability of tablets by 15.99% We also investigate how the effect of live chat is moderated by existing information cues: product sales volumes and seller feedback scores The substitutional or complementary patterns allow us to understand how live chat tools change the ways consumers evaluate product quality and seller credibility We observe a substitutional effect of live chat on seller feedback score such that sellers with low feedback scores benefit more from live chat conversations than sellers with high scores On the contrary, products with a high past sales volume sell better after live chat, indicating a reinforcement effect Such findings deepen our understanding of the difference between various information cues in electronic markets As one of the first systematic studies to investigate live chat, our paper contributes to web trust conceptual frameworks with empirical analyses and sheds light on practical decisions faced by e-vendors and platform designers Keywords: live chat, reputation systems, feedback score, sales volume, online word-of-mouth, e- commerce, electronic market Electronic copy available at: https://ssrn.com/abstract=2846661 Introduction The lack of trust in the computer-mediated environment has long been regarded as the biggest challenge facing e-vendors1 (Gefen et al 2003; Kim et al 2009; Pavlou and Gefen 2004) Both cognition-based factors and emotion-based factors can affect trust Cognitively, the spatial distance between sellers and buyers leads to uncertainty about product quality and seller credibility When consumers purchase online, they cannot experience the product or service before making a payment or have their questions answered in a timely fashion Further, without face-to-face interactions with the seller, consumers not know whether sellers are honestly disclosing the true quality of products and whether they genuinely care about the customers’ welfare (Komiak and Benbasat 2006; Park et al 2005) This leads to higher perceived risk and lower purchase intentions Emotionally, shopping online is less enjoyable due to the lack of human interaction In brick-and-mortar stores, the salesperson can identify the customers’ needs and provide tailored recommendations (Perrault and Brousseau 1989) Such adaptive assistance is likely to create pleasure, facilitate trust between the transacting parties, and encourage purchasing (Park et al 2005) In online shopping, consumers face a static set of information cues, and the lack of interaction reduces purchase intentions (Park and Lennon 2006) Live chat has the potential to combat the abovementioned issues in computer-mediated marketplaces because it provides a synchronous conversation that reduces information asymmetry and fosters a strong customer relationship Live chat tools allow e-vendors to communicate with customers in a real-time fashion When the live chat function is embedded in an online store, a customer can initiate a conversation with a human web assistant by clicking the “chat” button on the webpage Live chat tools have shown their effectiveness in improving customer satisfaction and boosting purchase probability According to a survey conducted by Forrester Research (Strothkamp et al 2010), “Around 44% of online consumers say that having questions answered by a live person while in the middle of an online purchase is one of the most important features a Web site can offer.” AT&T claimed to have achieved the highest satisfaction when their chat function is used (J.D Power Ratings 2013) Abt Electronics experienced conversion rates as high as 20% after introducing live chat (Wagner 2010) Wells Fargo claimed to see both high satisfaction and a An e-vendor is a supplier that sells goods or service online We interchangeably use e-vendor and sellers When referring to sellers, we imply the context of an online marketplace Electronic copy available at: https://ssrn.com/abstract=2846661 double-digit increase in conversion rates with the use of live chat tools (Strothkamp et al 2010) Despite such industry reports and anecdotes, no empirical examination has been conducted to quantify how live chat affects the decision to purchase Systematic studies are needed to address econometric challenges and accurately quantify the causal effects of live chat This work aims to bridge this gap Thus, the first goal of our research is to quantify the effect of live chat on the decision to purchase A comprehensive understanding of live chat requires an investigation of its interaction with existing information cues in online marketplaces Specifically, we focus on two information cues: seller feedback score and product past sales volume Seller feedback score has been studied extensively in the reputation system as a predominant mechanism that facilitates trust between online sellers and buyers, making it possible for strangers to transact online (Ba and Pavlou 2002) Reputation systems document feedback after transactions, serving as a type of electronic word of mouth (eWOM) (Resnick and Zeckhauser 2002) Seller feedback scores are the sums of all numerical ratings of products sold by the same seller Consumers use this measure to gauge sellers’ credibility Throughout the paper, we use reputation and feedback scores interchangeably Although reputation systems drive sellers to behave in a trustworthy manner (Walden 2000), they can also generate market inefficiency when sellers enter the market sequentially Due to buyers’ dependency on reputation measures, new sellers’ initial cost of establishing a reputation is high This potentially deters high-quality sellers to enter the market (Ba and Pavlou 2002) Feedback scores have also been criticized for inducing opportunistic behavior When an online marketplace makes the entire feedback history available to buyers, sellers, after establishing an initial reputation for honesty, may cheat occasionally (Bakos and Dellarocas 2011) Moreover, a seller may gain positive feedback from selling cheap products and use that reputation to attract buyers who intend to buy expensive products (Su et al 2006) This renders the difficulty for customers to effectively make purchase decisions based on reputation, and tools that can substitute for reputation are highly desired Thus, our second goal is to examine the substitutional or complementary patterns between live chat and seller feedback scores Another information cue is product past sales volume E-vendors usually document past sales volume on product pages to assist with consumers’ decision making (Cheung et al 2014) A higher past sales volume is believed to indicate higher quality because it reflects the collective evaluation of the crowd (Ye Electronic copy available at: https://ssrn.com/abstract=2846661 and Fang 2013) According to information cascade theory (Duan et al 2009), consumers likely form their decisions based on others’ choices, making online adoption a preferential process of attachment to a product, whereby the “rich get richer” (Banerjee 1992; Bikhchandani et al 1992) The presence of live chat may influence such a preferential attachment process because it provides an additional channel for signals from sellers to buyers Thus, our third goal is to evaluate the substitutional or complementary patterns between live chat and past sales volume To achieve these research goals, we use large-scale granular data from Alibaba, one of the largest online marketplaces in the world The data are from March 2013 to June 2013 and focus on a homogeneous market of Apple and Samsung tablets The data provide an ideal means to address our research goals First, Alibaba has a reputation system similar to that of eBay, where feedback is collected after each transaction The feedback is aggregated into reputational measures, including feedback scores, and made publicly consumable More importantly, Alibaba displays the last month’s sales volume for every product Second, Alibaba has developed its own live chat tool, Trade Manager, for sellers and buyers to communicate instantly This tool is readily installed on the web page and is downloadable as a software application for free Third, the data are obtained from the Alibaba server, with consumers’ browsing, live chatting, and purchasing histories Such granularity allows us to recover the sequence of users’ behavior and unpack the impact of live chat After controlling for potential endogeneity issues with respect to consumers’ chatting decisions, product price premiums, and the information cues, we obtain a set of findings We find that (1) live chat has a significantly positive impact on purchase On average, a live chat conversation will increase purchase probability by 15.99% However, this number will be overestimated if not controlling for the endogeneity; (2) live chat has a substitution effect in regard to seller-level feedback scores such that sellers with low feedback scores benefit more from live chat than sellers with high feedback scores; and (3) live chat has a reinforcement effect on past sales volume A product with a high sales volume is more likely to be purchased after a seller-buyer communication as compared to a product with a low sales volume Through our extended analyses and robustness studies, we show that the positive effect of live chat in boosting rates of purchases results from its ability to effectively signal benevolence and strengthen the seller-buyer relationship Electronic copy available at: https://ssrn.com/abstract=2846661 Our findings contribute to the interdisciplinary subject of online marketplaces First, our work is among the first empirical studies to take into consideration the endogeneity problem with feedback scores and past sales volume figures Prior studies with secondary data mainly treat feedback scores or sales volumes as exogenous, potentially overestimating their effects (Tadelis 2016; Ye and Fang 2013) Second, we take into consideration the endogeneity of the consumer’s decision to use live chat We develop novel instrumental variables (IVs) leveraging the clickstream data and the differential time costs between weekdays and weekends Such IV constructions can be widely applied to other studies, given the increasing availability of clickstream data Third, our results contribute to the literature on information systems and consumer behavior theories in the marketing literature, as we examine a new channel that potentially alters consumers’ perception of existing information cues that have been widely studied in reputation systems and information cascade literature Finally, the findings of our analyses can inform decisions about platform design and seller effort allocation The manuscript is organized as follows In Section 2, we review past works In Section 3, we lay out the theoretical background and develop our hypotheses We describe our data in Section In Section 5, we present our model We discuss the results in Section and show the extended analyses in Section We show the robustness checks in Section and conclude our paper in Section Literature Review In this section, we review the conceptual framework of trust models in order to map the informational artifacts we study onto previous discussions Although it is difficult to accurately draw a one-to-one mapping from empirical study to the theoretical framework, this discussion helps us show the contributions of our study and how it aligns with past work We will review several related theories and empirical findings in order to establish the nature and interrelationships of three phenomena: the seller-level variable of reputation; the product-level variable of past sales volume; and our key variable of interest, live chat conversations 2.1 Web Trust Model Ajzen and Fishbein (1975) proposed the Theory of Reasoned Action (TRA) to define and describe the relationships among beliefs, attitudes, intentions, and behaviors McKnight et al (2002), applying TRA and integrating perceptions from multiple disciplines, proposed a TRA-based Web Trust Model to Electronic copy available at: https://ssrn.com/abstract=2846661 conceptualize the trust between an online customer and an e-vendor In this integrative model, trust is defined as a composite concept that includes disposition to trust, institution-based trust, trusting beliefs, and trusting intentions Disposition to trust refers to people’s general tendency to depend on others, and institution-based trust is the perceived trustworthiness of an online platform These two factors are shown to be antecedents to trusting beliefs and trusting intentions Trusting beliefs imply specific consumer perceptions of an e-vendor’s attributes, including competence, benevolence, and integrity They are causally related to trusting intentions, i.e., the willingness to depend on the e-vendor, and eventually, to trusting-related behaviors Trusting beliefs as defined in the TRA-based Web Trust Model are similar to the “perceptions of the other” defined in Kee and Knox (1970) and “cognitive perceptions of trustworthiness” discussed in Mayer et al (1995) In the trust-based consumer decision-making model proposed in Kim et al (2008), antecedents of consumer-perceived risk are categorized into cognition-based, affect-based, experience-based, and personality-oriented factors In Kim’s framework, e-vendors’ reputation and buyers’ feedback are considered affect-based antecedents of perceived risk Our study follows the Web Trust Model in specifying well-delineated trusting beliefs: competence, benevolence, and integrity (Bhattacherjee 2002) Competence is related to the trustee’s capability to deliver the product on time and in quality condition, benevolence is about how much the trustee cares about the truster’s interests, and integrity reflects the honesty of the trustee Integrity and benevolence are considered ethical traits, with benevolence driven by altruism (Mayer et al 1995) and integrity possibly driven by utilitarian considerations The perception of seller competence, on the other hand, is often considered together with perceived risk as a cognitive antecedent to trust-related behaviors by consumers In the Web Trust Model, trusting beliefs have received the most attention from information system researchers This is because trusting beliefs can be affected by the design of web vendor systems By introducing reputation systems and displaying past sales volumes, e-vendors provide venues for customers to acquire the information they need to evaluate the e-vendors and cultivate trust Another reason for the focus on trusting beliefs is the proliferation of online marketplaces in which many e-vendors compete in the same market, within which customers hold a unified disposition to trust and an institution-based trust Understanding the process of building trusting beliefs is critical both for e-vendors to improve their conversion rates and for platform owners to leverage electronic tools like live chat to stay attractive to sellers and buyers Electronic copy available at: https://ssrn.com/abstract=2846661 2.2 Reputation Systems Online reputation systems are considered the best technological means for building trust in electronic markets (Dellarocas 2003) In an eBay-like reputation system, consumers leave a feedback rating after each transaction The feedback can be positive (+1), neutral (0), or negative (-1), and the sum of the feedback ratings is the net feedback score, one of the most popular components in a seller’s feedback profile Empirical evidence has found that most feedback ratings are positive because leaving negative feedback is costly for consumers; thus, the net feedback score is highly correlated with the number of all ratings (Resnick and Zeckhauser 2002) The findings regarding the relationship between feedback scores and price, however, are inconclusive Dewan and Hsu (2001) find that a higher feedback score is related to a higher price in the product category of postage stamps A similar positive relationship is found between feedback scores and price in the product category of dolls (McDonald and Slawson Jr 2000) Studies by Kauffman and Wood (2000) and Lucking‐Reiley et al (2007) found an insignificant relationship between feedback scores and price in the product category of coins on eBay Finally, with respect to sales, feedback scores seem to have a positive effect on subsequent purchases (Resnick and Zeckhauser 2002), but such a link is weak (Cheung et al 2014) 2.3 Information Cascade Past sales volume is another piece of information that can affect customers’ purchasing behavior, according to information cascade theory Information cascade refers to the phenomenon that, when consumers are uncertain about product quality, they follow the adoption decision of preceding individuals (Duan et al 2009) According to information cascade theory, past sales volume drives more subsequent sales, resulting in a preferential attachment process Ye and Fang (2013) investigate the role of past sales volume on new purchases, using data from eBay and Alibaba, and confirm the predictions of information cascade theory They also used eye-tracking systems in a lab experiment to show that consumers spent more time looking at the area of the page where past sales volume is displayed and that they are more likely to purchase when past sales volume is high Similarly, empirical evidence from Alibaba shows that historical sales volume has a positive impact on subsequent sales (Li et al 2010) Liu (2006) finds that box office revenue comes mainly from the volume of word of mouth (WOM) rather than the valence Similar results were also discovered in experimental settings Huang and Chen (2006) controlled for historical sales volume at Electronic copy available at: https://ssrn.com/abstract=2846661 different levels and found a positive effect of past sales volume on future sales Zhu and Zhang (2010) confirmed the role of review volume on product sales and further examined the moderating roles of product popularity and user experience Park et al (2007) examined the moderating role of involvement in how reviews affect purchase Our work contributes to this stream of literature by investigating the substitutional or complementary patterns between past sales and live chat in terms of the impact on purchase decisions A limited number of papers have compared the impact of seller reputation and product past sales volume Cheung et al (2014) compared the effectiveness of information cues from transaction data versus cues from reviews and concluded that “actions speak louder than words,” i.e., many people buying a product has more effect on conversion rates than many people writing a review for the product Another study, which uses Alibaba data, showed that historical sales volume has a significantly positive effect, while feedback score has a weak influence on subsequent sales (Peng and Song 2014) Our work contributes to this stream of research because we not only confirm the role of feedback scores and past sales volume in affecting purchase decisions but also delineate their interaction patterns with live chat 2.4 Live Chat as a Web Site Function Behavioral information systems (IS) research looks at the role of online agents in facilitating trust in electronic markets Åberg and Shahmehri (2000) developed a live chat prototype system and performed a usability study with a field trial They showed that participants in their study were enthusiastic about the live chat feature and enjoyed the human touch in their shopping experience In an empirical study matching survey data to a conversation database, live chat was found to improve newcomer adjustment in the online banking industry (Köhler et al 2011) Qiu and Benbasat (2005) compared text-based live assistance with a system using text-to-speech voice and three-dimensional avatars in order to determine the best user interface design for live chat Basso et al (2001) compared an online store with no real-time interpersonal communication with other stores offering different forms of real-time communication These studies all show the effectiveness of live assistance in building customer relationships and boosting sales In addition to live chat using human agents, computer-based recommendation agents have been studied Research has shown that communications, even with computer-based avatars that are purely technological artifacts, help consumers develop positive product attitudes (Holzwarth et al 2006; Komiak and Benbasat 2006) Further, Köhler et al (2013) examined the agent-customer interaction content and found a differential effect of Electronic copy available at: https://ssrn.com/abstract=2846661 functional content versus social content In online labor markets, Hong et al (2018) showed with granular data that live chat can serve roles of information provision and private information signaling To the best of our knowledge, however, no prior work has investigated how the effects of live chat interact with the effects of other well-studied information cues such as feedback scores and past sales volumes Our work contributes to this stream of studies by examining the substitutional or complementary patterns between live chat and existing information cues 3.1 Theoretical Foundation Live Chat Live chat can be utilized to address a lack of consumer trust, through two primary mechanisms First, the communication provides the e-vendor an opportunity to build an interpersonal relationship with customers (Köhler et al 2013; Maes 1994; Qiu and Benbasat 2005) By patiently listening to the customers’ concerns and responding in a professional and dedicated way, sellers signal to buyers their benevolence, a key factor of trusting beliefs Second, live chat effectively reduces the perceived risk of purchase by mitigating information asymmetry through information provision (Hong et al 2018) The real-time question and answer allow customers to acquire product information that is not displayed on the website, making consumers more confident that the product will meet their needs This lowers perceived risk and the need for other product quality signals 3.2 Reputation (Feedback Scores) A couple of factors explain why feedback scores have been found to be a weak signal of product quality, leading to a weak relationship with purchase (Ye and Fang 2013) Sellers’ reputation, as reflected in their feedback scores, is an aggregate measure derived from their past sales of all products (Pavlou and Dimoka 2006) Since buyers rarely leave negative ratings due to the high cost of doing so (Resnick and Zeckhauser 2002), this measure in effect serves as a proxy for the sellers’ overall volume of transactions on the marketplace Thus, feedback scores are usually considered a measure of sellers’ trustworthiness that encompasses honesty, competence and benevolence A seller usually offers a portfolio of products with varying levels of risk A customer may see that the seller has a high feedback score and equate this with a high reputation However, this reputation may have been built through the sale of completely different products from the one the customer intends to purchase This problem becomes worse if the customer Electronic copy available at: https://ssrn.com/abstract=2846661 intends to buy high-risk products For example, high feedback score accumulated primarily through selling iPad cases does not provide sufficient evidence that the seller is competent and honest as a vendor of tablets Nevertheless, the seller’s feedback score provides a good signal of the seller’s benevolence Benevolent sellers are ones who act in the customers’ interest and are likely to provide high-quality presales and post-sales services These sellers honor customers’ trust and genuinely care about their welfare Since benevolence is a trait inherent to the seller, a seller who is considered benevolent when selling iPad cases is likely to behave the same way when selling iPads If anything, the seller will behave more benevolently when selling iPads because the profit margin is likely higher Therefore, when customers shop for relatively high-risk products, feedback scores are a better signal for sellers’ benevolence rather than for honesty or competence 3.3 Past Sales Volume Unlike the seller feedback score, which aggregates data at the seller level, past sales volume reflects data at the level of a distinct product Throughout this paper, a distinct product is considered a specific model sold by a certain seller The same model sold by different sellers is considered a different product According to information cascade theory, past sales volume conveys the strongest signal about the quality of a product (Cheung et al 2014; Peng and Song 2014) Since any previously reported dishonest behavior or unsatisfactory delivery would have stopped a seller from achieving a high sales volume (Cabral and Hortacsu 2010), a consumer is likely to believe that the seller is honest in product information disclosure and competent in delivering this product as promised if the product has a high sales volume.2 Past studies show that people tend to update their trusting beliefs after their first-hand experience or observing secondhand WOM from third parties (Bhattacherjee and Premkumar 2004) Therefore, the product quality signal from past sales volume is unlikely to be substituted by a live chat conversation because the seller himself has little incentive to reveal the low quality of his product if there is any 3.4 Hypothesis Development Seller-buyer interaction has long been studied in the marketing literature (Sheth 1976) Research has shown that salespeople can increase the probability of a consumer purchase by establishing trust and satisfaction We would like to note that the high sales volume also signals benevolence just as the feedback score does These two information cues are two sides of the same coin They both drive purchase and share many similarities, yet differ significantly in abilities to signal competence and honesty 10 Electronic copy available at: https://ssrn.com/abstract=2846661 a product page over a weekend when they are less restricted in time Another important insight from the chat equation is that customers are more likely to initiate live chat when the seller’s reputation, product’s sales volume, and price are low Specifically, feedback scores, past sales volumes, and premiums all have negative and significant coefficients ( 13 =-0.0241, 14 = -0.0740, 15 = -0.00314, with p-values all less than 0.001) We not interpret other equations because they are reduced form instead of structural However, the correlations between the instruments and the endogenous variables show patterns that confirm our identifying strategy First, from the premium equation, we can see that IVpremiumij has a positive coefficient (  31 = 0.00885, p < 0.001), which means that a competitor’s price rise for similar models will drive a price hike for the focal model Second, from the sales equation, we can see that a positive shock (first-difference) results in a higher sales volume (  21 = 0.160, p < 0.001) Finally, as seen in the reputation equation, sellers spend extra effort to reach the next reputation category when they are close to the reputation boundary This can be found from the positive and significant coefficients of IVrepbound While we only report the model specified with interactions Chatij×Reputationij and Chatij×Salesij in the manuscript, we show in Appendix B the model specification without interactions to confirm the positive average effect of Reputationij and Salesij, as suggested in the literature We also would like to point out that the marginal effect of Reputationij is negative with the presence of a live chat, as the coefficient of the interaction between Reputationij and Chatij ( 5full = -0.0633, p-value