Programmatic advertising in online retailing: consumer perceptions and future avenues

Purpose – Digital advertising enables retailers torely onlarge volumes ofdata on consumers and even leverage artificial intelligence (AI) to target consumers online with personalised and context-aware advertisements. One recentexampleofsuchadvertisementsisprogrammaticadvertising(PA),whichisfacilitatedbyautomaticbiddingsystems.GiventhatretailersareexpectedtoincreasetheiruseofPAinthefuture,furtherinsightsontheprosandconsofPAarerequired.ThispaperaimstoenhancetheunderstandingoftheimplicationsofPAuseforretailers. Design/methodology/approach – A theoretical overview is conducted that compares PA to traditional advertising, with an empirical investigation into consumer attitudes towards PA (an online survey of 189 consumers using an experimental design) and a research agenda. Findings – Consumer attitudes towards PA are positively related to attitudes towards the retailer. Further, perceived ad relevance is positively related to attitudes towards PA, which is moderated by (1) consumer perceptions of risks related to sharing their data with retailers online and (2) consumer perceptions of AI ’ s positive potential. Surprisingly, the disclosed use of AI for PA does not significantly influence consumer attitudes towards PA. Originality/value – This paper contributes to the literature on technology-enabled services by empirically demonstrating that ad relevance drives consumer attitudes towards PA. This paper further examines two contingencies: risk beliefs related to data (i.e. the source of PA) and perceptions of AI (i.e. the somewhat nebulous technology associated with PA) as beneficial. A research agenda illuminates central topics to guide future research on PA in retailing.


Introduction
Retailers increasingly engage in digital advertising as consumers migrate to online channels (Hennig-Thurau et al., 2010;Larivi ere et al., 2013). The use of programmatic advertising (PA) in this regard entails the ability to target consumers online in real time with personalised messages, with the help of automated purchasing of ads (Samuel et al., 2021;White and Samuel, 2019). Hence, it combines ad personalisation and the automation of advertising placement. PA spending grew globally from 68.2 billion US dollars in 2017 to 155 billion US dollars in 2021 [1]. It is predicted that by 2025, 84% of all digital ad spending in the United States will be processed via PA [2].
PA is portrayed as an ideal method to market products online (Gonzalvez-Cabañas and Moch on, 2016), because it provides the potential to fit offered products to consumer needs and secure an instantaneous response from consumers (Hoban and Bucklin, 2015;Lee and Shin, 2020). It also offers efficiency and lower costs due to automation (Miklosik et al., 2019) and enables reaching customers on the move through their smartphones. Overall, the better targeted and more attractive ads help online retailers reach consumers optimally and ultimately gain higher revenues, and the higher relevance and frictionless customer journeys are beneficial to customers as well (Malthouse et al., 2019). However, there are several critical voices that warn that targeting in general may lead to suboptimal spending, where customers who are already loyal are targeted (Nelson-Field et al., 2012;Sharp et al., 2009). This implies that retailers need to go beyond current fan-based targeting methods (e.g. Facebook likers, email or mobile apps targeting extant customers) to more needs-based targeting, which is allowed by PA. However, with PA, retailers lose control over the context where the ad will be placed, because it may land on any online website.
Against this background, we identify and address an important research gap related to consumer reactions to PA. Despite the importance of PA for practitioners and its potential to reshape online retailing, this fast-developing phenomenon has received limited research attention (Samuel et al., 2021). Extant research has focused on identifying the general benefits, characteristics and intricacies of PA (Araujo et al., 2020;Helberger et al., 2020) from a business-to-business perspective (i.e. concentrating on PA adopters and PA platforms; White and Samuel, 2019). On the other hand, studies on how consumers think and behave relative to PA are scarce. Little is known about consumers' attitudes towards PA (as an advertising practice) and their responses to retailers delivering highly personalised ads through big data and analytics (Samuel et al., 2021). Nevertheless, such insights are needed particularly in the current times, as consumers tend to be more attentive to data-related risks overall and increasingly concerned with privacy online (Kabadayi et al., 2019).
There are reasons to assume that advertisements generated through PA may evoke mixed responses in consumers (Samuel et al., 2021). Drawing on rich findings from online advertising research (Liu-Thompkins, 2019) one reason could be the so-called personalisation paradox (Aguirre et al., 2015). On the one hand, consumers may perceive PA ads as highly relevant due to their high degree of personalisation and context-embeddedness (due to the use of data which generates more fitting ads); on the other hand, the high relevance follows from the use of personal data to target the consumer with the ad. Ad relevance in PA should by default be highbut how will this relevance influence attitudes towards PA as a practice (as it indirectly signals that data are used) and the attitude towards the retailer? The current knowledge gap is a crucial drawback because a negative attitude towards the practice might also reflect on consumers' attitude towards the retailer.
This paper aims to enhance the understanding of the implications of PA use for retailers with the help of a theoretical overview that compares PA to traditional advertising, combined with an empirical investigation into consumer attitudes towards PA and a research agenda. In the empirical study, we focus on the personalisation tension of PA (Samuel et al., 2021), which refers to how consumers react to retailers using increasingly sophisticated technologies to deliver them highly personalised ads online. The technology studied is artificial intelligence (AI), which is "a system's ability to interpret external data correctly, to learn from such data, and to use those learnings to achieve specific goals and tasks through flexible adaptation" (Kaplan and Haenlein, 2019, p. 17). The empirical findings show that perceived ad relevance is positively related to attitudes towards PA. This relationship is moderated by risk beliefs associated with online data disclosure and perceptions of AI as beneficial. When ad relevance is perceived as high, consumers' high risk beliefs are not problematic, but when ad relevance is perceived as low, high risk beliefs weaken the attitude towards PA. If consumers see AI as beneficial in general, it only strengthens the relationship between ad relevance and the attitude towards PA. Against our expectations, the disclosed use of AI for PA does not significantly influence consumer attitudes towards PA. Our findings suggest that consumers implicitly assume that AI is employed when they are informed that their data are used to bring them personalised ads.
We continue the paper with a discussion of online advertising and PA and the latter's complexities for online retailers. We then introduce the conceptual model and develop hypotheses on the general effect of ad relevance on attitudes towards PA, with two moderating effects, after which we present our empirical study, the analysis and results. The paper ends with a discussion and future research recommendations.

Conceptual background
Data and personalisation in online advertising Advertising is facing changes in terms of the "constant addition of (new) media and formats, the evolution of (new) 'consumer' behaviors related to advertising, and a growing acknowledgment of extended effects of advertising" that imply changes for the future of advertising (Dahl en and Rosengren, 2016, p. 335). These developments are driven by the higher availability of increasingly detailed consumer data, which has enabled advertisers to reach the most interested consumers online (i.e. computational advertising) rather than, or in addition to, reaching large audiences offline (i.e. mass advertising; Malthouse et al., 2018). Customer data may be first-party data, such as in the purchase of an item from the retailer. Data may also be sold; for example, a retailer may buy data from a housing brokerage firm to identify movers, in which case the data are second-party. Third-party data in turn refers to information that is collected by firms selling it professionally (Malthouse et al., 2019), and this type of data (often in combination with first-and second-party data) is required to target customers online.
Using data based on consumers' online behaviour to show them highly relevant, i.e. personalised advertising is generally referred to as online behavioural advertising (see review in Boerman et al. (2017)). Personalization refers to "tailoring of message content and delivery based on data collection or covert observation of users, to increase the personal relevance of message" (Bang and Wojdynski, 2016, p. 868). Personalisation is one of the key topics in online advertising research and has been studied in relation to how different types of consumer data (e.g. consumer preferences, interests, and past and present behaviour) are used and processed to (re)target advertising (see Liu-Thompkins (2019) for a review). A typical example is online display advertising which takes the form of behavioural retargeting based on consumers' browsing behaviour (Bayer et al., 2020). Traditionally, this was accomplished through browser cookie data, which can be first-or third-party data and may capture detailed information on the individual consumer (Palos-Sanchez et al., 2019), including age, gender, location and preferences (Gonzalvez-Cabañas and Moch on, 2016).
Based on this data, advertisements are personalised, and higher degrees of personalisation imply indirect consumer benefits, e.g. more relevant and/or frictionless interactions. The benefits may also be direct, such as personalised information that helps the consumer adjust behaviour immediately (Malthouse et al., 2019), e.g. geolocation data directing the consumer to the closest store. Kumar and Gupta (2016) proposed that high level of personalisation and the associated improved relevance would become more prevalent in future advertising and even required by customers.
PA and its personalisation tension PA has emerged as "an automated big data system that allows organizations (predominantly retailers) to bid for the privilege to publish personalized online advertising in the right place, to the right people, at the right time" (Samuel et al., 2021, p. 2). The PA system entails interaction between different groups of actors: PA adopters, PA platform developers and consumers (White and Samuel, 2019). PA aims to facilitate the generation of real-time ads that match the interests of individual consumers at the exact moment when they are most likely to make a purchase or click on an ad (Palos-Sanchez et al., 2019;Yang et al., 2017); this offers means for retailers to connect to their potential/existing customers during the purchasing journey. Table 1 presents a comparison between traditional advertising and PA. The latter entails important implications for retailers and other marketers who aim to advertise their services and communicate with their customers. Samuel et al. (2021) discuss three tensions stemming from the social, technological and economic complexity of the PA system: personalisation (i.e. the need for more data to deliver more personalised ads), efficacy (i.e. the need to adopt this novel approach to advertising without understanding its true impact) and mechanisation (i.e. the need for automation to reap speed and efficiency benefits). Henceforth, we focus on the personalisation tension, as it reflects increasing consumer concerns about the use of personal data and privacy (Cooper et al., 2022;Rus-Arias et al., 2021). The data collection and use may vary from simple cookiebased data collection and behavioural tracking to top players using big data-driven AI (e.g. machine learning and custom bidding algorithms) to increase an advertising campaign's success (Samuel et al., 2021).
Reflecting the importance of personalisation tension, a survey by the World Federation of Advertisers [3] found that consumer privacy/sensitivity is a primary challenge in utilising PA data. Along similar lines, Palos-Sanchez et al. (2019) argue that PA may be invasive, because beyond the use of cookies and geolocation, PA employs algorithms to determine user interests to target them with relevant ads later, even while visiting pages unrelated to the original site where those interests were identified. Recently, Google announced that their Chrome browser will no longer support third-party cookies, which further emphasises fundamental changes in how online advertising deals with tracking and targeting consumers using data (Cooper et al., 2022). Hille et al. (2015) identify consumer privacy concerns as "consumers' apprehensions regarding how online companies collect and use their personal data" (p. 3). Consumers may experience concerns in three areas: firms collecting personal information, consumers' control over the use of personal data and consumer awareness of privacy practices (Malhotra et al., 2004). If consumers become concerned about their data, they may refuse to disclose personal data online, provide fictitious data or even avoid websites they fear misuse their data (Bandyopadhyay, 2009).

Model and hypothesis development
To understand consumer perceptions of PA, we propose a model ( Figure 1) with ad relevance as the main antecedent of customer attitudes to PA, with risk beliefs and perceptions of AI as beneficial as moderators, and attitude towards the retailer sponsoring the ad as the outcome. Next, we discuss the hypotheses.
We propose that ad relevance is positively correlated with consumer attitudes towards PA, because it is likely that consumers who find an ad relevant also attribute some of its usefulness to PA (as retailers are using their data to provide a highly personalised ad). In online advertising, perceived ad relevance is found to predict consumer responses Kim and Huh, 2017;Liu-Thompkins, 2019). Supporting this hypothesis, Palos-Sanchez et al. (2019) proposed a direct relationship between consumer attitudes and relevance/usefulness of PA. Moreover, research in service settings (i.e. recruiting) shows that when the use of personal data leads to a positive outcome, the perception of privacy invasion is lower than that for individuals not experiencing a positive outcome (Fusilier and 1980). This implies that PA is likely to be more positively viewed if relevance is high. Consequently, we propose that ad relevance leads to consumers' being more positively attuned towards PA: H1. Ad relevance has a positive impact on consumer attitudes towards PA.
Although consumer attitudes towards PA may be higher if ad relevance is high, if risk beliefs are high, we propose that the attitude towards PA may decrease. In other words, we suggest that consumers' general risk beliefs related to sharing their data online (Malhotra et al., 2004) moderate the influence of ad relevance on consumer attitude towards PA. This is because ad relevance in an online setting implies that the advertising retailer knows or appears to know about consumer preferences/needs. The riskier the consumers perceive handing over their information to be, the more likely it is that ad relevance will raise suspicions about PA. Earlier research has reported a negative correlation between risk beliefs and consumer intentions to share data with firms (Li et al., 2011;Malhotra et al., 2004), along with consumer privacy concerns leading even to ad avoidance (Ham, 2017;Jung, 2017). However, in the era of PA, much of the sharing may take place elsewhere, prior to the focal firm targeting the customer. Hence, we propose that risk beliefs about sharing data with online retailers moderate the relationship between ad relevance and consumer attitude towards PA: H2a. Higher risk beliefs weaken the relationship between the ad relevance and consumer attitudes towards PA.
Moreover, we propose that the perceptions of AI as beneficial moderate the relationship between ad relevance and consumer attitudes towards PA. Studies (Liljander et al., 2006;Parasuraman, 2000) have shown that consumers' positive view of a particular technology influences their attitudes towards using that technology. While studying customer intentions to adopt AI services, Flavi an et al. (2021) found a positive impact of technology optimism on attitudes towards technology use. Whereas Flavi an et al. (2021) and others studied general technological optimism, we investigate customer perceptions of AI as beneficial (Tussyadiah and Miller, 2019). We propose that if consumers perceive AI as beneficial, ad relevance has a heightened impact on attitude towards PA. This is due to ad relevance being viewed as a reflection of positive (AI) technology outcomes, leading to more positive attitude towards PA. Hence, we postulate the following: H2b. Higher perceptions of AI as beneficial strengthen the relationship between the ad relevance and consumer attitudes towards PA. Finally, consumer attitudes towards PA reflect consumer perceptions of the practices employed to show them ads (e.g. the extent to which it was acceptable that consumer data were used to show highly personalised ads) (Jin and Lutz, 2013). Although Schwaig et al. (2013) proposed that consumers' general attitudes towards information use practices lead to consumer intentions to block the use of their data, it is also likely that if consumers are positively attuned towards PA, it reflects positively on their attitude towards the retailer. Therefore, we hypothesise the following: H3. Consumer attitudes towards PA positively impact their attitude towards the retailer.
As depicted in Figure 1, our model also includes several covariates (e.g. age, gender, education and shopping frequency).

Study design
To collect data, we employed a 1X2 between-subject experimental design in which we manipulated the employment of AI for PA. Participants were instructed to imagine that they wanted to start their own business, an online store for plants, for which they were planning to set up their own website. Participants who indicated they could imagine themselves in this situation were then randomly assigned to one of two conditions: PA without AI and PA with AI. In both conditions, participants were asked to imagine casually browsing through their social media and coming across an advertisement for an online course on how to set up a website. Participants were then told that the online retailer who had sponsored the ad could show them such personalised advertising because it used data that had been collected about them online (PA without AI) or because it used AI that analysed data that had been collected on them online (PA with AI). Participants who indicated they did not carefully read the information were not allowed to continue answering the survey. Our goal with this design was to simulate the PA's "best match" between a consumer in a specific context and a suitable ad (Yang et al., 2017) by offering a situation in which the consumers would be directly interested in the ad (Samuel et al., 2021). We kept the media and format of the ad as neutral as possible while mimicking the consumer context (i.e. consumers planning to set up their own website and seeing an ad for such services on social media). We also intentionally did not provide any additional information as to which data were collected and by whom, nor what kind of AI was employed or how. This is because we only wanted to sensitise participants to the general idea of PA without or with AI rather than its specifics which typically elude consumers. The online retailer employed in our study was fictitious to avoid any potential associations or relationships with existing online retailers.
We employed established scales (see Appendix 1) to measure the attitude towards the retailer (adapted from MacKenzie and Lutz (1989)), attitude towards the PA (adapted from Schwaig et al. (2013)), the ad relevance (adapted from Laczniak and Muehling (1993)), data risk beliefs (adapted from Malhotra et al. (2004)) and the extent to which respondents perceive AI as beneficial (adapted from Tussyadiah and Miller (2019)). Multiple screening questions (e.g. a CAPTCHA task to avoid bots) and attention checks were implemented in the survey to ensure high response quality (e.g. in a set of questions, one of the items was "I am a robot from outer space," and respondents who agreed to that statement were automatically excluded from the survey).

Participants
Participants were recruited through Amazon Mechanical Turk, an established data collection platform for social sciences (Goodman and Paolacci, 2017). Respondents residing in the United States with an approval rating above 95% were asked to participate in the study for a $1.00 compensation. Out of 200 respondents who requested compensation, 11 had to be rejected for entering an invalid completion code. Thus, our sample consisted of 189 respondents: 60% were male, 40% were under 35 years old and 56% held a bachelor's degree. Further, 94% of the respondents used social media at least once a day, and 61% shopped online at least once a week. Participants were assigned randomly to either the PA without AI (n 5 97) or the PA with AI (n 5 92) scenario.

Manipulation checks
We asked the participants in both scenarios two manipulation check questions (each on a seven-point Likert scale anchored by "Strongly disagree" to "Strongly agree"). Independent samples t-tests show that there was no statistically significant difference between the two conditions on the data-focused manipulation check ("ONLINE RETAILER uses my data to show me personalised advertising"): M PA without AI 5 5.73 (SD 5 0.92), M PA with AI 5 5.66 (SD 5 1.14), p 5 0.59. However, a statistically significant difference was found on the AI manipulation check ("ONLINE RETAILER uses Artificial Intelligence (AI) to show me personalised advertising"): M PA without AI 5 4.88 (SD 5 1.52), M PA with AI 5 5.90 (SD 5 1.18), p < 0.01. Interestingly, while the PA with AI manipulation functioned in the envisioned direction (i.e. respondents exposed to the PA with AI scenario had a significantly higher mean than respondents in the PA without AI scenario), the mean for respondents in the PA without AI condition is still high. This implies that although the scenario made no mention of AI, respondents have a higher-than-neutral perception of AI being used when data are employed for personalised advertising.

PLS-SEM results
We estimated our conceptual model (see Figure 1) in Smart PLS v3.3.3 using the consistent PLS algorithm with 5,000 bootstraps (complete bootstrapping, bias-corrected and accelerated bootstrap; Hair et al., 2017). On each exogenous variable (i.e. attitude towards PA and towards the retailer), we controlled for the impact of our manipulation (0 5 PA without AI condition; 1 5 PA with AI condition), age (0 5 Younger than 35; 1 5 35 and older), gender (0 5 Female; 1 5 Male), education (0 5 Less than a Bachelor's degree; 1 5 At least a Bachelor's degree) and shopping frequency (0 5 Shops online weekly; 1 5 Does not shop online weekly). Two models were estimated, a moderation-free model to test H1 and H3 (i.e. the direct effects) and a moderation model to test H2 (i.e. the interaction effects). Additional analyses were then carried out to explore the role of AI in PA.

Measurement model
As can be seen in Appendix 2, convergent validity is established since the outer loadings for each construct are above the threshold of 0.70, and the average variance extracted (AVE) is above the threshold of 0.50 (Hair et al., 2017). Internal consistency reliability is also established since, for each construct, composite reliability and Cronbach's alpha values are above the threshold of 0.60 (Hair et al., 2017). Finally, discriminant validity is established based on the heterotrait-monotrait (HTMT) criterion, as all HTMT ratios are lower than the 0.85 threshold, and none of the bias corrected confidence intervals for any relationship in the model includes the value 1 (Hair et al., 2017). In sum, the measurement characteristics of the constructs employed in our analysis are adequate, so we can proceed to assessing the results of the structural model.

Structural model
The inner variance inflation factor (VIF) values for all combinations of endogenous and exogenous constructs are below the threshold of 5, indicating that collinearity among the predictor constructs is not a critical issue in the structural model (Hair et al., 2017). The R 2 values of the endogenous latent variables are 0.73 in the moderation-free model and 0.81 in the moderation model for attitude towards the PA, and 0.80 in both models for attitude towards the retailer. To assess the predictive relevance of the model, we ran a blindfolding procedure with an omission distance [4] of 8 which yields Q 2 values considerably above zero: 0.49 in the moderation-free model and 0.53 in the moderation model for attitude towards PA, and 0.63 in both models for attitude towards the retailer. Appendix 3 provides an overview of R 2 and Q 2 and presents the f 2 and q 2 effect sizes for both models. While Hair et al. (2017) advise against the use of fit statistics in PLS-SEM, they condone a conservative approach to the standardised root mean square residual (SRMR) fit measure. The estimated model SRMR value in both our models is 0.04 below the 0.08 cut-off point, indicating good fit.

Hypotheses testing
To test our direct effect hypotheses, we consulted the bias corrected bootstrapped confidence interval for each path in the moderation-free model presented in Table 2: if 0 is not included in the confidence interval, the path coefficient is significant at 0.05 significance level. Results show a significant, positive path coefficient from ad relevance to attitude towards PA (0.44, [0.23, 0.62]) with a medium to strong effect (f 2 5 0.29) in support of H1. Attitude towards PA also shows a significant positive path coefficient (0.49, [0.27, 0.73]) to attitude towards the retailer with a medium to strong effect (f 2 5 0.33) in support of H3. The total effect of ad relevance on attitude towards the retailer (0.74, [0.56, 0.90]) is significant. Both the indirect path from ad relevance to attitude towards the retailer through attitude towards PA (0.22, [0.10, 0.36]) and the direct path from ad relevance to attitude towards the retailer (0.52, [0.31, 0.71]) are significant and positive. Thus, we find evidence for a complementary, partial mediation of ad relevance on consumer attitudes towards the retailer through their attitude towards PA. In terms of the control variables, there is a significant path coefficient from age to attitude towards retailer (i.e. older consumers are less positive about the retailer) and from online shopping frequency to attitude towards PA (i.e. consumers who shop less frequently online are more positive about the use of PA to show them ads).
To test the hypothesised interaction effects, we compute two interaction terms with the two-stage moderation procedure recommended for hypotheses testing in Smart PLS (Hair et al., 2017). We then consult the bias corrected bootstrapped confidence interval for each path in the moderation model presented in Table 3. Both interactions, ad relevance 3 risk beliefs 0.28, (0.15, 0.43) and respectively ad relevance 3 AI beneficial 0.15, (0.07, 0.23), show positive, significant path coefficients on attitude towards PA (with a strong f 2 5 0.36 and respectively medium f 2 5 0.16 effect). These significant interaction effects are depicted in Figure 2. Mirroring H1, the simple slopes in Figure 2 show a positive relationship between ad relevance and attitude towards PA (i.e. the more relevant the ad, the more positive are the consumer attitudes towards the use of PA to show them that ad). However, the simple slopes in Figure 2b show that for lower general data risk perceptions, the relationship between ad relevance and attitude towards PA is not strengthenedin support of H2a. In contrast, the simple slopes in Figure 2a show that for higher levels of AI perceived as beneficial, the relationship between ad relevance and attitude towards PA is strengthenedin support of H2b. Appendix 4 provides an alternative representation of the interaction effects with bar charts.

Additional analysis
A dummy variable for the manipulation (0 5 PA without AI condition; 1 5 PA with AI condition) was used as a control variable in all models, but it did not yield a significant influence on any of the outcome variables. Nevertheless, we conducted some additional analyses to test differences in the entire model for two subsamples: participants exposed to PA without and respectively with AI. This type of comparison is possible through multigroup analysis in Smart PLS (PLS MGA), a non-parametric significance test for the difference of group-specific results that builds on bootstrapping results (Hair et al., 2017). The results show a statistically significant difference (p < 0.01) in the path coefficients for the relationship between attitude towards PA and attitude towards the retailer. Specifically, the path A. Ad relevance x Risk beliefs  Attitude towards PA B. Ad relevance x AI beneficial  Attitude towards PA

Note(s): SD = Standard deviation;
In each visualization, the middle line represents the relationship between ad relevance and attitude towards PA for an average level of each moderator and the other two lines represent the relationship at higher (i.e., mean +1 SD) or lower (i.e., mean -1SD) levels of each moderator Simple slope analysis visualizations of interaction effects coefficient for this relationship in the moderation model estimated on the sub-sample exposed to PA without AI (n 5 97) is significantly higher than in the subsample exposed to PA with AI (n 5 92), while both paths are significant in their respective models.

Discussion
Our study examines PA, a way for retailers to reach potential consumers online with the efficient automised ad placement based on bidding (Samuel et al., 2021). To date, little is known about consumer attitudes towards PA and towards the retailers who employ PA. Hence, our results complement the largely conceptual field of PA research with a predominant business-to-business perspective (Araujo et al., 2020;Helberger et al., 2020;Samuel et al., 2021). In this experimental study, we exposed participants to a personalised online ad, creating a scenario emulating PA from a consumer's point of view. Our findings corroborate earlier studies (e.g. Kim and Huh, 2017) and show that if consumers find an online ad relevant, they are more likely to have a positive attitude towards PA having been used. It means that the relevance of an ad justifies the use of consumer data. In our study, ad relevance not only had an influence on the attitude towards PA but also had a direct influence on the attitude towards the retailer, supporting the important role of relevance that has been reported in other online advertising contexts (Kim and Huh, 2017;Hayes et al., 2020).
We further examined two contingencies, one related to the source of the PA (i.e. consumer's general risk beliefs about sharing data with retailers online) and the second related to a novel-yet-nebulous technology often associated with PA (i.e. consumer general perceptions of AI as beneficial). Our results show that risk beliefs moderate the relationship between ad relevance and attitudes towards PA. Particularly, the relationship between ad relevance and attitudes towards PA is weakest when risk beliefs are high and ad relevance is low. Our results also show that perceptions of AI as beneficial moderate the relationship between ad relevance and attitudes towards PA. Specifically, the more consumers perceive AI as beneficial, the stronger is the relationship between ad relevance and attitudes towards PA. These findings are in line with results from previous research (Flavi an et al., 2021;Liljander et al., 2006;Parasuraman, 2000) suggesting that consumers' positive views on a particular technology influence their attitudes towards using that technology. However, whereas earlier research has investigated consumers' own technology use, we demonstrate that these findings extend to retailers' use of technology (in terms of data being used to create PA) and reflect positively on consumer attitudes towards the retailer.
Consumers seem to assume that AI is used to personalise ads, not differentiating between data use and AI. Surprisingly, the explicit use of AI in PA did not have detrimental effects on consumer attitudes towards PA or towards the retailer, which is an optimistic finding for retailers aspiring to increase their use of AI in online advertising. Furthermore, the results suggest a significantly weaker impact of consumer attitudes towards PA on attitudes towards the retailer when AI is explicitly mentioned compared to when it is not. It may be that AI use awakes some concerns about the retailer behind the ad. This result hints at the need to study in more depth consumer perceptions of AI being employed for PA. Our manipulation checks show that when consumers are made aware that their data are used to show them ads (without an explicit mention of AI), they still have a higher-than-neutral perception that AI is involved. The level of transparency (informing participants in our scenarios that the highly personalised ad was created based on their personal data) may explain this finding. Nevertheless, that consumers think that AI is involved somewhat by default is unexpected since most retailers are just starting to explore using AI for PA.

Managerial implications
Since PA requires third-party data or use of data by third parties, it can be argued that consumer attitudes towards PA can potentially reflect negatively on attitudes towards the ad and the retailer. However, according to our findings, if retailers succeed in personalisation by showing relevant ads, then consumers will most likely have a positive attitude towards the retailer. Still, retailers need to respect consumers' need to protect their data, which requires considerations of the kinds of data to be collected, the purposes the data are used for, and with whom the data are shared (Martin, 2016); this will ensure that consumers are willing to share their data in the future (White and Samuel, 2019). If retailers wish to cultivate positive consumer attitudes by employing PA, apart from ensuring that ad relevance for the consumers is on a high level, they need to educate consumers about the general benefits of employing AI as well as monitor and address consumer-perceived risk related to data sharing.
Regarding the disadvantages of PA, whereas it aims to be temporally precise (identifying the customer at the point of purchase) at low cost, the advertiser loses control of the context in which the ad is placed. Especially for sensitive products (e.g. related to sexuality or health) or for retailers promoting a strong ideology or being at the high-end scale, choosing appropriate media may be a wiser option than letting bidding systems place the ads haphazardly online. This may be particularly prevalent in the current turbulent times, such as during a pandemic or political uprising, where potential landing pages may be detrimental to the retailers' brand image. Moreover, retailers may need to monitor how the new customers attracted with the help of PA score in the long run in terms of profitability and how existing customers perceive PA.
Avenues for future research Starting from our findings on ad relevance, data concerns and AI perceptions, and expanding with the recent literature on online advertising (see Table 1 in the conceptual background), we divide the research agenda into two major themes for the future of service retailing and PA use. These themes are briefly discussed below, with research questions suggested in Table 4.
Theme 1: Consumers in the era of PA and AI There are multiple interesting avenues for future research, ranging from consumer privacy concerns to other consumer responses. For example, ad relevance (one of the main drivers of consumer attitudes towards PA in our study) is seen as an outcome of data use and personalisation (Aguirre et al., 2015), and some research has investigated ad relevance as an antecedent to privacy concerns and ad avoidance (Jung, 2017), which may be particularly relevant in case of PA. Such negative consequences of PA offer one interesting avenue for further research along with more positive outcomes that PA may have. Furthermore, it would be useful to determine what kind of self-defence (e.g. throw-away profiles) or even sabotage (e.g. providing false data) methods consumers may engage in response to PA. Other interesting positive outcomes of PA, such as omnichannel loyalty and purchase behaviours, should be likewise studied.
While the use of AI in online retailing is likely to grow, there is currently little research into how consumers perceive retailers' investment in AI (e.g. in relation to data, algorithms, etc.) and to what extent they understand and accept its use (Puntoni et al., 2021). In our study, we examined the moderating role of perceptions of AI as beneficial, and we found a positive effect on attitudes towards PA. However, we also found that consumers are still somewhat confused about AI and that the border between AI and data remains blurry. An interesting area for future study is thereby how the (perceptions of) employment of such technologies in advertising influence consumers and in turn impact retailers.
Underlying the developments in advertising is datafication, which refers to "the collection, databasing, quantification and analysis of information, and the uses of these data as resources for knowledge production, service optimization, and economic value-generation"  (Flensburg and Lomborg, 2021, p. 1). Datafication raises several important questions regarding consumers' and citizens' reactions to use of data by service providers. Data serve as the fuel for PA, and in our research, we examined one related factor: the role of data risk beliefs. Future research can examine other consumer attitudes to different types of data being collected and consider how general attitudes towards datafication evolve (e.g. through longitudinal field studies). Intelligent advertising ("consumer-centered, data-driven, and algorithm-mediated brand communication" (Li, 2019, p. 333)) may be the next step of digital advertising and by inclusion PA. Intelligent advertising brings along a new type of personalisation by prescribing user needs and wants in real-time context to recommend offerings with the help of AI technologies such as machine learning and voice automation. Consumer attitudes towards these prescriptive techniques offer interesting research avenues.
In this regard, one particularly relevant field to be studied is the linkage between AI and sustainability (see, e.g. the items used to assess AI as beneficial in our study). One question could be: will consumers balance the positive and negative consequences of AI as a constructive force that helps solve big-scale problems, such as global warming, but also as a potential energy-consuming factor that causes problems? Simultaneously, the use of digital services requires large quantities of energy due to data storage, transfer and use. Hence, an equally relevant area for the benefit of retailers is the environmental impact of online advertising, including PA (e.g. P€ arssinen et al., 2018), because future retailers urgently need to understand and assess energy consumption and CO 2 emissions of their online activities. Simplifying the production systems, lowering the number of layers between creating and delivering ads, and reducing the data load by shortening and simplifying them would be some options for how to reduce the CO 2 footprint of online advertising.
Theme 2: AI-driven advertising and PA for retailers A pervasive theme is the impact that new advertising strategies have on retailers. Although targeting customers with the help of PA and AI is attractive due to cost-effectiveness, it may lead to negative outcomes. On the one hand, precision may entail a drastic drop in reach (Fulgoni, 2018;Nelson-Field et al., 2012). If only current customers receive the brand messages, the targeting may lead to stagnation of the customer base and reduce sales (Nelson-Field et al., 2012). On the other hand, microtargeting that increasingly predicts personal needs and wants entails ethical challenges, both in restricting customer choice and in terms of (over-)collection, use and potential sharing of data.
In the last decades, advertising has expanded from a controllable ecosystem with stable, selected partners to one where parties are brought together through automation and, with large user numbers and amounts of data, hold considerable power. This raises several questions regarding the political, social and practical influence of the data giants (e.g. Facebook/Meta, Google/Alphabet) in shaping the ecosystems as well as their obligations and rights vis-a-vis other, sometimes small, and possibly local, players, such as small and medium-sized retailers. In our research, we intentionally did not provide any additional information as to what data were collected and by whom (i.e. the retailer, a third party, etc.). Ethical questions about how consumer data are monetised arise (Breidbach and Maglio, 2020), and further research into how consumers react to trusted service providers potentially selling and buying their data is needed.
Development in advertising and technology is inevitable, and service retailers, similar to all marketers, must stay on track with technological advancements, the opportunities these offer, and other marketing strategies and tactics. NA 5 Not applicable f 2 effect size assesses an exogenous construct's contribution to a latent variable's R 2 value q 2 effect size assesses an exogenous construct's contribution to an endogenous variable's Q 2 value 0.02, 0.15 and 0.35 indicate a small, medium, and respectively large effect (Hair et al., 2017)