Subjective or objective: How the style of text in computational advertising influences consumer behaviors?

Abstract

In the era of computational advertising, advertising effectiveness can be measured by different metrics at different stages of the sales funnel. In the upper funnel, the click-through rate (CTR, the rate of click per impression) represents the attractiveness of the advertising; the conversion rate (CVR, the rate of conversion per click) in the lower funnel indicates the persuasiveness of the advertising. Achieving higher CTR and CVR may need distinct advertising strategies: improving CTR requires raising more consumers’ interest in the ad, which is more beneficial to publishers; boosting CVR needs the ad to inspire more consumers’ desire in the product (service), which is more profitable to advertisers. In order to study the performance of advertising texts in terms of the two dimensions and reconcile the two different goals, this paper draws on Speech act theory (SAT) in linguistics to classify advertising texts into three types (i.e., assertive, expressive, and directive), and analyzes how advertising texts can impact consumer behaviors. We further categorize the above three styles of advertising texts into subjective type (i.e., expressive and directive) and objective type (i.e., assertive). Based on a field study, we find that subjective advertising with more personalization leads to a higher CTR, while objective advertising with higher consistency with the brand information results in a higher CVR. The results suggest that firms with different marketing goals should utilize different styles of advertising texts to elicit desirable consumer behaviors during different stages of the sales funnel.

1. Introduction

Online advertising is witnessing radical changes due to the rapid development of the Internet and has evolved from interactive advertising, digital advertising to computational advertising. Computational advertising can be defined as a broad, data-driven advertising approach that relies on or is facilitated by enhanced computing capabilities, mathematical models/algorithms, and the technology infrastructure to create and deliver messages and monitor/surveil an individual's behaviors. The explosive growth in the volume, variety, and velocity of available data has changed how brands generate and deliver their content [1,2]. Computational advertising, such as native ads, provides more personalized forms and contents than traditional advertising and thus obtains more views and clicks [3]. In the online advertising market, the publishers are the supply side that provides platforms (e.g., social media) and opportunities to show ads to consumers, and the advertisers are the demand side that pays for displaying ads and attracting consumers.

However, it is getting more challenging for advertisers to use ads on social media platforms to persuade consumers to purchase because consumers are experiencing ads fatigue in every media context [4,5]. Computational advertising offers a new approach to address the challenge. As a goal-oriented communication between advertisers and consumers, computational advertising can be personalized to better meet consumers’ demands given their identities, and their different stages in the sales funnel. Therefore, computational advertising, just like mass media advertising, must possess four distinct characteristics: selling power, memory value, attention value, and readability. Advertising often conveys its pragmatic function through attractive external forms, arousing readers’ desire to buy and thus realizing the value of advertising.

The advertising with the same content but different expressions varies in outcomes [6,7]. The development of technology has enabled computational advertising to have richer metrics of advertising effectiveness than traditional advertising. Following prior literature, we use CTR and CVR as key performance metrics [8,9]. The computing advertising market is a two-sided market with publishers and advertisers. The publishers mainly use CTR to set a price per ad impression. In comparison, advertisers care about not only CTR but also CVR since CVR is linked to the actual revenue and receives more attention from advertisers [8,10]. From a consumer's perspective, CTR and CVR represent two sales funnel stages from ad impression to click and conversion in the online advertising context. CTR captures the upper funnel regarding whether the viewer notices and clicks on the ad after the ad is loaded and displayed to the viewer. CVR captures the lower funnel regarding consumers’ conversions or purchases after clicks [8].

Different from other literary styles, advertising text has the special function of drawing consumers’ attention and persuading them to buy in a very short time. Additionally, advertising text has the unique stylistic characteristics of vocabulary, syntax, and rhetoric. In other words, advertising text can be viewed as a speech act that enables consumers to grasp the advertiser's intention and take action to purchase or convert. Therefore, the mechanism can be explained by the speech act theory (SAT) to interpret the linguistic form - communicative intention - communicative effect of advertising expressions step by step. Speech acts refer to the performative function of communication, in which phrases are indistinguishably interwoven with actions. Consequently, the behavior that the message intends to prompt is central. Advertisers’ different intentions motivate consumers at several stages of the decision-making process. Based on SAT framework, speech acts are classified into five types: assertive, directive, expressive, declarative, and commissive [11]. Three of the five types are closely related to advertising on social media platforms, including offering objective information (i.e., assertive acts), exerting demands (i.e., directive acts), and conveying emotions (i.e., expressive acts) [12,13]. Specifically, in practice, even the same advertising appeal should be expressed differently based on consumers’ preferences.

Based on SAT, we can link the sales funnel with advertising expressions, and then the results are reflected in the two metrics. Expressive and directive advertising refers to the subjective part of the message, such as the feelings and expectations of the speaker, and consumers can easily understand the intention in the advertising. As a result, these two types of advertising are the communication perspective from consumers’ experiences and emotions and can be taken as subjective advertising. Furthermore, expressive and directive advertising can bring a higher CTR compared to assertive advertising. By contrast, assertive advertising provides an objective message through the communication perspective of products’ features. It can be classified as objective advertising in which the subjective intention of the speaker is more difficult to detect than expressive and directive advertising. However, objective advertising can enhance consumers’ certainty of the results when they make actual purchase decisions. Consequently, assertive advertising can lead to a higher CVR than expressive or directive advertising. The above reasoning suggests that advertisers should balance between these two metrics based on their advertising goals.

To investigate the impact of advertising texts on the effectiveness of computational advertising on social media platforms, we cooperated with a large game company that operates its own advertising exchange platform. The company produced 99 different advertising texts for a new product and launched its advertising campaign on a social media platform. Based on these unique data, we systematically evaluated the impact of advertising texts on consumers’ decision processes. This large amount of available data, including more than one million impressions and a multitude of viewers, enables an accurate analysis to compare the performance of the ads [14,15]. We classified these texts manually and checked inter-coder reliability. After encoding the variables, 30% of the data was filtered out by setting the daily input cost threshold of each advertising, and we verified the validity of the remaining 70% of the data. We further used the Heckman selection model to correct for the self-selection bias caused by the advertisers, and we used the Tobit model to correct the sparsity in the data. The results indicated that subjective advertising achieved a higher CTR, while objective advertising had a higher CVR.

2. Theoretical background

2.1. Two metrics in the sales funnel: CTR and CVR

The goals of computational advertising are to achieve a more efficient allocation of advertising resources through better targeting and improve effectiveness through enhanced ad relevance and personalization [1]. Social media platforms (e.g., Weibo, Facebook) can make more profits by setting different prices for Internet traffic. In the digital era, the flow of information has never been faster, and advertisers generally deliver accurate ads through real-time bidding [16]. Advertisers expect to save advertising costs and achieve better advertising results through the precise delivery of ads [17]. From the perspective of profits, the impact of ads can be represented not only by clicking through but also by the extent to which consumers are encouraged to make actual conversions [18].

We can better understand the advertising effectiveness through the two metrics of CTR and CVR in the sales funnel. A sales funnel, or a consumer's decision-making process, consists of the need for recognition, initial consideration, active evaluation, purchase, and post-purchase [19]. First, advertisers set up touchpoints based on historical consumer behaviors data by deciding when and where to show advertising to which consumers. Every time advertising is exposed, an impression is recorded [1,20]. Next, advertisers make personalized adjustments to the content through consumer information to attract consumers’ attention to facilitate consumers’ consideration and evaluation of products. When consumers narrow down their list of choices, advertisers expect to convince them that they are making the most suitable choices. If consumers are successfully persuaded, consumers will make purchases or conversions. Therefore, the advertiser's selection and adjustment of advertising content can enhance consumers’ decision-making process. Advertisers need to insert the product or brand information into consumers’ consideration set when consumers deliberate possible offerings to satisfy their needs or want [21,22].

In a consumer's decision-making process, there is an enormous gap between the evaluation and purchase [19,23]. What actions they will take when they purchase costs them a lot remains an important question [8,23]. We use CTR and CVR to represent the transition rate of the upper and lower sales funnel, respectively. In practice, publishers or media platforms, such as Weibo and Facebook, can use CTR to price advertising impressions in real-time bidding (RTB), indicating greater importance of CTR to the platform. In comparison, advertisers expect consumers to make purchases to create actual profits [2,8]. CTR represents attractiveness (i.e., from an ad impression to click), while CVR represents persuasiveness (i.e., from click to conversion) [8]. In summary, CTR and CVR, as commonly used metrics for measuring computational advertising effectiveness, are quite useful for publishers and advertisers because they can not only capture the attractiveness and persuasiveness of ads but also reflect the stages of the sales funnel. Prior research has studied the effect of advertising content on brand evaluation or purchase intention. As shown in Table 1, we attempt to study how the advertising texts impact these two metrics.

Table 1.

Study comparison with relevant literature.

Authors (Year)	Independent Variable		Dependent Variable		Theory	Data source	Key Findings
	Dimension	Coding Method	Evaluation	Action
Xu, A. J., & Wyer Jr, R. S. (2010).	Puffery in advertisements, media context, communication norms, and consumer knowledge	Human	Evaluations of the product		Source-Specific Communication Principles	Lab Experments	When an ad appears in a professional magazine or the ad appears in a popular magazine but readers perceive themselves to know less about the product than consumers at large, puffery increases the ad's effectiveness.
Jain, S. P., & Posavac, S. S. (2004).	Negative v.s positive comparative advertisements	Human	Evaluations, purchase intention		Consumers' advertiser attributions	Lab Experments	Advertisements perceived as carrying more negative/derogatory references to competition result in more counterarguments, fewer support arguments, lower believability, more associated perceived bias, and lower brand attitude scores.
Huang, M., Cai, F., Tsang, A. S., & Zhou, N. (2011).	Quality,Authenticity,Authority, Interestingness as WOM characteristics	Human	Resending Intention		Ripple effect, resender, and social interaction	Lab Experments	All of the WOM characteristics, which were investigated (quality, authority, authenticity and interestingness) have a significant positive impact on resenders' resending intention.
Noriega, J., & Blair, E. (2008).	Native v.s second language used in advertisings	Human	Purchase Intention		Social Cognition	Lab Experments	A native-language advertisement is more likely to lead to more positive attitude measures and behavioral intentions
Chakraborty, A., & Harbaugh, R. (2014).	Puffery by salespeople and advertisers	N.A.	Purchase probability		puffery as a cheap talk game	Game theory	Puffery pulls in some buyers who value product attributes that are talked up or emphasized while pushing away other buyers who infer that the attributes they value are relative weaknesses.
Tellis, G. J., MacInnis, D. J., Tirunillai, S., & Zhang, Y. (2019).	Information-Focused, Drama and Commercial Content in video ad	Human	Shares of online content		Self-serving motivations, social engagement	Field Data	Information-focused content has a significantly negative effect on sharing. Various drama elements arouse emotions. Prominent placement of brand names hurts sharing. Emotional ads are shared more on general platforms, and the reverse holds for informational ads.
Villarroel Ordenes, F., Grewal, D., Ludwig, S., Ruyter, K. D., Mahr, D., & Wetzels, M. (2019).	Brand-generated messages (assertive, expressive, or directive) in social media conversations	Human and automatic,corpus in English		Share of posts or retweets	Speech act theory	Field Data	The use of rhetorical styles (alliteration and repetitions) and cross-message compositions enhance consumer message sharing. As a further extension, an image-based study demonstrates that the presence of visuals, or so-called image acts, increases the ability to account for message sharing.
Current study	Subjective(expressive,directive) v.s Objective(assertive)	Human and automatic,corpus in Chinese	CTR	CVR	Speech act theory	Field Data	Results show that subjective advertising texts can lead to a higher CTR. Objective advertising texts can lead to a higher CVR.

2.2. Speech act theory in advertising

Speech act theory (SAT) is considered as the tool and method which people use to express their intentions and has become the basis for studying real-world language applications [24,25]. The theory explains how speakers (in verbal or written form) use words to convey information and carry out actions. Research on speech acts has evolved from classifying phrases and sentences (e.g., assertive, expressive, directive [25,26]) to integrating rhetoric (e.g., figures of speech [27]), intertextual meta acts (e.g., across phrases [28]), and image acts [29]. Conceptually, marketing research uses SAT as a tool to understand speakers’ intentions based on their words [30,31] or to assess genuine intentions through the cues in reviews [32,33].

A speech act is the basic unit of verbal communication, whereas traditional linguistics has frequently used symbols, words, or sentences as communication units. SAT theory believes that speech acts carry and convey messages which are more than an emotional description or objective information. For example, the advertising “that is so funny, you should try” can be considered as either a command or recommendation, depending on the scenario. The advertising context is not only a description of the situation but also conveys the advertiser's expectations or intentions for consumers to take action. According to Austin (1965) [24], speech acts can be understood as “by saying or in saying something we are doing something” (p.12), with speakers performing three parallel speech acts simultaneously: locutionary, illocutionary, and perlocutionary. The definitions and examples of each type are summarized in Table 2.

Table 2.

The definition and example of three types of speech acts.

	Locutionary	Illocutionary	Perlocutionary
Definition	The act of speaking and the sentences	The intention of speaker conveyed by sentences	The effect of sentences on receiver
Example: "That is so funny, you should try"	The speaker is very happy and hope that others will participate in	The intention here is sending out an invitation.	The receiver would join in the game if he is persuaded.

Locutionary acts are meaningful utterances (or sentences) in which people speak or state something. For example, by saying “that is so funny, you should try,” the speaker expresses a pleasant mood, hoping that other people will join in.

Illocutionary speech acts can be divided into five different forms depending on the context of the communication: assertive, directive, expressive, declarative, and commissive speech acts [11]. Assertive acts emphasize a description of the objective world and rarely use emotions; they provide correct or incorrect information, such as product details (e.g., “Our computer uses Intel processors.”) [7]. Directive acts refer to the words and sentences used when the issue calls for action or exhibits demand information. The emphasis is on guiding the consumer behavior, such as using phrases that make them act (e.g., “Let us get started! The price will be 12% off tomorrow.”). Expressive acts are descriptions of the personality or attributes of the speaker and the audience, emphasizing the expression of feelings (e.g., “I feel so excited about the product.”) and opinions (e.g., “The clothes looking good.”) [25]. Declarative acts refer to messages released by the speaker to inform the audience. These words concentrate on the fact that the information cannot be changed or argued by the listener, such as a decision that can have a direct impact on the listener (e.g., “You are fired!”). Obviously, marketers or advertisers lack the authority to perform such declarative acts on consumers in advertising texts [7]. Commissive acts create a future obligation; providers might employ them in response to a request (e.g., “door-to-door delivery is guaranteed.”) [12]. Promises made by commissive acts, however, are often difficult to achieve. In the context of computational advertising, when consumers notice the content of native advertising (paid ads made to resemble the media format in which they appear) on media platforms, they might feel deceived, have a negative attitude, and be unwilling to believe the promise. Therefore, commissive acts are not often used in promoting products [7]. This paper focuses on assertive, expressive, and directive acts in the practice of advertising campaigns. On social media platforms, to maintain the unity of the advertising form and background, native advertising adopts the form of a headline and an advertising image. This process facilitates our judging the type of advertising texts as assertive, expressive, or directive.

In the advertising context, the advertisers act as the speakers, the consumers act as the listeners, and advertising texts can be regarded as different speech acts. There are two perspectives from which advertisers may communicate with consumers: the perspective of products’ features or the perspective of consumers’ experiences and emotions. When advertisers want to introduce products or deliver brand messages, assertive advertising would be more suitable, as it provides messages in the speech act that can be judged true or false. In this situation, the advertising text can be regarded as an objective statement about a brand or product, using specific factual information to clarify the profile of brand attributes. Objective ads are informational phrases or sentences without emotion or judgment and have weaker connections to the speaker's circumstances. Therefore, assertive advertising is classified as objective advertising. Objective advertising makes claims that associate the brand with a tangible product feature and include specific factual information to substantiate the association between brand and attributes [34,35]. When advertisers care more about consumers’ experiences and emotions, expressive and directive ads are likely to be chosen, as they have stronger connections with the speaker's circumstances and facilitate the listeners’ grasp of the advertisers’ intentions. In these two types of speech acts, the listener's understanding of advertising texts is more context-dependent and open-ended than assertive acts [36].

Compared with assertive acts, expressive advertising and directive advertising enable the consumers to feel the varying emotions or intentions of the speaker more easily. By using these two speech acts, advertisers aim to stimulate a consumer's feelings or show the benefits of taking specific actions to attract consumers. Therefore, these two speech acts can be seen as subjective advertising, which is a kind of advertisers’ communicative strategy and arouses the subjective part of a consumer's perlocutionary acts.

The final purpose of advertising is to persuade consumers to buy the advertised products or services. Previous research has used SAT to examine the impact of different speech acts on consumers. For example, Villarroel Ordenes et al. [7] studied the impact of different types of speech acts on consumers’ sharing of brand information in 2018. Compared to the above literature, this research differentiates in two dimensions. First, the type of information is changed: on social media platforms, the intentions to present the brand information is different from that of traditional advertising. Brand information is published through official accounts, consumers are guided to follow and spread the information, and publishers expect consumers to pay attention to the information as much as possible. In contrast, computational advertising (e.g., native advertising) requires the content to be integrated into the media background as much as possible so that consumers cannot distinguish between the actual and the ad. Second, the dependent variable is different. From the perspective of the sales funnel, sharing is a post-purchase behavior, whereas what we need to study is how advertising plays a role in advancing the consumer's decision-making process. The locutionary act (i.e., written or spoken) that SAT proposes to use delivers the speaker's intentions through their communication style (e.g., word choice, grammar use, and sentence structure). SAT can also be broadly used as an analysis tool for sentences, paragraphs, and pictures. In this study, computational advertising on social media platforms usually uses a sentence along with a picture, and thus we only need to classify the sentences into speech acts, which reduces the uncertainty of the classification. It can be deduced that the illocutionary act is most closely related to advertising discourse.

Consumers can perceive, to varying degrees, the advertising intentions in different advertising texts. That is to say, the perlocutionary act is reflected in various stages of the sales funnel by influencing the consumer's psychological state; further, it is embodied in the chosen metrics for computational advertising. Therefore, SAT can be used to conduct advertising research to fill the gap described above [33].

2.2.1. Impact of advertising types on CTR

Speech acts manifest in both phrases and sentences and vary in the extent to which they elicit responses. Directive and expressive acts can better convey the speaker's state, including more dramatic and rich emotions. For instance, during the introduction of the product, subjective advertising emphasizes the pleasure or benefits one can obtain from the product and the specific benefits that the product can bring. In this situation, more changeable and personalized advertising texts elicit entertainment, better attract consumers’ attention and motivate evaluation of advertising contents. In contrast, objective advertising emphasizes objective and unchangeable information but does not show the effect or pleasure the product can bring, making it more difficult for such advertising to establish a dialogue with consumers. CTR serves as an indicator of attention and advertising evaluation, and it reflects the attractiveness of advertising. Therefore, in the upper sales funnel, subjective advertising may better facilitate consumers’ interests in the advertising content and generate more clicks [6]. Therefore, we state the following hypothesis:

H1: Compared to objective advertising, subjective advertising is more effective for getting a higher CTR.

2.2.2. Impact of advertising types on CVR

In the process of turning an intention into behavior, the certainty of interest, which refers to people's confidence in the results of their decisions, plays a critical role [37]. For consumers, interests with high certainty are likely to bridge the interest-behavior gap, increasing the likelihood of taking action [38]. In the process from click to conversion, the objective message conveyed by assertive speech acts (i.e., objective advertising) might allow consumers to increase the certainty of the coming results, and, following their decision to click, consumers would be more likely to convert or purchase. On the contrary, expressive and directive speech acts lack objectivity, and the affective or actional phrases are used to deliver subjective feelings related to product usage. Although directive information can express the advertiser's intentions, it is limited by the length of the advertising and rarely explains the revenue sources. The two speech acts have advantages in conveying key advertising intentions; however, in the conversion process of convincing consumers from interest to purchase, it is difficult to assure consumers with certainty about the outcome of purchasing and thus trigger consumers’ purchase in the lower sales funnel (i.e., CVR). Therefore, it can be deduced that objective advertising can generate a higher CVR than subjective advertising. Thereby, we state the following hypothesis:

H2: Compared to subjective advertising, objective advertising would generate a higher CVR.

3. Method

3.1. Data

The data set for our study came from a big game company in China that runs different kinds of games, and we used the data of a new game. This new game's marketing campaign relied on digital channels, including an online APP store and social networks, to attract new consumers and persuade them to download it. The ad campaign began on February 2, 2018, and lasted 169 days. The ads on the Weibo platform included 99 different ad texts and 249 different combinations of ad texts and pictures. From the Weibo platform, we found a total of 12,948 lines of data. The total advertising impressions on the Weibo platform reached 119,583,199 times, and the total investment was 1,466,310.74 RMB. A total of 1,359,541 ad clicks resulted in 18,326 conversions.

Since the data contained massive information, we removed those meaningless lines from the data for a precise description. The mean value of CTR of all campaigns was 1.14%, and each ad cost 113.25 RMB per day. When the expenditure of an ad was excessively low, we assured that no click or conversion happened. This led to the sparseness of the data, in which the advertising results contained a large number of zeros, and would affect the accuracy of the model's outcomes. We set the threshold of expenditure at 0.68 RMB; the proportion of ads below this threshold was 30%. We obtained the data set of 9,711 lines which was sufficiently useful and much less noisy. The ad impressions for the model below reached 119,519,545 times, and the total investment amount was 1,465,460.2 RMB. These ads obtained 1,358,964 clicks and 18,321 conversions. Filtering 30% of data had little impact on the daily impressions, clicks, and conversions. The number of advertising impressions covered a sufficient number of people to accurately analyze and measure the effectiveness of the advertising [14,15]. The correlations of variables are shown in Table 3.

Table 3.

Descriptive statistics and correlations.

Variable's name	Min	Max	Mean	SD	CTR	CVR	Money	Expose	Type	Download	Media	Design	Emotion	Days	Week	Workdays
CTR	.0000	.3683	.009893	.0108	1
CVR	.0000	1.0000	.012346	.0462	-.041⁎⁎	1
Money	.69	13733.86	151.05	510.80	.046⁎⁎	.039⁎⁎	1
Expose	18	1181816	12319.06	41043.37	.041⁎⁎	.020*	.940⁎⁎	1
Type	0	1	.47	.499	.136⁎⁎	-.066⁎⁎	.048⁎⁎	.036⁎⁎	1
Download	0	1	.22	.413	.113⁎⁎	-.092⁎⁎	-.147⁎⁎	-.151⁎⁎	.192⁎⁎	1
Media	0	1	.80	.397	.131⁎⁎	-.068⁎⁎	.098⁎⁎	.118⁎⁎	.460⁎⁎	-.052⁎⁎	1
Design	0	1	.18	.383	.118⁎⁎	-.004	.023*	.032⁎⁎	.021*	-.168⁎⁎	.094⁎⁎	1
Emotion	.006	.9984	.644	.3551	-.096⁎⁎	.002	-.130⁎⁎	-.104⁎⁎	-.218⁎⁎	.028⁎⁎	-.109⁎⁎	-.004	1
Days	0	1	.66	.472	-.049⁎⁎	-.007	-.026*	-.025*	.012	.012	.000	-.006	-.005	1
Week	1	7	4.04	2.011	.020	.004	.035⁎⁎	.041⁎⁎	-.011	-.008	-.014	.004	.007	-.601⁎⁎	1
Workdays	1	223	55.17	36.989	.102⁎⁎	-.085⁎⁎	-.244⁎⁎	-.265⁎⁎	.210⁎⁎	.371⁎⁎	.231⁎⁎	-.039⁎⁎	-.068⁎⁎	.064⁎⁎	-.036⁎⁎	1

Since we had time-series data, we could first describe the tendency of advertising effectiveness. Here we used the sum of expenditures, clicks, and conversions of all advertising campaigns in one day to calculate average CTR and CVR. While reducing the investment of advertising, CTR and CVR in the later period of campaigns changed more dramatically. In addition, the average values of CTR and CVR in a fixed time window may be different. The descriptions of CTR and CVR are given in the two figures below: (Fig. 1, Fig. 2)

Fig. 1.

Tendency of CTR.

Fig. 2.

Tendency of CVR.

3.2. Advertising effect

Before modeling the data, we classified the text according to SAT. Two independent groups comprising ten part-time adult students who enrolled in business courses and had already learned about SAT took part in the work. Interrater agreement was high (γ = 0.90), and disagreements were resolved through discussion. The resulting data consisted of 99 advertising texts, of which 39 (39.39%) texts were categorized as expressive; 13 (13.13%) as directive; and 47 (47.47%) as assertive. In the cleaned data, objective ads accounted for 53.5%; and subjective ads for 46.5%, showing almost equal data distribution. As a result, we had 52 (52.53%) subjective advertising texts and 47 (47.47%) objective advertising texts. After the classification was completed, we found that the emphases of the two types of ad texts were different. Objective advertising that was more biased toward objective texts contained more information associated with the brand, whereas subjective advertising contained more narratives associated with the product. Fig. 3 uses WordCloud to show the kinds of texts. The objective texts were more about objective information of the brand, such as the firm's name (e.g., “Wang Yi”) and the type of product (e.g., “Mobile Game,” “Wu Xia”). Conversely, subjective advertising texts consisted more of telling stories and expressing emotions, such as stories of roles like “NPC,” “Monk,” “To Bathe,” and “Hug.”

Fig. 3.

WordCloud of two text types.

Apart from the advertising texts, the data also included the date of the advertising campaign, picture and video, daily cost of the advertising, number of exposures, number of clicks, and number of conversions. We followed the same procedure as above to code the picture designs. Because eye gaze direction might affect advertising effectiveness [39], we divided the pictures into direct gaze (i.e., looking directly at the viewer) and averted gaze (i.e., looking away from the viewer). Moreover, since emotions in ads can easily affect the consumers’ choices, we added emotions of the text as a control variable. To avoid anthropic factors affecting the classification results, we used the API provided by Baidu PaddlePaddle to assess the values of text emotions and obtained a continuous variable. Descriptions and operationalizations of the variables in our research are shown in Table 4:

Table 4.

Constructs and measures.

Variable's name (name in equation)	Variable description
Dependent variable
CTR	CTR=clicks/exposures
CVR	CVR=conversions/clicks
Independent variables
Type	The Speech act theory classification of advertising text, where the objective type is assertive and is recorded as 0; the subjective type contains expressive and directive apeech acts is recorded as1.
Download	The design of ad, Direct download as 0; Indirect download as 1.
Media	The design of ad, Text with pic as 0; Text with video as 1.
Design	The design of ad pic and video cover, people in pic facing the customer as 0; people in pic not facing the customer as 1.
Control variables
Emotion	The emotion of text calculated by Baidu PaddlePaddle.
Days	The number of days from the day when this ad is exposed.
Week	The weekday when this ad is exposed.
Workdays	The day is recorded as 0 when it is holiday; the day is recorded as 1 when it is workday.
Instrumental variables
Continuity	1, when the ad is launched in the following day; 0, when the ad is not launched in the following day.
Money	The money costed by this ad in this day.
Exposure	The impressions of this ad in this day

We faced two major challenges in deriving causal estimates for the impact of advertising text style on advertising effectiveness. First, self-selection bias related to advertising campaigns may exist because companies choose to use certain advertising and avoid using others. A company might select ads with better effectiveness and not spend additional money on those with worse effectiveness. We, therefore, used the two-stage Heckman's selection correction procedure to address the influence of any such factors in the model [40,41]. Similar to prior work, we used advertising's effectiveness on day one to measure the firm's advertising adoption on day two [42]. We added a new variable, “continuity,” in our model to represent this adoption. The value of this variable was set at 1 when the ad was launched on day two and at 0 when it was not launched on Weibo. Second, the sparsity of data made it impossible to use linear regression. We used the Tobit model to solve the problem of a large number of zeros. The inverse Mills ratio (IMR) could be derived from the Heckman selection model and was inserted into the Tobit model to correct for the sparsity.

In the first stage, we regressed the continuity variables on all independent variables and instrumental variables to obtain the IMR. Regarding these company-level variables, we kept all the independent variables to the same day. In addition, we added the advertising effect (CTR, CVR) into the regression to correct for self-selection bias. We used the expenditure and impressions of advertising as instrumental variables because a consumer could not know the cost or impression numbers of an ad while the company could. A consumer's reaction would also not be influenced by the information.

In the second stage, we added the IMR as a control variable to the regression model while excluding the variables that served as instrument variables in the first stage. The results remained stable, and the interactions showed more information about ad design. The coefficient of IMR accounted for potential self-selection bias affecting the outcome variables. Our results showed that the coefficient of IMR was negative and significant (p < .001), suggesting that the selection correction term adjusted the identified effects downward. We employed Tobit models to resolve the zero observation (left-censoring) problem in many cues, permitting a more stable result [43,44]. Because of the specific data structure, we used log transformation to bring CTR and CVR closer to normal distributions. The Heckman model equations can be written as:

$$\begin{array}{c}\text{Prob}\left(\text{continuity}=1\right)\hfill \\ ={\alpha }_0+{\alpha }_{11}\text{Money}+{\alpha }_{12}\text{Exposure}+{\beta }_{11}\text{Type}+{\beta }_{12}\text{Download}\hfill \\ +{\beta }_{13}\text{Media}+{\beta }_{14}\text{Design}+{\gamma }_{11}\text{Emotion}+{\gamma }_{12}\text{Workdays}+{\gamma }_{13}\text{Days}\hfill \\ +{\gamma }_{14}\text{Week}+\mu \text{CT}{{\mathrm{R}}^{\prime }}_{\mathrm{n}}\left(\text{or}\text{CV}{{\mathrm{R}}^{\prime }}_{\mathrm{n}}\right)\hfill \end{array}$$

where Money and Exposure are two instrumental variables; ${\text{CTR}}_{\mathrm{n}}^{{}^{\prime }}$=ln(100*$CT{R}_n$+1) and ${\text{CVR}}_{\mathrm{n}}^{{}^{\prime }}$=ln(100*$CV{R}_n$+1) represent the ad campaign effect on the nth day. The second stage of the Heckman model can be written as below:

$$\begin{array}{c}\text{CT}{{\mathrm{R}}^{\prime }}_{\mathrm{n}+1}\left(\text{or}\text{CV}{{\mathrm{R}}^{\prime }}_{\mathrm{n}+1}\right)=\hfill \\ {\mathrm{f}}_1({\beta }_0+{\beta }_{21}\text{Type}+{\beta }_{22}\text{Download}+{\beta }_{23}\text{Media}+{\beta }_{24}\text{Design}+{\gamma }_{21}\text{Emotion}+\hfill \\ {\gamma }_{22}\text{Workdays}+{\gamma }_{23}\text{Days}+{\gamma }_{24}\text{Week}+\text{IM}{\mathrm{R}}_{\text{CTR}}\left(\text{or}\text{IM}{\mathrm{R}}_{\text{CVR}}\right))\hfill \end{array}$$

and the interaction model is:

$$\begin{array}{c}\text{CT}{{\mathrm{R}}^{\prime }}_{\mathrm{n}+1}\left(\text{or}\text{CV}{{\mathrm{R}}^{\prime }}_{\mathrm{n}+1}\right)\hfill \\ ={\mathrm{f}}_1({\beta }_0+{\beta }_{31}\text{Type}+{\beta }_{32}\text{Download}+{\beta }_{33}\text{Media}+{\beta }_{34}\text{Design}\hfill \\ +{\rho }_{31}\text{Type}*\text{Download}\hfill \\ +{\rho }_{32}\text{Download}*\text{Media}+{\rho }_{33}\text{Download}*\text{Design}+{\rho }_{34}\text{Type}*\text{Design}\hfill \\ +{\rho }_{35}\text{Media}*\text{Design}+{\gamma }_{31}\text{Emotion}\hfill \\ +{\gamma }_{32}\text{Workdays}+{\gamma }_{33}\text{Days}+{\gamma }_{34}\text{Week}+\text{IM}{\mathrm{R}}_{\text{CTR}}\left(\text{or}\text{IM}{\mathrm{R}}_{\text{CVR}}\right))\hfill \end{array}$$

where ${\mathrm{f}}_1$(.) is the one-sided Tobit functional form; ${\text{CTR}}_{\mathrm{n}+1}^{{}^{\prime }}$=ln(100*$\text{CT}{\mathrm{R}}_{\mathrm{n}+1}$+1) and ${\text{CVR}}_{\mathrm{n}+1}^{{}^{\prime }}$=ln(100*$\text{CV}{\mathrm{R}}_{\mathrm{n}+1}$+1) represent the ad campaign effect on (n+1)th day; $\text{IM}{\mathrm{R}}_{\text{CTR}}\text{and}\text{IM}{\mathrm{R}}_{\text{CVR}}$ are the inverse Mills ratios from the first stage of the two models, respectively.

The outcomes are shown below:

Table 5.

Estimation results after Heckman model correction.

	Estimation of main effects		Estimation of main effects and interactions
	CTR	CVR	CTR	CVR
Type	0.039841384***	-0.257017516***	0.037983016**	-0.121558079*
Download	0.06016596***	-1.986727353***	0.291395549***	-1.723613596***
Media	0.130382718***	0.534510824***	0.176875925***	0.570444102***
Design	0.110419482***	0.117880307*	-0.085007282*	0.47340821**
Type*Download			0.166716318***	0.277690827
Download*Media			-0.398295349***	-0.571480439.
Download*Design			-0.543253828***	0.029430413
Type*Design			0.044937727.	-0.60497704***
Media*Design			0.239951207***	-0.102135034
Emotion	-0.179971658***	-0.482966493***	-0.139428493***	-0.465769616***
Weekdays	-0.026466262.	0.022690075	-0.028576614.	0.014880346
Days	0.000767153***	-0.009178186***	0.000767432***	-0.009467195***
IMR	-0.739282433***	-7.74512611***	-0.72960343***	-7.406875017***

Notes:.p<0.10,*p<0.05,**p<0.01,***p<0.001

In the estimation of main effects, both H1 and H2 were verified. Subjective advertising text led to a higher CTR (${\beta }_{21}^{\text{CTR}}=0.039841384$, p<0.001), and objective advertising text led to a higher CVR (${\beta }_{21}^{\text{CVR}}=-0.257017516$, p<0.001). Evidently, the design of advertising can impact advertising effectiveness to a different extent. When the advertising did not directly provide the download button, consumers were more likely to click the advertising (${\beta }_{22}^{\text{CTR}}=0.06016596$, p<0.001); by contrast, when the advertising did provide the download button directly, consumers were more likely to download (${\beta }_{22}^{\text{CVR}}=-1.986727353$, p<0.001).

In the estimation of main effects and interactions, both H1 and H2 were verified. Subjective advertising text led to a higher CTR (${\beta }_{31}^{\text{CTR}}=0.037983016$, p<0.01) and objective advertising text led to a higher CVR (${\beta }_{31}^{\text{CVR}}=-0.121558079$, p<0.05). We also found that other elements contained in an ad could cause fluctuations in results. The results showed that ads in the format of texts with videos performed better on both CTR (${\beta }_{33}^{\text{CTR}}=0.176875925$, p<0.01) and CVR (${\beta }_{33}^{\text{CVR}}=0.570444102$, p<0.01). Another interesting result was that when people in advertising pictures gazed at the consumer rather than looking at each other or turning around, consumers were more likely to download (${\rho }_{34}^{\text{CVR}}=-0.60497704$, p<0.001), a finding consistent with prior research [39]. The design of the download button might also make a difference. By hiding the direct download button, an ad could presumably better assimilate into the media context, leading to a higher CTR (${\beta }_{32}^{\text{CTR}}=0.291395549$, p<0.001); while showing the direct download button led to a higher CVR (${\beta }_{33}^{\text{CVR}}=-1.723613596$, p<0.001). The download button also interacted with other advertising elements. Results were more significant on CTR (${\rho }_{31}^{\text{CTR}}=0.166716318$, p<0.001;${\rho }_{32}^{\text{CTR}}=-0.398295349$, p<0.001;${\rho }_{33}^{\text{CTR}}=-0.543253828$, p<0.001) than CVR (${\rho }_{31}^{\text{CVR}}=0.277690827$, n.s;${\rho }_{32}^{\text{CVR}}=-0.571480439$, p<0.1;${\rho }_{33}^{\text{CVR}}=0.029430413$, n.s). The results indicated that ad design could easily affect the click-through, but it is more difficult to affect consumers’ decisions about download or purchase.

4. General discussion

4.1. Theoretical contribution

Computational advertising via real-time bidding has received increasing attention from advertisers [45,46]. However, the difference in advertising effectiveness caused by different texts of the same product advertising has been insufficiently understood [23,47]. By distinguishing the narrative patterns of advertising and using the conceptual analysis of SAT, this research contributes to investigating the effect of computation advertising on social media platforms. While prior literature focused on cues of sincere reviews and brand messages for sharing, this study considers the classification of advertising texts and describes a framework for categorizing advertising texts based on SAT. Advertising texts are divided into subjective and objective in the advertising context [34,36]. In the context of computational advertising, advertisers can collect both immediate feedback and consumer behaviors in advertising campaigns; correspondingly, new advertising metrics, such as CTR and CVR, have been established. In addition to the meanings of the two metrics mentioned in prior research in computational advertising marketing [8], we link CTR and CVR to the sales funnel concept. We use CTR to measure an ad's attractiveness and CVR to measure its persuasiveness. The study finds that subjective advertising texts are conducive to advertising attraction, and objective advertising texts are more persuasive for consumers to take action. Our following conclusions discuss theoretical support of the findings and provide managerial suggestions about how publishers and advertisers may deliver different contents to achieve advertising effectiveness measured by two metrics. Specifically, the theoretical contributions are as follows:

(1) This study expands SAT in the advertising context. Based on the five categories of speech acts (i.e., assertive, directive, declarative, expressive, and commissive acts) in SAT, the computational advertising environment scope of SAT is expanded. Previous research focused on English brand messages for sharing. In the Chinese advertising context on the social media platform, we divide advertising texts into two types: subjective and objective. Subjective advertising texts are more personalized, while objective advertising texts are more consistent with brand messages. Each style has its advantages. Subjective advertising texts have a stronger ability to attract consumer's attention, while objective advertising texts can better motivate consumers to purchase.

(2) This study systematically explores two metrics of advertising effectiveness, namely, attractiveness and persuasiveness. Previous research on advertising on social media platforms merely focused on advertising's appeal to consumers. The effect of advertising has been enhanced through social media platforms [48,49]. However, previous research did not systematically analyze which kind of advertising texts will contribute to which goal. With the development of computational advertising, both attractiveness and persuasiveness have the appropriate measurement. Establishing a link between the concept of SAT with the consumer decision-making process in a sales funnel, this study uses CTR as a measure of attractiveness and CVR as a measure of persuasiveness. This research finds that subjective advertising has a higher ability of attraction, namely, a higher CTR; objective advertising has a higher ability of direct persuasion, namely, a higher CVR.

4.2. Management contribution

In our era of the digital economy, consumer attention has become a scarce resource that companies must acquire. On social media platforms, information flows and new topics constantly emerge and diminish. In this context, the design of computational advertising content should be more personalized. In the two-sided market of advertising, publishers or media platforms price Internet traffic by CTR, while advertisers need conversions (i.e., CVR) to make profits. From the perspective of consumers, CTR represents the degree to which an ad can arouse their interests; CVR represents its persuasive effect toward making a purchase. Although advertisers know that different texts can have distinct effectiveness, they have not understood the mechanism behind advertising texts. Nor have they known how to adapt their advertising strategy to achieve a specific advertising effect. This study examines the impact of different advertising texts on effectiveness based on the frame of SAT and using CTR and CVR as metrics. We propose the following strategies:

(1) Publishers want to reach more viewers, make a higher price of internet traffic, and gain a higher CTR. The usage of subjective advertising should be promoted to achieve the goals above, and the usage of objective advertising should be limited. When using subjective advertising, publishers should add dramatic descriptions of product details in the advertising texts, stimulate the emotions, highlight the pleasure obtained from the product, and make consumers grasp the speakers’ intentions easily. Stimulating consumers’ emotions or experiences is key, as is avoiding messages with an authoritative or mandatory tone.

(2) Advertisers want to earn profits through more purchases or conversions. CVR represents an ad's ability to persuade consumers to download or buy. To increase sales and gain profits, advertisers should use objective advertising, post news, and emphasize information about brands or agencies that lead to a higher CVR.

4.3. Limitations and future research

This study uses SAT to analyze the effectiveness of advertising on social media platforms and draws some conclusions. However, there are still some issues worth studying in the future. There are two remaining limitations that are worth further investigation. The first is the data issue regarding the collection of consumer data and effectiveness data. This study explores the impact of different types of advertising texts on advertising performance based on consumer behaviors data. In the future, consumer data with more dimensions, such as duration of watching ads or shielding ads, should be used to verify our hypothesis regarding the consumer's decision journey. Since the data used in this study is for a fixed product and a fixed media platform, more advertising data of different types of products would enable additional comparison and analysis; moreover, the validity of our conclusions regarding different product types could be examined. Second, future researchers could extend the use of SAT to other forms of media content. While this study looks at interactions of ad elements, it will be interesting to classify other rich media elements like videos using the frame of SAT.

Declaration of Competing Interest

The authors declare that they have no conflict of interest.

Biographies

Minxue Huang is a second-class professor and doctoral supervisor of School of economics and management, Wuhan University, an outstanding youth in Humanities and Social Sciences and a distinguished professor of Luo Jia, the director of Department of marketing and tourism management, the vice president of big data Research Institute of Wuhan University, the director of China marketing engineering and innovation research center, a professional editor in chief of Journal of Marketing Science, an excellent talent of the Ministry of education in the new century and a young Changjiang Scholar, the Deputy Secretary General of Marketing Research Association of Chinese universities. He mainly engaged in the teaching and research of marketing management and e-commerce. In recent years, he has published more than one hundred papers in the important Chinese journals and SSCI journals such as Journal of Marketing, Journal of Marketing Research, Journal of Business Research, Electronic Commerce Research and Applications. He has presided over one key project, four general projects and one overseas cooperation project of NSFC's major research program.

Tong Liu is a Ph.D. candidate at the School of Economics and Management, Wuhan University. He received his Bachelor's degree (2011) and received Master's degree (2017) from the School of Mathematics and Statistics, Wuhan University. His research focus on computational advertising, artificial intelligence and algorithms of recommendation systems.