How TV Ads Influence Online Shopping

Jura Liaukonyte,1 Thales Teixeira,2 Kenneth C. Wilbur3

April 7, 2014

Media multitasking distracts consumers’ attention from television advertising, but it also enables immediate and measurable response to advertisements. This paper explores how the content of television advertising influences online shopping. We construct a massive dataset spanning $4 billion in advertising expenditures by 20 brands, online shopping behavior at those brands’ websites, and content measures for 1,269 distinct television commercials. We use a quasi- experimental research design to estimate how advertising content influences changes in online shopping data within two-minute pre/post windows of time. We also measure the effects within two-hour windows of time using a difference-in-differences approach. The findings show that direct-response tactics increase both web traffic and purchase probability. Information-based arguments and emotional content actually reduce traffic but increase sales among those who visit the brand’s website. Imagery content reduces direct traffic but does not affect purchase probability. These results imply that brands seeking to attract multitaskers’ attention and dollars must select their advertising copy carefully according to their objectives.

Keywords: Advertising content, difference-in-differences, internet, media multitasking, online purchases, simultaneous equations model, quasi-experimental design, television.

1 Dake Family Assistant Professor, Cornell University, Dyson School of Applied Economics and Management, https://faculty.cit.cornell.edu/jl2545. 2 Assistant Professor of Business Administration, Harvard Business School, http://www.hbs.edu/ faculty/Pages/profile.aspx?facId=522373. 3 Assistant Professor, University of California, San Diego, Rady School of Management, http://kennethcwilbur.com.

The authors thank comScore, the Cornell University Dyson School Faculty Research Program, Dake Family Endowment, and the Division of Research and Faculty Development of the Harvard Business School for providing the funds to acquire and build the dataset in this research. Teixeira thanks Elizabeth Watkins for research assistantship. Wilbur thanks Duke University for employing him during part of the time this research was conducted. We are grateful to Donald Lichtenstein, Chris Oveis, Catherine Tucker, the editor, area editor, two anonymous referees and numerous seminar audiences for their helpful suggestions. Authors contributed equally. 1. Introduction

As computers have grown smaller and more convenient, simultaneous television and internet consumption (“media multitasking”) has increased rapidly (Lin, Venkataraman and Jap 2013).

Numerous studies have reported large increases in media multitasking; among them, Nielsen

(2010) claimed that 34% of all internet usage time occurred simultaneously with television consumption. Meanwhile, television usage has not fallen, with Americans still watching about five hours per day. In fact, time spent with television and time spent with internet are positively correlated at the household level (Nielsen 2011).

One might therefore suspect that television can effectively engage online shoppers. But do multitaskers engage with television ads or does simultaneous media consumption steal consumer attention away from commercials? Numerous studies suggest that engagement is possible. Among them, Nielsen (2012) found that 27% of US viewers had looked up product information for a TV advertisement, and 22% had looked up advertised coupons or deals advertised on TV. Ofcom (2013) reported that 16% of UK consumers had searched for product information or posted to a social network about a television advertisement.

The current paper studies how the content of television advertising influences online shopping. It aims to contribute to the literature on cross-media effects by answering the following questions: can TV advertising trigger online shopping? If so, how does it work and what type of content is most effective?

Recent research (Zigmond and Stipp 2010, Lewis and Reiley 2013, Joo et al. 2014) has used online search data to show that search engine queries to Google and Yahoo respond almost instantaneously to television commercials. However, to our knowledge, no past research has looked at the effects of television advertising on direct website traffic or online purchase data. 1

This paper not only establishes that online shopping responds to television advertising, it also investigates how those effects depend on advertising content.

To uncover these issues, we merged two large databases of television advertising and internet usage, and then created a third database of advertising content. The ad data represent

$4.1 billion spent by 20 brands in 5 product categories to air 1,269 distinct advertisements

365,017 times in 2010. The contents of these advertisements were coded to assess the extent to which each one incorporated direct response tactics, arguments, emotional content and imagery.

Finally, the advertising data were supplemented with comprehensive, passively measured brand- level website traffic and sales data from a daily sample of 100,000 consumers.

Advertising response studies are notoriously plagued by endogeneity. To address this, we employ a quasi-experimental research design in conjunction with narrow two-minute event windows (Chaney et al. 1991). For each ad insertion, online shopping variables are measured within a narrow window of time prior to the advertisement. This “pre” period serves as a baseline against which the ad’s effect is measured. The same variables are measured again in a

“post” window of the same length immediately following the ad’s insertion. Systematic differences between the pre- and post-windows are attributed to the ad insertion. The identification strategy is similar to the regression discontinuity approach of Hartmann, Nair and

Narayan (2011).

We also measure advertising effects on online shopping in broader two-hour windows of time. In order to partial out unobserved category-time interactions, we use online shopping on nonadvertising competitors’ websites as control variables in a difference-in-differences regression framework.

2

We find clear evidence that television advertising influences online shopping. Direct response content increases direct website visitation (e.g., directly using a URL) with a smaller corresponding decrease in search engine referrals (e.g., indirectly via a search engine). It also raises conversion probability. Arguments and emotional content reduce direct traffic while simultaneously increasing purchase probabilities; the net result of these two offsetting effects is positive for most brands. Imagery content reduces direct traffic and does not significantly change purchase probabilities. In sum, the results suggest that advertisers must select advertising content carefully according to their objectives.

The paper proceeds by reviewing literature on TV advertising and proposing a simple conceptual framework. It then describes the data, model specification and the results. A general discussion of the implications for television advertisers concludes.

2. Background Literature and Conceptual Framework

Our work is directly related to research on multimedia advertising effectiveness. Several recent studies found evidence of synergistic effects between television advertising and internet advertising on offline sales (Kolsarici and Vakratsas 2011, Naik and Peters 2009, Naik and

Raman 2003, Ohnishi and Manchanda 2012). Dagger and Danaher (2013) built a single-source, customer-level database of ten advertising media and retail sales for a large retailer. They found that single-medium advertising elasticities were highest for catalogs, followed by direct mail, television, email and search, suggesting that direct-response channels are more effective at increasing short-term sales than other advertising channels.

.

3

The sum of the evidence suggests that significant cross-media effects exist. However, researchers are just starting to understand how the content of advertising in one medium might influence consumers’ behavior in another. In an early effort, Godes and Mayzlin (2004) showed that online discussions of new television programs had explanatory power in a dynamic model of those program’s ratings. More recently, Gong et al. (2013) designed a field experiment to measure the causal impact of tweets and retweets on ratings of a television program. They found that the content of promotional messages on the internet influenced the number of people estimated to view the promoted television program.

2.1 TV Advertising and Online Behavior

Television ads are valuable for generating awareness, knowledge and interest in new products. A direct consequence is that effective television ads may lead viewers to seek out more information about these products and brands (Rubinson 2009). To date, the most studied online behavior among TV viewers has revolved around searching for advertised brands and products using search engines (e.g., Joo et al. 2014).

Lewis and Reiley (2013) found that advertisements during the Super Bowl tend to trigger online searches for the advertised brands immediately, within one minute, with smaller effects noticeable up to an hour after the ad’s broadcast time. However, their analysis did not include direct traffic to the brand’s website or online purchases, making it impossible to separate interest in the ad’s entertainment value from interest in the advertised product. They suggested that

“other user data such as site visitation and purchase behavior could provide a more holistic perspective…” This paper follows up on this suggestion.

Following this observation, we posit in Figure 1 that consumers have two major decisions in response to TV ad exposure. First, they choose whether to visit the brand’s website or not. If 4

the brand’s website is very salient, this action may be achieved by a direct route, such as entering the website address directly into the browser or clicking a bookmark. If the brand website is unknown or not salient, the consumer might instead need to visit a search engine and then click a referring link to the brand’s website. Second, upon arrival at the website, the consumer decides whether to purchase or not.

[Figure 1 about here]

Zigmond and Stipp (2010, 2011) offered several case studies showing that large increases in Google searches for branded keywords corresponded to the precise timing of brands’ TV ads aired during the Olympics. They speculated that heterogeneity in search response to TV ads was partly due to the brand and partly due to the ad content. They reasoned that new-product ads should generate more online search while call-to-action ads should generate fewer searches and more direct website visits. Therefore, we allow for differential effects of TV ad content on consumers’ two major routes of visitation to the brand website. Next, we review the literature on television advertising content.

2.2 Typology of TV Ad Content

Prior research has claimed that advertising content is a first-order determinant of advertising effects. For example, Wind and Sharp (2009) said that “the most dramatic influence on short- term effect is creative copy.”

Tellis (2004) summarized the advertising literature by explaining that advertising effects can be classified as either behavioral or attitudinal. Behavioral effects act instantaneously, at the moment of exposure, or shortly thereafter. Attitudinal effects operate by changing the consumer’s attitudes and memory over a longer period of time. Using this simple dichotomy, prior research has categorized ads into those that predominantly seek a behavioral response and 5

those that predominantly seek to influence attitudes. An ad need not focus on just one purpose; many TV ads exhibit some elements of both types. However, in practice they are negatively correlated as ad time is expensive and different tactics are used to reach these two goals.1 Ads that primarily seek to elicit behavioral responses are normally called “direct-response” (Danaher and Green 1997) while those that intend to cause attitudinal changes are often called “brand- image” (Peltier et al. 1992).

Direct-response ads possess three characteristic elements. In order to elicit a behavioral response from the viewer, they provide (i) a solicitation of a specific action(s), (ii) supporting information to encourage a decision, and (iii) a response device or mechanism to facilitate action

(Danaher and Green 1997, Bush and Bush 1990). About 20% of the TV ads in the U.S. are estimated to be primarily direct-response (Danaher and Green 1997, Peltier et al. 1992). The literature has shown that these ‘gimmicks’ are indeed effective at eliciting immediate responses.

On the other hand, brand-image ads are used to reinforce or change attitudes regarding how consumers perceive the brand. They do so by appealing to two processing mechanisms, the cognitive and the affective system. Brand-image ads constitute about 75% of all TV ads in the

U.S. (James and Vanden Bergh 1990).

Ads that involve cognitive (or central route of) persuasion do so through the use of arguments. These argument-based ads persuade by appealing to reason and relying on evidence about the product, the price and brand information whereby viewers evaluate the merits of the proposed arguments against their counterarguments (Petty and Cacioppo 1986, Tellis 2004).

1 A related literature uses similar typology and focuses on the trade-off between informative and persuasive roles of advertising (e.g., Ackerberg 2001, Anderson et al. 2013, Bagwell 2007, Ching and Ishihara 2012). 6

Ads that appeal to the affective (or peripheral) system attempt to persuade customers of the brand’s value either through the use of emotionally engaging content (Gross and Thomson

2007, Hajcak and Olvet 2008) or through visual imagery. We term the former emotion-based ads as they attract attention and engage viewers by using emotion-inducing content such as creative stories, warmth and humor tactics (Teixeira et al. 2012, Tellis 2004). On the other hand, the use of multiple perceptual or sensory representations of ideas (predominantly visual) is intended to excite the senses using sensory stimuli, concrete words, and vivid pictures. This approach, in turn, evokes visual imagery processing in consumers and incites a process of memorization, intent formation, or affect (MacInnis and Price 1987, Peltier et al. 1992). We refer to this as imagery-based ad content.

[Figure 2 about here]

While all the tactics may be used within the same advertisement, constraints (e.g., air time or production budget) typically require advertisers to focus primarily on one technique.

Figure 2 summarizes the four types of TV advertisements. Next, we use this classification to develop expectations about how ad content affects online shopping.

2.3 TV Ad Content and Online Shopping

We expect the effect of TV ads on online shopping to be driven by media multitasking, an activity in which consumers divide attention between the television set and a secondary screen, the computer.2 Therefore, we expect the level of attention needed to process each type of advertising content to influence how that content affects online shopping behavior (Teixeira

2 ComScore only measured internet usage only on desktops and laptops in 2010; at that time it had not yet developed tracking technology for tablets or smartphones. 7

2014). Because there is no extensive literature on which to base formal hypotheses, we only provide informal conjectures.

In thinking through the possible influence of TV ads on online shopping, it is necessary to consider the role the brand’s website might play. Broadly speaking, the brand website can serve two roles: it could be a channel for selling (e.g. providing product information and further persuasion), or it could be a channel for order fulfillment. Ad content that stimulates interest without providing much information would be more effective in conjunction with a brand website that is a channel for selling. Ad content that provides extensive information would be more effective in conjunction with a website that is a channel for order fulfillment. For example,

Hans et al. (2013) showed that some claims in text search ads are more effective for generating click-through (promoting the site) while others generate less traffic but are better at increasing conversions (persuading to buy). Similarly, Wu et al. (2005) found that some magazine ad formats were more effective at generating site traffic while other formats brought less traffic to the site but that traffic converted at higher rates. Anderson and Renault (2006) formally modeled this trade-off; in equilibrium, a rational consumer’s willingness to incur a search cost (e.g. visit a website) is greater when the firm provides partial information about product attributes and price than when it provides full information.

Although we do not observe brand website content in the dataset described below, these thoughts about the role of the website helped to shape our expectations about how TV ad content might influence online shopping. Similar to magazine and search engine ads, TV commercials may attempt to persuade viewers to visit a brand’s website or to make a purchase online. While both approaches might result in a purchase, it is important to distinguish the ad’s ability to

8

generate traffic from the ad’s ability to generate sales. Next, we relate the four types of TV ad content to their expected impacts on website visitation.

We expect direct-response ads to increase website visitations because, by their nature, they are created to cause consumers to act immediately. Immediate action lends itself well to media multitasking as these ads actively encourage their viewers to make use of another medium to respond (e.g., “call now,” “go online,” “visit us,” etc.). We also expect this time of advertising to make the web address more salient in consumers’ minds, leading to a greater impact on direct traffic than on search engine referrals. Argument-based ads, on the other hand, make use of content that requires heightened attention and cognitive processing. For this reason, some viewers might not be motivated to exert the necessary effort to process the arguments in the ad and this will reduce the likelihood that media multitaskers actively seek additional information from the advertiser on the Internet by directly visiting the website or via search engines. This is not to say that argument ads do not trigger interest. Rather, we expect that the desire to act quickly is much less than from direct-response ads, which induce an impulsive act (Doyle and

Saunders 1990, Wood 2009).

As for emotion-based ads, they do not require an intense cognitive processing. We expect emotional ads to increase both routes to website visitations by media multitaskers as they do not require heightened attention to process the message. Further, by changing attitudes, emotions act as a trigger for action (Gross and Thomson 2007). Lastly, while imagery-based ads also generate affect, they can reduce people’s desire to go online as the sensory stimulation is likely to keep viewers’ attention focused on the TV screen and less on other competing media. Thus, by evoking strong visuals and sensory stimulation, viewers may feel less compelled to switch from

9

television, a stimulating and fast-paced medium, to the internet, a slower and self-paced medium

(Berlyne 1971). Next, we conjecture the impact of the four types of TV ad content on purchase.

We expect that, by focusing on consumer actions, direct-response ads that use the web as a fulfilment channel will increase online purchases above and beyond that which results from more website visits. Argument-based ads are expected to increase online purchases as well, but because they focus on the product and brand. As for the affective-laden ads, emotion-based ads should also increase purchases as they provide peripheral cues that entertain and persuade viewers to evaluate the brand favorably (Teixeira et al. 2013). Contrary to the other ads however, we expect that imagery-focus ads will reduce the viewer’s likelihood of purchasing online in the short run as imagery offers a positive sensory experience that acts as a palliative substitute for actual product consumption, delaying purchase (MacInnis and Price 1987). In the next section we describe the data, sample selection and key measures used in the empirical model.

3. Data

We merge two large datasets of television advertising and internet behavior in 2010 and construct a database of advertising content. Given the huge databases involved, the analysis focuses on 20 brands in five product categories with extensive online shopping activity: dating, pizza delivery, retailers, telecommunications, and travel.

3.1. Web Traffic and Transactions Data

Online traffic and transactions data were collected from comScore Media Metrix. ComScore used proprietary software to passively track all web usage on a sample of two million internet- connected desktops and laptops. The data contained information about the Uniform Resource

Locators (URL), date and time of each web page visited. Due to the substantial costs of data

10

retrieval, comScore randomly selected 100,000 machines each day and only retrieved internet usage data from these machines (Coffey 2001).

ComScore reports the web browsing data at the level of the user/website session.

Consistent with standard industry practice, a new session is recorded when a user first initiates a page view from a particular domain (e.g., Amazon.com) after not viewing any page from that domain in the past 30 minutes. The choice of 30 minutes is commonly made because many users stop looking at webpages without closing a browser tab, so some assumption is required about the point at which the user stopped interacting with the site.

For each user/website session, comScore reported an anonymous user ID, the domain name (brand website), domain name of a referral website (if any), and the exact date and start time. Further, comScore identified paid transactions by analyzing the structure of confirmatory

URLs for all but a few brands it tracked.3

The internet usage database has several limitations that are important for interpreting the results below. First, the data are a daily cross-section drawn from a panel, but not a panel in and of themselves. Therefore, we analyze the data by aggregating users’ session data within specific windows of time; we refer to traffic and transactions as the aggregate counterparts to individual visitation/browsing and purchasing decisions. Second, the data do not track individuals across computers (a common issue in internet usage data). Third, at the time the data were collected, comScore only measured internet usage on desktops and laptops; it had not yet developed tracking technology for tablet computers. In 2010, smartphone penetration was 22% and major brands of tablet computers had just come on the market; both devices were generally less

3 Prior research in marketing has analyzed comScore data from 2002-2004 (e.g., Moe and Fader 2004, Park and Fader 2004, Montgomery, Li, Srinivasan, and Liechty 2004, Danaher 2007, Johnson et al. 2004). 11

suitable for online shopping than desktops and laptops (Nielsen 2010). By 2014, smartphones and tablets had become more capable and their respective penetration rates had risen to 65% and

29% (Nielsen 2014). One might suspect that the effects estimated in this paper are a conservative estimate of the current importance of online response to television ads.

Figure 3 summarizes the online shopping data by plotting traffic and transactions within each product category by hour of the day. In four of five categories, brand website traffic and transactions are surprisingly flat throughout the day, from about 9 A.M. until 9 P.M., with a peak in the early evening at 7 P.M. Eastern Time. The exception is pizza, which has a more pronounced peak in online shopping at dinnertime.

[Figure 3 about here]

3.2. Television Advertising Data

Television advertising data were recorded by Kantar Media. Kantar continuously monitored all national broadcast and cable networks in the U.S and identified advertisements using codes embedded in networks’ programming streams. Each unique combination of a commercial message, television channel, date and time is referred to as an advertising “insertion.” For each insertion in 2010, the database reports the commercial’s duration, the brand, the date and start time (in hours, minutes, and seconds E.S.T.), and an estimated cost of the insertion. Cost estimates were reported to Kantar by the networks after ads aired and are commonly used by large advertisers to plan upcoming media buys. The data also record the specific product advertised within the advertisement, as many brands advertised multiple different products.

Finally, the database report several properties of the program into which the ad was inserted: the

“property” (defined as a national television network or program syndication company), program

12

name, program genre, the number of the slot during the commercial break when the appeared, and the number of the commercial break within the program.4

The data included more than 750,000 insertions of 4,153 unique advertising creatives in national networks. We dropped the bottom 5% of creatives by total expenditure, and all insertions whose estimated cost to broadcast was less than $1,000, as these corresponded to channels and dayparts with very small audiences. These two refinements reduced the number of insertions by about half but eliminated just 6% of total observed ad spending. The final estimation sample consists of 365,017 insertions of 1,269 unique advertisements accounting for

$4.1 billion of TV ad spending by 20 brands in 2010.

Like the online shopping activity in Figure 3, Figure 4 shows numerous advertising insertions occurred between about 9 A.M. and 7 P.M. The number of ad insertions dropped but advertising expenditures rose considerably during the prime time hours of 8-10 P.M.

[Figure 4 about here]

3.3. Research Design, Model-Free Evidence and Descriptive Statistics

We measure brand-specific shopping variables twice for each ad insertion and each window length. The baseline rates of online shopping variables are measured in a “pre” window of time just prior to the insertion of the advertisement. These same variables are measured again in a

“post” window of time just after the ad starts. Any systematic differences between the online shopping variables measured in the “pre” and “post” windows will be attributed to the advertisement itself.

4 The database did not report program name, genre, break number or slot number for 36,805 ad insertions carried by a particular group of program syndication companies. We decided to drop these 10% of insertions from the sample. The results of primary interest (tables 5 and 6) are essentially invariant to including or excluding these insertions. 13

The online shopping variables of interest are brand website traffic, either direct or via search engines, and transactions. They are defined as follows.

Direct Traffic (DIR): the number of new sessions on a brand’s website that were initiated by direct means (e.g., URL entry or clicking a bookmarked link) within a particular time window.

Search Engine Referrals (SE): the number of new user sessions on a brand’s website that were initiated by search engine referrals within a particular time window. Six search engines (AOL,

Ask, Bing, Google, MSN and Yahoo) are included, accounting for 99% of U.S. searches.5

Transaction Count (TC): the number of new sessions on a brand’s website that are initiated within a particular time window and where a transaction is completed within 24 hours. Purchase decisions may take much longer than site visits, as they may be delayed by time spent reading reviews, researching competing options or consulting other members of the household. Thus, a one-day window was employed similarly to Blake et al. (2013).

It is important to note that the difference between sessions and pageviews (described earlier) ensures that the same machine will not be counted in both the pre and post windows in the two-minute data. If a given machine initiates a new session during the two-minute “pre” window, comScore’s definition of a session ensures it will not be counted again in the two- minute “post” window, as 30 minutes have not elapsed between pageviews.6

Several exploratory analyses were conducted using subsets of the data. In one, we plotted traffic to brand websites corresponding to different ad creatives. Figure 6 shows Amazon.com traffic for two distinct ads: (a) “available now” and (b) “Kindle.” The data showed a large spike

5 A limitation of this the data is that this measure does not indicate when the user initiated the search. 6 It is possible but highly unusual for a single machine to be counted in both the “pre” and “post” windows in the two-hour dataset. If a machine’s last visit to a brand webpage is more than 30 minutes prior to an ad insertion, and then the machine is observed to visit the brand’s webpage again during the two hours following the ad insertion, those will be counted as one new session during the two-hour “pre” window and one new session during the two- hour “post” window. 14

in the minute following the start of the ad and a small, enduring increase thereafter. The magnitude of these lift patterns seemed to depend on the ad content, highlighting the importance of more formal investigation of the impact of ad content on web visitations.

The second exploratory exercise involved plotting browsing activity within shorter time windows for a wider selection of brands. Figure 7 illustrates this for Target and JC Penney’s.

Most of the immediate uptick in browsing activity was observed within two minutes after the ad, with some effects persisting up to two hours after the ad. A similar pattern appeared for all of the brands analyzed in this manner. This is how we chose the two particular window lengths of two minutes and two hours.7 The online appendix offers a concrete example of how the online shopping variables are measured within each of these two windows.

[Figures 6 and 7 about here]

Table 1 provides advertising and online shopping data for the 20 brands in the dataset.

The average brand created 64 different commercials to advertise 7 distinct products, and spent

$204 million to air those commercials 18,251 times. Consumers initiated 49,402 direct sessions on the average brand’s website, with an additional 23,061 sessions coming from search engine referrals. 6.3% of the machines that were observed to initiate those sessions completed a paid transaction or subscription within 24 hours. Table 2 offers some back-of-the-envelope

7 Although an ideal approach would be to gauge the sensitivity of the analysis to the length of the window chosen, this was judged to be infeasible due to computational costs. This was an unusually complex data merge; to our knowledge, it has not previously been offered by any commercial research firm. Due to the sheer size of the datasets, our merge routine required 3*1013 computational queries and about 45 days to run. Section 5 indicates some agreement in the results based on the two chosen window lengths, suggesting that small adjustments the window lengths might be unlikely to change the qualitative findings. 15

calculations showing that, under conservative assumptions, 13.5% of the online purchases in the data may have been a direct result of in-sample advertisements.8

3.4. Television Advertising Content Data

The third dataset was created specifically for this paper by coding the contents of the 1,269 TV advertisements. Most prior academic efforts to analyze advertising content have manually coded a few dozen ad creatives9. Our data collection effort contained 1,269 unique ad creatives, 21 ad content items per creative and spanned multiple brands and categories. Given the size of the task at hand, we opted for a three-step procedure involving item coding, assessment of reliability and classification validation.

We first used the literature to identify and define the four ad types (Direct-response,

Argument-based, Emotion-based and Imagery-based) by which each TV commercial in our dataset to be classified. This ad typology is defined and presented in Section 2.2 and summarized in Figure 2.

Using these definitions, we selected 21 ad content elements to code from prior academic analyses of advertising content. All ads were coded on the basis of these items. Given the large number of ads to be coded, we recruited ten coders and assigned each ad to only one coder who viewed it multiple times and coded it on the basis of the items chosen. A subsample of ads was later re-coded by a new group of six independent coders following the same procedure to measure inter-coder reliability. Finally, we submitted the proposed classification along with the

8 We advise strong caution in interpreting these calculations as they rely on several untested assumptions and they are intended solely as an illustration. However, they do suggest that a significant number of the observed transactions were caused by TV ads in the sample. 9 Unusually large exceptions are Buijzen and Valkenburg (2004), who identified the presence of 41 types of humor in 316 advertisements, and Anderson et al. (2013) and Liaukonyte (2013), who coded the product attributes communicated by 1,571 OTC pain medication ads.. 16

items pertaining to each ad type to an expert panel of 14 academics to validate or refute the groupings. The details of each step follow.

Item selection. 21 question items were created to measure the prevalence of the four types of content features in each ad. The features were chosen to identify direct-response elements

(e.g., call to go online, online contact information, call to purchase), arguments (e.g., product- related, price-related, brand-related), emotion-inducing elements (e.g., story, humor, warm feeling content) and sensory elements (e.g., visually pleasing, sensory stimulation).10

Feature coding. Ten research assistants were trained to code the advertisements. Coders were instructed to watch each ad at least twice and then answer the 21-item questionnaire for that ad. During coding, they could watch, pause and rewind the ad as many times as needed. If they still remained unsure about how to code a particular ad, they were instructed to inform a research associate. Over 99% of ads were coded completely the first time. Coders worked independently, were paid hourly, and instructed not to work more than two hours at a time in order to avoid respondent fatigue.

Coding reliability. A separate group of six assistants were hired to code a random sample of 150 ads for eight of the brands (12% of the original 1,269) following the same procedure. We dropped two survey items (“Is the product demonstrated in the ad?” and “Is the focus of the ad more on the product or on the brand?”) due to low inter-coder reliability. The percentage match among the remaining 19 survey items was 78%. We judged this figure to be acceptable given the subjective nature of some of the survey items and the coders’ inability to resolve discrepancies through discussion.

10 19 of the 21 survey items used binary response scales (presence/absence of element). For two items a three-point scale (predominance of one element, predominance of another element or neither) was used. 17

Classification validation. In order to validate the choice of items used for each ad type, we surveyed 14 academics from top-tier schools around the world who are experts on consumer behavior research. We asked whether each item was applicable, somewhat applicable or not applicable to the ad type that it was associated with. Only one of the items (“Would you judge this to be an expensive or cheap ad to make?”) had a high rate of disagreement with the original classification, at 50%, and was therefore dropped from the study. On average, the academics surveyed agreed with the applicability in 97% of the remaining item/grouping combinations, with every item-specific agreement score exceeding 85%. In the end, 18 survey items were used to create indices based on the sum of each advertisement’s item responses within each group.

The survey items by ad type are provided in Table 3.

[Table 3 about here]

Descriptive Statistics. Table 4 describes how brands differ in their use of advertising content. For example, Papa John’s made the heaviest use of direct-response ads in the sample, while

Victoria’s Secret ads rated the lowest on this type. However, while there are differences across brands, standard deviations across creatives within a brand are sometimes comparable to the standard deviations across the entire sample. In sum, every brand used every type of ad content in its advertisements.

[Table 4 about here]

4. Model and Estimation

We model the causal effects of TV advertisements on three online shopping variables—search engine referrals (SE), direct traffic (Dir) and transaction counts (TC)—using a system of linear equations. Let i index advertisement insertions. Each insertion i promotes a particular brand and 18

product in a particular product category and corresponds to a particular date and time; we denote

these bi , pi , ci and ti , respectively.

POST POST POST POST Let Yi  (SEi ,Diri ,TCi ) be a vector of the three online shopping variables for brand measured within a window of time (either two minutes or two hours) immediately

PRE following . Let Yi denote a vector of the same three variables for brand measured in a window of time (of the same duration) immediately preceding insertion i.

In explaining our approach to estimating causal effects, we adapt the notation of Angrist and Pischke (2009). We distinguish the online shopping variables we observe after an ad

POST insertion, denoted Y1i , from the same online shopping variables we would have observed in

POST the same “post” time window had insertion i never occurred, denoted Y0i . In other words,

is the baseline level of online shopping while is this baseline plus the treatment effect of the television advertisement.

In the absence of an ad insertion, then we would expect the online shopping variables

(Y POST ) to be influenced by their past realizations (Y PRE ), as well as brand effects (  ), 0i i bi observed category-time interactions ( X ), and possibly unobserved category-time interactions ( citi

 ). We make the conventional assumption that the conditional expectation of Y POST is a linear citi 0i function of these variables:

Y POST  Y PRE Y    X Y   U Y (1) 0i i bi bi citi citi i where Y and Y are matrices of parameters to be estimated and U Y is a random deviation of bi i

POST the realization of Y0i from its conditional expectation. We supplement the model in equation 19

(1) to allow the traffic variables (SE and Dir) to directly impact sales (TC), as this relationship is directly implied by the definition of TC:

Y POST   Y PRE Y    X Y   U Y (2) 0i bi i bi bi citi citi i wherein  is a 3x1 vector with zeroes in its first two rows and SE POST ST  Dir POST DT in its bi i bi i bi third row, and  ST and  DT capture the likelihood that search-engine referral or direct visitation bi bi

to brand bi ’s website is observed to result in a purchase within 24 hours.

We use i to refer to the treatment effect of advertisement insertion i and assume it also enters linearly:

E[Y POST | Y PRE ,  , , X , ]    E[Y POST | Y PRE ,  , , X , ]. (3) 1i i bi ti citi citi i 0i i bi ti citi citi

Therefore, the conditional expectation of the shopping variables in the “post” window is

Y POST      Y PRE AY    X Y   U Y . (4) i i bi i bi bi citi citi i

In the following subsections, we describe each term in equation (4) in full detail. Before doing so, we focus on the assumptions required for consistent estimation of the parameters using data from each of the two time windows (two minutes and two hours).

4.1. Endogeneity Concerns

In order to obtain consistent estimates of the parameters without further modification of the model, it must be the case that the unobserved category/time interactions (  ) are orthogonal to citi

the treatment effect ( i ). The possibility that they may be correlated arises from the idea that

20

brands may be able to anticipate unobserved category/time interactions and purchase ad insertions at particular times to profit from those unobserved category/time interactions.11

Fortunately, when we use online shopping data measured in two-minute windows of time, typical advertising business practices ensure that this is not a concern. The reason is that, when a brand buys a television commercial, it pays for a network/quarter-hour combination, for example ESPN between 8:45:00-8:59:59 P.M. on January 1, 2010. The advertiser does not know the specific time within the quarter-hour that the ad will air, for three reasons. First, the actual timing of commercial break within that quarter hour is not specified in the contract between the network and the advertiser. Second, unless the advertiser has taken the unusual step of purchasing a specific slot within the break, it does not know which slot its advertisement will occupy (Wilbur et al. 2013). Third, 80% of advertisements are sold during the May “upfront” market, 3-15 months prior to the ads’ air dates. Advertisers and networks often do not even know what programs the ads will be on the air, much less the specific times that the ads themselves will run.

For all three reasons, it strains credibility to argue that an advertiser could time a specific ad insertion to profit from changes in online shopping behavior between a two-minute “pre” window and an immediately following two-minute “post” window. Any systematic differences in online shopping variables between the two-minute pre and post windows should be directly attributable to the treatment effect, because it is not possible for a brand to time an ad insertion within any four-minute block of time. In other words,  and  are uncorrelated in the two- biti i

11 For example, suppose that many people subscribe to online dating sites on Monday afternoons with the hope of finding dates during the following weekend. If a dating website was aware of this behavior, then the website might tend to air particular types of ads on Mondays more than at other times. It then would be difficult to distinguish dating-Monday-afternoon effect from a treatment effect for a dating website on a Monday afternoon. 21

minute window data. Equation (4) can be estimated directly using Two-Stage Least Squares

(2SLS) using the online shopping variables measured in two-minute intervals.

However, when using online shopping data measured in two-hour windows of time, standard industry practices provide no such assurances; one might suspect that advertisers choose treatment effects  with some knowledge of unobserved category-time trends  . In this case, i citi we use a difference-in-difference approach to control for unobserved category trends that vary over time.

The idea is to use changes in online shopping variables for brands that (i) are in the same

product category as brand bi , and (ii) did not advertise during the sample period, to control for unobserved category-time interactions corresponding to each insertion in the sample. Let

Z POST  (SE POST ,Dir POST ,TCPOST ) be a vector of the three online shopping variables for a set of ci ci ci ci

control brands in category ci measured in a two-hour window of time following ad insertion i, and let Z PRE be its counterpart in the two-hour period preceding ad insertion i. We denote the set ci

of control brands corresponding to product category ci as zi and discuss the selection of those control brands below.

Because the control brands did not advertise at all, including at time t i , and therefore did not experience any treatment effect, the conditional expectation of Z POST takes the same form as ci

POST Y0i :

Z POST  K  Z PRE AZ    X Z   U Z . (6) i zi i bi zi citi citi i

22

where the first two elements of K are zero and the third is SE POST ST  Dir POST DT . To zi zi zi zi zi finalize the difference-in-difference regression model, we subtract equation (6) from the

POST conditional expectation of Y1i in equation (4) so that the unobserved category-time interactions (  ) drop out of the model: citi

Y POST  Z POST    K  (Y PRE AY  Z PRE AZ )    X B  U (7) i i i i i bi i zi i citi i where K  K  K ,      , B  BY  BZ and U  U Y U Z . The system of i bi zi i bi zi i i i equations described by (7) is estimated via 2SLS using the two-hour window data.

We chose the control brands by selecting the largest brands within each product category that did not advertise on television. That resulted in the following sets of control brands: (1)

Dating: OKCupid.com, Plentyoffish.com; (2) Retailers: Abercrombie, Roamans, American

Eagle, Children’s Place; (3) Telecom: Letstalk, Cricket, Wirefly; (4) Travel: Airtran,

ChoiceHotels, Cheaptickets, JetBlue. For the fifth product category (pizza), we were not able to find any non-advertising brands that seemed to offer services comparable to pizza delivery.

Therefore, the difference-in-difference regressions exclude advertising insertions for pizza.

4.2. Exogenous Variables

The model contains two types of explanatory variables: baseline (in  and X ) and treatment bi citi

(in i ). The baseline variables include brand fixed effects and a rich set of interactions between category dummies and time variables. For each of four intervals of time—week of the year, day of the week, hour of the day, and minute of the hour—we interact a fixed effect for each time interval with a product category dummy. Therefore each online shopping variable equation in conditional expectation (2) contains 714 categorical variables: an intercept, 19 brand effects, 259

23

category-week interactions, 34 category-weekday interactions, 119 category-hour interactions, and 299 category-minute interactions. The purpose of including such a flexible conditional expectation function is to avoid mistakenly attributing systematic variation in the online shopping variables to advertising insertions.

The treatment effect i includes the advertising content variables described in section

3.4, and an additional 266 fixed effects for nearly everything we are able to observe about the advertising insertion: the product that was advertised (137); the property (national network or syndication company) in whose program the ad was inserted (96); the program genre during which the ad was inserted (15); the number of the commercial break during which the ad was inserted (9); and the number of the slot during the break during which the ad ran (9).

Moreover, the treatment effect contained additional controls: the estimated expenditure on the advertisement, interacted with a brand dummy; the sum of all prior observed expenditures

on advertising creative ci (to control for possible ad wearout); the total expenditure by the brand on other insertions during insertion i’s PRE and POST windows; and within-category competitors’ total ad expenditures during insertion i’s PRE and POST windows. These final four variables are included to control for clustering, that is, occurrences of neighboring insertions.12

Investigations showed that the results of primary interest (in Tables 5 and 6) are qualitatively unaffected by the inclusion or exclusion of these controls.

4.3. Total Effects on Transactions

A quantity of particular interest is the total effect of advertising content on transactions. The model allows each advertising content element j to affect sales (TC) of brand b directly and

12 To illustrate the clustering concern, suppose AT&T inserts an advertisement on CBS at 8:41:00 P.M. and another ad on ESPN at 8:42:00 P.M. In this case, the POST window of the CBS ad will include some traffic caused by the ESPN ad and both the PRE and POST windows of the ESPN ad will include some traffic caused by the CBS ad. 24

indirectly through either route of visitation (SE and Dir). Therefore, the total effect of ad content element j on brand b transactions is

SE ST Dir DT TC TotalEffectbj  j b  j b  j . (5)

The right-hand side of equation (5) sums the direct effect of ad content element j on search engine referral traffic, multiplied by the brand-specific effect of search engine referrals on transactions; the direct effect of ad content element j on direct traffic, multiplied by the brand- specific effect of direct traffic on transactions; and the direct effect of ad content element j on transactions. Standard errors of these total effects are calculated by bootstrapping from the asymptotic distribution of the parameter estimates. The next section presents the econometric findings.

5. Results

The first question to consider is whether advertising influences online shopping. Table 5 reports the proportion of the variation explained in the three online shopping variables by several alternative models.

[Table 5 about here]

Several conclusions emerge. First, in all cases, the model shows a greater ability to explain direct traffic than search engine referrals, perhaps because of the time required to conduct an internet search and evaluate the results. Second, if we include only data about the advertising insertion treatment effect (excluding all baseline variables), the model can explain

48.7% and 62.2% of the variation in search engine referrals and direct traffic, respectively, in the two-hour window data, and 3.5% and 13.6% of those variables in the two-minute window data.

Third, the baseline model (excluding treatment effect variables) explains more of the variation in 25

the dependent variables than the treatment effect alone. When we add the treatment effect to the baseline variables, the model does indeed show an increased ability to explain all three dependent variables in both the two-minute and two-hour datasets, thereby answering the first question (whether TV advertising influences online shopping) in the affirmative.

Finally, we estimated a model with category-specific treatment effects. This was done by interacting the treatment effect variables with category dummies, thereby increasing the number of treatment effect parameters by factors of 5. However, the R-square statistics showed no meaningful increase when we included the category-specific treatment effects. Evaluating all models, the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) are both minimized by the model with the common treatment effect, suggesting that the richer category-specific parameterization does not justify the increase in model complexity.13

Therefore, we proceed by presenting and interpreting the findings from the common treatment effects model, starting with advertising content.

5.1. Effects of Television Advertising Content on Online Shopping

Table 6 presents the effects of TV advertising content on direct traffic, search engine referral traffic, and transactions. There are twelve such effects—four advertising content variables times three online shopping variables—within each of the two regressions. In the two-hour regression,

9 of the 12 parameters are statistically significant at the 95% confidence level, while 6 of the 12 are significant in the two-minute regression. Four effects correspond in sign and significance level between the two regressions, and there are no cases of contradictory findings between the two-minute and two-hour window regressions, indicating some convergent validity. We proceed

13 The model was also subjected to a random 80% hold-out validation exercise. This was done to check for overfitting because of the large number of fixed effects included. The results indicated a high degree of reliability. The R-square and RMSE statistics were comparable between the full sample, a model estimated with a random 80% subsample, and the predictions from that latter model when compared to the remaining 20% validation subsample. 26

by discussing the two-hour results, as summarized in Figure 8, as the slightly longer time window includes more consumer search and purchase activity.

[Table 6 and Figure 8 about here]

Direct-response. Ads that make heavy use of direct-response tactics are found to have three effects. First, they reduce the number of new sessions at the brand website initiated by search engine referrals. Second, as expected, they increase the number of visitors coming through direct means. The positive effect on direct visitation is more than six times larger than the drop in search engine referrals. Taken together, these two results suggest that direct-response tactics both bring new visitors to the site and encourage direct means of visitation over indirect means

(similar to Joo et al. 2014). This is likely a positive consequence for the brand, as it suggests that television advertising makes the brand more salient to consumers, helping them to bypass search engines and thereby reducing the “toll” the brand has to pay for search engine referral traffic.

Third, direct-response content is found not only to increase traffic overall, it leads to a higher purchase probability among those who visit the website. These effects combine to create a positive, significant total effect of direct response advertising on purchases for all 17 brands.

Argument and Emotion. Arguments and emotion-based content in ads are associated with two seemingly contradictory effects: they simultaneously reduce traffic to the website while increasing purchase probability among those who do visit. The most likely explanation for these phenomena is that this type of advertising content is effective at resolving consumer uncertainty about whether the advertised product fits their needs. In such a case, low-fit consumers would forego visiting the brand website, while high-fit consumers would visit and buy, as suggested by

Wu et al. (2005) and Hans et al. (2013). The positive effect on purchase probability outweighs the negative effect on traffic for most brands, leading to positive total effects argument-based 27

content on sales which are statistically significant for most brands. The same is true for emotion- based content.

Imagery. Imagery content is associated with reduced direct visitation to the website in the two- hour dataset. We suspect the reason for this is the effect of imagery on multitasking; intense images are normally used in television advertisements to arrest the viewer’s attention and to discourage them from disengaging with the medium. This suggests a possible downside for brands that would benefit from triggering multitasking behaviors, as this is the only type of ad content to trigger a negative total effect on transactions for most brands.

5.2 Additional Results

Table 7 presents additional predictors of the treatment effect of an advertising insertion.

Advertising insertions that appear within particular genres lead to significantly different conversion rates. Relative to the excluded program genre (Animation), the largest increase in purchase probability were observed for insertions during live sports. However, our ability to interpret this result is limited. More research will be needed to determine whether the effect on purchases comes from program genre itself, which may affect viewer engagement, or whether they are attributable to the particular target audience attracted by those programs.

The break number and slot number results are more easily interpreted. The results indicate that ad breaks that occur later in the program generate fewer new website sessions than the first break in the program, with little or no apparent impact on purchase probability. There does not appear to be any systematic variation in traffic or transactions across slots within commercial breaks within the two-hour regression, though later ad slots produced fewer direct visits in the two-hour regression (two-minute results are omitted for brevity).

28

The data show interesting findings for advertisements that have been aired repeatedly.

More past spending on an advertisement is associated with a reduced ability to generate new direct traffic, but a higher purchase probability among those who do visit. This goes in line with the dual purpose of TV ads for multitaskers: to generate an impulsive action to visit the website and to build up a desire to go buy the brand.

[Tables 7-9 about here]

Table 8 shows that television networks vary in their effects on internet shopping. For example, ads carried on CNBC are associated with reduced visitation across all three shopping variables, while those appearing on Adult Swim or E! show a large increase in new brand website traffic. As was the case for program genre, it is not clear whether these effects are caused by the networks per se or by the types of viewers the networks attract.

Multi-product brands often have distinct creatives for different products. In those cases, the product advertised sometimes has a substantial effect on how the advertisement influences online shopping. The products found to have statistically significant effects are shown in Table 9.

They come mostly from the telecommunications category, wherein brands often market broad lines of related products. It is perhaps unsurprising that these products may vary in their ability to bring consumers to the brand website, as various plans offer substantially different value propositions to consumers.

Table 10 presents autocorrelation and cross-equation correlation parameters. These effects should not be interpreted as causal, but they do describe some interesting patterns. For most brands, website traffic (through either route) in the POST window is positively correlated with traffic counts in the PRE window. Similarly, transactions in the POST window are

29

positively associated with traffic in the POST window for most brands, with larger effects for direct traffic than for search engine referrals.14

[Table 10 about here]

6. Discussion

The debate in the advertising industry has focused on the potentially negative effects of media multitasking: distracting consumer attention from advertisements. In this research, we hope to emphasize a potentially positive aspect: the viewer’s “second screen” enables an immediate and measurable response to television advertising. The question then becomes, how can brands alter their traditional television advertising efforts to influence online shopping?

The purpose of this paper is to estimate how television advertising content affects traffic to the advertiser’s website, immediately or shortly after it appears on TV, and subsequent transactions. This research contributes to the literature on cross-media advertising effects by showing how brands can benefit from increased media multitasking, particularly the consumer habit of simultaneously watching TV and browsing the internet.

The results showed that television advertising influences online shopping. In particular, ads that use direct response tactics are found to increase both direct traffic and conversion probability with a smaller reduction in search engine referrals. Argument- and emotion-based content reduce traffic while simultaneously increasing conversions. Imagery is found to reduce direct visitation without changing purchase probability.

6.1. Implications for Advertising Management

14 The effect of advertising expenditure is included as a control variable, but is 1 difficult to interpret. Recall that the regression includes a large number of category/time interactions, plus dummies for TV property and program type. Due to ad pricing practices, the estimated cost of each insertion is highly correlated with these variables. 30

Managers have to make three major decisions in planning their advertising campaigns: how much to spend, how to spend it (i.e., what media to use) and what to say (i.e., what ad content to use). This research deals primarily with the last question. Although ads may contain countless executional elements, the majority of advertisers use only a handful of broad conceptual categories. We have identified how four such categories influence online shopping.

Perhaps the most striking finding is the fact that advertising content can have opposite effects on traffic and transactions. Internet sales data are typically sparse and highly variable

(Lewis and Rao 2013), so an advertising manager who wants to optimize the online effects of her

TV advertising budget might naturally consider using website traffic as a success metric.

However, our results suggest that such a success metric might lead to precisely wrong conclusions if she is using argument- or emotion-based ad content, as these types of ad content reduce traffic while simultaneously increasing total sales for most brands.

One clear recommendation is that advertisers seeking to influence media multitaskers in the short run should not make heavy use of imagery in their ads. While they may work for consumers of a single medium, these ads have no identifiable cross-media benefits.

These recommendations should be applied with caution as they only apply to the two- hour windows within which we monitored online responses. Also, the total effects on sales was found to vary across brands for each type of ad content. Finally, our data only measure online sales, so our estimates do not distinguish incremental sales from those that may cannibalize traditional channels such as offline retail or telephone contact.

6.3. Caveats and Future Research

As with all research, our analysis is subject to caveats. The first one is that, despite the massive size of our database, we do not directly observe which ads were viewed or attended to by which 31

households. We designed the research to prevent this unobserved factor from influencing our findings; however, because we did not control what TV or online content consumers were exposed to, we avoid attempting to associate TV advertising content and online shopping over time horizons longer than two hours. An especially important question is how brand and brand image advertising will change in an era of media multitasking? Future research will ideally be able to study multi-brand, multi-media single-source panel data of advertising exposure, online traffic and sales, and look for longer-term effects. We hope our findings help to stimulate further interest in creating and sharing such resources.

Another limitation of our approach is that we do not observe brands’ website content or online advertising efforts. While these variables are held constant by our research design, one would naturally expect they influence the probability that a consumer purchases after browsing a brand’s website. A next logical step would be to quantify the effect of website content on sales, and investigate how it might interact with television advertising content. It would also be desirable to collect data on additional TV ad content variables to uncover additional insights.

In conclusion, brand managers have to deal with two effects of media multitasking. On the one hand, it divides consumer attention and diverts it away from advertising. On the other, the rise of online commerce enables a more immediate reaction to traditional advertising. This paper shows how marketers can design their traditional television advertising to influence online shopping, by managing the related goals of maximizing website traffic and transactions. In the long run, we expect that marketers will develop a more sophisticated understanding of how communication efforts in various media impact consumers at different stages of the purchase funnel. We hope our findings can offer some initial progress in that direction by showing how brands can achieve new goals with old methods. 32

References

Ackerberg, D. A. 2001. Empirically Distinguishing Informative and Prestige Effects of Advertising. RAND Journal of Economics, 32 (2) 316-333. Anderson, S.P., F. Ciliberto, J. Liaukonyte. 2013. Information content of advertising: Empirical evidence from the OTC analgesic industry. International Journal of Industrial Organization 31(5) 355-367. Anderson, S.P., R. Renault. 2006. Advertising Content. The American Economic Review, 39(1) 305-326. Angrist, J.D., J.-S. Pischke. 2009. Mostly Harmless Econometrics. Princeton University Press: Oxfordshire, UK. Bagwell, K., 2007. The Economic Analysis of Advertising. In: Armstrong, M. and Porter, R. (eds.), Handbook of Industrial Organization, North Holland, Amsterdam, 3 1701–1844. Berlyne, D.E. 1971. Aesthetics and Psychobiology. East Norwalk, CT: Appleton-Century-Crofts. Blake, T., C. Nosko, S. Tadelis. 2013. Consumer heterogeneity and paid search effectiveness: A large scale field experiment. Mimeo, University of Chicago. Buijzen, M. and Valkenburg, P. M. 2005. Parental mediation of undesired advertising effects. Journal of Broadcasting & Electronic Media 49(2): 153-165. Chaney, Paul K., T. M. Devinney, R.S. Winer. 1991. The impact of new product introductions on the market value of firms. Journal of Business 64(4) 573-610. Ching, A.T.,Ishihara, M. 2012. Measuring the Informative and Persuasive Roles of Detailing on Prescribing Decisions. Management Science, 58(7) 1374-1387 Coffey, S. 2001. Internet Audience Measurement: A Practitioner's View. Journal of Interactive Advertising 1(2) 10-17. Dagger, T., P.J. Danaher. 2013. Comparing the Relative Effectiveness of Advertising Channels: a Case Study of a Multimedia Blitz Campaign. Journal of Marketing Research, 50(4) 517-534. Danaher, P.J., B.J. Green. 1997. A comparison of media factors that influence the effectiveness of direct response television advertising. Journal of Interactive Marketing 11(2) 46-58. Doyle, P., J. Saunders. 1990. Multiproduct advertising budgeting. Marketing Science 9(2) 97- 113. Godes, D., D. Mayzlin. 2004. Using Online Conversations to Study Word-of-Mouth Communication. Marketing Science 23(4) 545-560. Gong, S., J. Zhang, P. Zhao, X. Jiang. 2013. Tweets and Sales: A Field Experiment. Mimeo, Massachusetts Institute of Technology. Gross, J.J., R.A. Thomson. 2007. Emotion Regulation: Conceptual Foundations. Handbook of Emotion Regulation, James J. Gross, ed. New York: Guilford Press 3-26. Haans, H., N. Raassens, R. van Hout. 2013. Search engine advertisements: The impact of advertising statements on click-through and conversion rates. Marketing Letters 24(2) 151-163. Hajcak, G. and Olvet, D. M. 2008. The persistence of attention to emotion: brain potentials during and after picture presentation. Emotion 8(2): 250. Hartmann, W., H. Nair, S. Narayanan. 2011. Identifying Causal Marketing Mix Effects Using a Regression Discontinuity Design. Marketing Science 30(6) 1079-97.

33

James, E. L., B. G. Vanden Bergh. 1990. An information content comparison of magazine ads across a response continuum from direct response to institutional advertising. Journal of Advertising 19(2) 23-29. Johnson, E. J., W. Moe, P. Fader, B. Steven, J. Lohse. 2004. On the depth and dynamics of online search behavior. Management Science, 50(3), 299-308. Joo, M., K.C. Wilbur, B. Cowgill, Y. Zhu. 2014. Television Advertising and Online Search. Management Science, 60(1) 56-73. Kolsarici C., D. Vakratsas. 2011. The complexity of multi-media effects. Marketing Science Institute Working Paper Series 2011 Report No. 11-100. Marketing Science Institute. Lee, D., K. Hosanagar, H. S. Nair. 2013. The effect of advertising content on consumer engagement: Evidence from Facebook. Mimeo, University of Pennsylvania. Lewis, R , J. Rao. 2013. On the near impossibility of measuring returns to advertising. Mimeo, Microsoft Research. http://www.justinmrao.com/lewis_rao_nearimpossibility.pdf. Rossiter, J.R. 1982. Visual imagery: applications to advertising. Advances in Consumer Research 9(1) 101-106. Lewis R., D. Reiley. 2013. Down-to-the-Minute Effects of Super Bowl Advertising on Online Search Behavior. ACM Conference on Electronic Commerce 639-656. Liaukonyte, J. 2014. Is Comparative Advertising an Active Ingredient in the Market for Pain Relief? Working Paper, School of Applied Economics and Management, Cornell University. Lin, C., S. Venkataraman, S.D. Jap. 2013. Media Multiplexing Behavior: Implications for Targeting and Media Planning. Marketing Science 32(2) 310-324. MacInnis, D.J., L.L. Price. 1987. The role of imagery in information processing: Review and extensions. Journal of Consumer Research. 13(4) 473-491. Moe, W. W. and Fader, P. S. 2004. Capturing evolving visit behavior in clickstream data. Journal of Interactive Marketing 18(1): 5-19. Montgomery, A. L., Li, S., Srinivasan, K. and Liechty, J. C. 2004. Modeling online browsing and path analysis using clickstream data. Marketing Science 23(4): 579-595. Naik, P. A., K. Peters. 2009. A Hierarchical Marketing Communications Model of Online and Offline Media Synergies. Journal of Interactive Marketing, 23 288-299. Naik, P. A., K. Raman. 2003. Understanding the Impact of Synergy in Multimedia Communications. Journal of Marketing Research, 13(4) 25-34. Nielsen (2010) Three Screen Report, 1st Quarter 2010. White paper, http://www.nielsen.com/us/en/newswire/2010/what-consumers-watch-nielsens-q1-2010-three- screen-report.html, accessed August 8, 2013. Nielsen (2011) 40% of Table and Smartphone Owners Use Them While Watching TV. http://www. nielsen.com/us/en/newswire/2011/40-of-tablet-and-smartphone-owners-use-them- while-watching-tv.html, accessed August 8, 2013. Nielsen (2012) State of the Media: Advertising & Audiences Part 2. http://nielsen.com/content/dam/corporate/us/en/reports-downloads/2012-Reports/nielsen- advertising-audiences-report-spring-2012.pdf, accessed April 2014. Nielsen (2014) The U.S. Digital Consumer Report. White paper, http://www.nielsen.com/ us/en/reports/2014/the-us-digital-consumer-report.html, accessed March 14, 2014. Ofcom (2013) Communications Market Report 2013. White paper. Ohnishi, H., P. Manchanda. 2012. Marketing Activity, Blogging and Sales. International Journal of Research in Marketing 29(2012) 221-234. 34

Park, Y. H. and Fader, P. S. 2004. Modeling browsing behavior at multiple websites. Marketing Science 23(3): 280-303. Peltier, J.W., B. Mueller, R.G. Rosen. 1992. Direct response versus image advertising: Enhancing communication effectiveness through an integrated approach. Journal of Direct Marketing 6(1) 40-48. Petty, R. E., and Cacioppo, J. T. 1986. The elaboration likelihood model of persuasion. Springer New York. Rubinson, J. 2009. VIEWPOINT-The New Marketing Research Imperative: It's about Learning. Journal of Advertising Research 49(1) 7. Teixeira, T.S. 2014. The Rising Cost of Consumer Attention: Why You Should Care, and What You Can Do about It. Harvard Business School Working Paper, No. 14-055. Teixeira, T.S., R. Picard, R. el Kaliouby. 2014. Why, when and how much to entertain consumers in advertisements? A web-based facial tracking field study. Forthcoming Marketing Science. Teixeira, T.S., M. Wedel, R. Pieters. 2012. Emotion-induced engagement in internet video ads. Journal of Marketing Research 49(2) 144-159. Television Bureau of Advertising (TVB). 2013. TV Basics. http://www.tvb.org/media/file/TV_Basics.pdf, accessed August 8, 2013 Tellis, G. J. 2004. Effective advertising: Understanding when, how, and why advertising works. Sage Publications. Tellis, G.J., R.K. Chandy, P. Thaivanich. 2000. Which ad works, when, where, and how often? Modeling the effects of direct television advertising. Journal of Marketing Research 37(1) 32- 46. U.S. Census. 2012. E-stats. http://www.census.gov/econ/estats/2010/2010reportfinal.pdf, accessed May 2013. Wilbur, K.C., L. Xu, D. Kempe. 2013. Correcting Audience Externalities in Television Advertising. Marketing Science, 32(6): 892-912. Wind, Y.J., B. Sharp. 2009. Advertising empirical generalizations: implications for research and action. Journal of Advertising Research, 49(2) 246-252. Wood, L. 2009. Short-term effects of advertising: Some well-established empirical law-like patterns. Journal of Advertising Research 49(2) 186-192. Wu, J., V. J. Cook Jr, E. C. Strong. 2005. A two-stage model of the promotional performance of pure online firms. Information Systems Research 16(4) 334-351. Xu, L., K.C. Wilbur, S. Siddarth, J. Silva-Risso. 2014. Price Advertising by Manufacturers and Dealers. Management Science, forthcoming. Zigmond, D., H. Stipp. 2010. Assessing a New Advertising Effect: Measurement of the Impact of Television Commercials on Internet Search Queries. Journal of Advertising Research 50(2) 162-168. Zigmond, D., H. Stipp. 2011. Multitaskers may be advertisers' best audience. Harvard Business Review 12(1/2) 32-33.

35

Table 1. Descriptive Statistics Kantar Comscore Unique Total Ad Search Direct Advertised Ad Trans- Conversion Sector Brand Ad Spendin Engine Visit Products Insertions actions Rate Creatives g ($MM) Referrals Sessions Dating Chemistry 1 4 2,264 7 2,196 2,220 15 0.34% eHarmony 1 48 16,124 53 7,793 19,430 58 0.21% Match.com 1 22 7,192 25 24,629 40,500 384 0.59% Pizza Domino's 5 50 23,904 150 7,210 12,203 10,806 55.66% Papa John's 14 25 6,709 55 3,019 10,817 6,740 48.71% Pizza Hut 13 65 22,783 168 5,388 10,227 3,373 21.60% Retailers Amazon 1 3 672 8 151,745 244,022 35,134 8.88% JC Penney's 13 99 18,159 174 23,208 32,596 5,137 9.21% Macy's 14 159 21,378 211 18,607 29,790 1,953 4.04% Overstock 1 12 3,866 18 17,108 20,112 1,396 3.75% Sears 16 161 26,512 210 16,816 22,266 919 2.35% Target 28 125 20,540 278 52,337 65,369 1,720 1.46% Victoria's Secret 1 16 4,657 51 10,605 21,662 4,755 14.74% Telecom. AT&T 8 180 83,919 1,126 41,928 166,335 3,889 1.87% Sprint 5 46 22,920 484 12,110 37,053 836 1.70% Verizon 7 168 57,442 826 24,731 138,073 8,206 5.04% Travel Expedia 3 30 12,938 66 15,972 46,832 1,611 2.57% Orbitz 1 12 4,876 17 6,461 22,791 786 2.69% Priceline 1 21 7,464 55 8,619 22,995 1,064 3.37% Southwest 3 23 698 106 10,732 22,747 2,026 6.05% Total 137 1,269 365,017 4,088 461,214 988,040 90,808 Average 7 63 18251 204 23061 49402 4540 6.27%

Table 2. Back-of-the-envelope Sampling Calculations

Variable Value Comment Number of total purchases in the sample 90,808 Source: comScore ComScore daily sample size 100,000 Source: comScore Number of computers in the US 470,850,000 Assumption: 1.5 desktops and laptops per person Sampling probability 0.02% Calculation Total ad spend $ 4,088,000,000 Source: Kantar estimates Cost/person/exposure $0.01 Assumption; typical TV ad cost Average response rate to direct response ads 0.10% Assumption; typical for DR TV ads Average response rate to brand ads 0.01% Assumption; 1/10 typical DR TV ad response rate Proportion of direct response ads 20.00% Assumption; Danaher and Green (1997) Proportion of ad responders who buy online 50.00% Assumption; 47% of our sample is online only Number of person-exposures to sample ads 408,800,000,000 Calculation Number of person-exposures delivered to comScore Calculation; assumes no correlation between panelists 86,821,705 sampling probability and ad exposures Number of purchases by comScore panelists stimulated by sample ads 24,310 Calculation: avg. response rate * num. exposures Number of online purchases caused by sample ads 12,155 Calculation: num. purchases * proportion online

36

Table 3. Survey Items by Ad type Direct Response Is there a call to go online (e.g., shop online, visit the web)? Is there online contact information provided (e.g., URL, website)? Is there a visual or verbal call to purchase (e.g., buy now, order now)? Does the ad portray a sense of urgency to act (e.g., buy before sales ends, order before ends)? Is there an incentive to buy (e.g., a discount, a coupon, a sale or “limited time offer”)? Is there offline contact information provided (e.g., phone, mail, store location)? Is there mention of something free? Argument Does the ad mention at least one specific product (e.g., model, type, item)? Is there any visual or verbal mention of the price? Does the ad show the brand or trademark multiple or few times? Emotion Is the ad intended to be emotional? (You may not agree. But was that the intention of the ad?) Does the ad give you a warm feeling about the brand? Does the ad tell a story (e.g., with characters, a plot, an ending)? Is the ad creative, clever? Is the ad intended to be funny? (You may not agree. But was that the intention of the ad?) Imagery Does this ad provide sensory stimulation (e.g., cool visuals, arousing music, mouth-watering)? Is the ad visually pleasing? Is the ad cute? (e.g., baby, puppy, animated characters)

Table 4. Ad Content Descriptives by Brand Num. Direct Resp. Argument Emotion Imagery Advertise Sector Brand Unique Ad min=0,max=7 min=0,max=3 min=0,max=5 min=0,max=3 d Products Creatives avg. (st.dev.) avg. (st.dev.) avg. (st.dev.) avg. (st.dev.) Dating Chemistry 1 4 3.8 (2.0) 1.0 (1.0) 1.5 (0.9) 0.3 (0.4) eHarmony 1 48 3.9 (1.2) 1.2 (0.5) 2.2 (0.9) 1.1 (0.9) Match.com 1 22 1.9 (0.8) 0.5 (0.6) 2.9 (1.0) 1.5 (1.1) Pizza Domino's 5 50 3.6 (1.5) 2.9 (0.4) 2.0 (1.2) 1.5 (0.7) Papa John's 14 25 5.6 (1.0) 3.0 (0.1) 1.0 (0.9) 1.4 (0.5) Pizza Hut 13 65 3.1 (1.3) 3.0 (0.2) 1.3 (1.4) 1.5 (0.7) Retailers Amazon 1 3 3.6 (1.1) 2.7 (0.7) 3.0 (0.0) 2.1 (0.4) JC Penney's 13 99 3.4 (1.7) 2.0 (0.8) 0.8 (1.0) 1.4 (1.0) Macy's 14 159 3.6 (1.7) 2.1 (1.1) 0.9 (1.4) 1.2 (0.9) Overstock 1 12 2.7 (1.0) 2.3 (0.7) 2.3 (0.8) 1.8 (0.8) Sears 16 161 3.8 (1.2) 2.0 (0.8) 1.7 (1.4) 1.0 (0.8) Target 28 125 1.3 (1.0) 1.2 (1.0) 2.1 (1.1) 1.7 (0.9) Victoria's Secret 1 16 0.9 (0.8) 1.2 (0.7) 0.3 (0.6) 1.6 (0.5) Telecom. AT&T 8 180 2.5 (1.6) 1.5 (1.0) 2.4 (1.2) 1.6 (0.9) Sprint 5 46 2.9 (1.2) 2.0 (0.8) 1.7 (0.9) 1.3 (0.7) Verizon 7 168 3.2 (1.9) 1.9 (1.0) 1.4 (1.2) 1.3 (0.6) Travel Expedia 3 30 3.4 (1.5) 1.5 (0.8) 1.6 (1.2) 1.6 (1.0) Orbitz 1 12 1.4 (0.6) 1.2 (0.4) 2.1 (1.1) 0.8 (0.8) Priceline 1 21 2.6 (1.5) 1.6 (0.6) 2.4 (1.2) 1.0 (0.5) Southwest 3 23 2.8 (1.2) 0.8 (0.9) 2.4 (0.9) 0.9 (0.6) Average (st. dev.) across all insertions 3.0 (1.7) 1.9 (1.0) 1.8 (1.3) 1.4 (0.8) 37

Table 5. R-Square Statistics by Online Shopping Variable and Model Specification

2 Hours (Diff.-in-Diff.) 2 Minutes (Quasi-Diff.) R-square R-square SE DIR TC AIC BIC SE DIR TC AIC BIC Treatment Effect Only 0.487 0.622 0.087 0.037 0.138 0.002

Baseline Only 0.693 0.868 0.206 862,555 887,457 0.044 0.172 0.154 2,030,240 2,050,468

Baseline + Treatment Effect 0.698 0.873 0.212 858,202 891,283 0.055 0.191 0.158 2,024,186 2,051,297

Baseline + Category-Specific 0.698 0.873 0.212 858,215 891,809 0.055 0.191 0.158 2,024,205 2,051,695 Treatment Effect Number of Observations 277,573 328,212

Table 6. Effects of Ad Content on Online Traffic and Transactions

2 Hours (Diff.-in-Diff.) 2 Minutes (Quasi-Diff.) SE DIR TC SE DIR TC Direct -.0170*** 0.106*** .0165*** -.0010** .0046*** .0005 Response (.0069) (.0142) (.0048) (.0007) (.0012) (.0005) Argument -.0260** -.2970*** .0360*** .0000 -.0110*** 1.02e- (.0119) (.0246) (.0083) (.0013) (.0021) (.0009) Emotion .0036 -.0980*** .0262*** .0003 -.0010 .0001 (.0078) (.0160) (.0054) (.0008) (.0013) (.0005) Imagery -.0120 -.1680*** -.0040 -.0020** -.0130*** .0020*** (.0108) (.0223) (.0075) (.0011) (.0019) (.0008) Standard errors in parentheses. *** p<0.01, ** p<0.05

38

Table 7. Additional Treatment Effects 2 Hours (Diff.-in-Diff.) Break # 2 Hours (Diff.-in-Diff.) Slot # 2 Hours (Diff.-in-Diff.) Program Type SE DIR TC (in prog.) SE DIR TC (in break) SE DIR TC Documentary -.2310** -.0020 0.177*** 2 .0391 .0264 -.0130 2 -.0090 -.0350 .0316 (.1130) (.2310) (.0782) (.0268) (.0551) (.0186) (.0282) (.0580) (.0196) Drama/Adventure -.2060 .0867 0.139*** 3 -.0210 -.1270** -.0110 3 -.0310 -.0560 .0147 (.1080) (.2220) (.0749) (.0286) (.0588) (.0199) (.0285) (.0586) (.0198) Entertainment -.0220 -.1430 0.202*** 4 -.0710** -.1090 -.0130 4 .0006 -.0830 .0003 (.1090) (.2240) (.0758) (.0309) (.0636) (.0215) (.0294) (.0604) (.0204) Feature Film -.1510 -.1600 0.143*** 5 -.0510 -.0940 -.0500** 5 -.0470 -.1250** .0282 (.1060) (.2180) (.0735) (.0346) (.0712) (.0241) (.0310) (.0637) (.0215) Instruction/Advice -.1590 .1090 0.153*** 6 -.1170*** -.1200 -.0140 6 .0091 -.0840 -.0140 (.1250) (.2560) (.0865) (.0428) (.0880) (.0297) (.0331) (.0681) (.0230) News -.1280 .0776 0.148*** 7 -.0380 .0743 .0364 7 -.0150 -.0900 -.0280 (.1120) (.2300) (.0776) (.0493) (.1020) (.0343) (.0366) (.0754) (.0255) Olympics .1920 2.335*** -.4350*** 8 -.1380** -.2190 .0294 8 .0007 -.0360 .0749*** (.2190) (.4510) (.1520) (.0588) (.1210) (.0409) (.0424) (.0872) (.0295) Other -.3360*** -.1890 0.188*** 9 -.1510** -.2500 -.0520 9 .0599 .0579 .0418 (.1290) (.2650) (.0894) (.0674) (.1390) (.0468) (.0506) (.1040) (.0352) Sitcom -.1370 .0819 0.134*** -.0630 -.3320*** -.0080 .0952*** .0193 .0580** (.1020) (.2100) (.0709) 10 or more (.0463) (.0952) (.0321) 10 or more (.0458) (.0941) (.0318) Slice of Life -.2080 -.2080 .1000 (.1070) (.2190) (.0740) Sports -.3430*** -.2110 0.231*** Past Spend -5.68e-10 -3.60e-08*** 5.58e-09*** (.1180) (.2430) (.0821) on Creative (1.61e-09) (3.32e-09) (1.12e-09) Suspense/Mystery -.0060 .3590 0.132*** (.1120) (.2300) (.0776) Talk -.2700** -.8310*** 0.162*** Own "Pre" -9.92e-07*** -3.21e-06*** 1.03e-07*** Comp. "Pre" 1.13e-07*** -2.63e-07*** 1.07e-07*** (.1140) (.2350) (.0792) Window (3.52e-08) (7.25e-08) (2.45e-08) Window Spend (1.84e-08) (3.79e-08) (1.28e-08) Unknown -.2870** -.0610 .1640 Own "Post" 3.46e-07*** -3.94e-07*** -1.09e-07*** Comp. "Post" 9.17e-08*** 3.81e-07*** 2.12e-08 (.1460) (.3000) (.1010) Window (3.63e-08) (7.47e-08) (2.50e-08) Window Spend (1.90e-08) (3.91e-08) (1.32e-08) Variety Music -.1410 -.0820 .1150 (.1220) (.2510) (.0849)

Table 8. Effects of Television Network (Significant Effects Only) 2 Hours (Diff.-in-Diff.) 2 Hours (Diff.-in-Diff.) 2 Hours (Diff.-in-Diff.) TV Network SE DIR TC TV Network SE DIR TC TV Network SE DIR TC ADSM 0.273*** 0.520** 0.115 FNEW -0.0836 -0.415** 0.0328 NGC 0.0196 -0.461** -0.0771 (0.104) (0.215) (0.0725) (0.0942) (0.194) (0.0654) (0.0931) (0.192) (0.0647) AMC 0.309*** -0.0456 0.00471 FX 0.0909 0.486*** 0.130*** SPK -0.0683 -0.334** 0.0859 (0.0989) (0.204) (0.0687) (0.0687) (0.141) (0.0478) (0.0652) (0.134) (0.0453) BET -0.0709 -0.321** 0.0692 GALA 0.0565 -0.451** 0.0892 TLC 0.0435 0.347** 0.0113 (0.0727) (0.149) (0.0505) (0.0900) (0.185) (0.0625) (0.0780) (0.160) (0.0542) BRAV 0.0677 0.316** 0.0613 HIST -0.0158 -0.330** 0.0744 TNNK 1.386*** 0.396 -0.300 (0.0766) (0.158) (0.0532) (0.0702) (0.145) (0.0488) (0.335) (0.690) (0.233) CNBC -0.457*** -2.504*** -0.237*** MNTV 0.106 1.205** -0.225 TOON 0.560** 0.863 0.160 (0.124) (0.255) (0.0861) (0.273) (0.561) (0.190) (0.246) (0.506) (0.171) CNN -0.109 -0.781*** 0.0726 MTV 0.0887 0.376*** 0.0657 TRU 0.290*** 0.114 0.107 (0.0888) (0.183) (0.0617) (0.0613) (0.126) (0.0426) (0.0957) (0.197) (0.0664) CW -0.100 -0.732*** 0.0316 NAN -0.115 -0.359** 0.0610 TWC -0.127 0.641*** -0.159** (0.0969) (0.199) (0.0673) (0.0827) (0.170) (0.0574) (0.106) (0.219) (0.0738) E! 0.197*** 0.372*** -0.0394 NBC 0.0911 0.326** 0.0203 USA 0.0254 -0.0228 0.102** (0.0658) (0.135) (0.0457) (0.0775) (0.159) (0.0539) (0.0685) (0.141) (0.0476) ESP2 0.188** -0.0916 0.0307 NFLN 0.517*** 1.750*** -0.583*** VH-1 0.0731 0.253** 0.108*** (0.0950) (0.195) (0.0660) (0.156) (0.320) (0.108) (0.0594) (0.122) (0.0413) ESPN 0.296*** 0.201 -0.0908 (0.0856) (0.176) (0.0595)

39

Table 9. Effects of Advertised Product (Significant Effects Only) 2 Hours (Diff.-in-Diff.) 2 Hours (Diff.-in-Diff.) Product SE DIR TC Product SE DIR TC AT&T : Consumer Wireless Service -0.0535 1.620*** -0.183*** Sprint Everything Data Family Plan : -1.671** 1.870 -0.539 (0.0777) (0.160) (0.0540) Consumer Wireless Service (0.664) (1.366) (0.461) AT&T : Pre-Paid Wireless Service 0.327 5.794*** -0.871** Sprint Everything Data Plan : Consumer -1.634*** -0.551** -0.103 (0.539) (1.109) (0.374) Wireless Service (0.117) (0.241) (0.0774) AT&T Go Phone : Pre-Paid Wireless -1.761*** -0.0111 0.201 Sprint Wireless Service : Consumer -1.621*** -0.565** -0.167** Service (0.481) (0.989) (0.334) Wireless Service (0.108) (0.223) (0.0709) AT&T Inc : Corporate Promotion -0.900*** 2.040*** -0.0614 Verizon : Business Wireless Service -0.101 0.442 -0.513*** (0.107) (0.220) (0.0742) (0.124) (0.256) (0.0833) AT&T Mobile TV : Consumer Wireless 0.218 5.226*** -0.753*** Verizon : Consumer Wireless Service -0.305*** 1.555*** -0.758*** Service (0.135) (0.278) (0.0939) (0.0952) (0.196) (0.0623) AT&T Unlimited Calling Plan : Consumer 0.340** 4.724*** -0.789*** Verizon : ISP/TV/Wireless 0.135 -0.571 -1.172* Wireless Service (0.149) (0.307) (0.103) (0.920) (1.892) (0.638) Eharmony.com : Online 0.594*** 3.028*** -0.0189 Verizon Communications : Corporate 0.562*** 0.0360 -0.755*** (0.186) (0.413) (0.128) Promotion (0.173) (0.357) (0.118) Match.com Dating Service : Online 1.906*** 4.989*** 0.0164 Verizon Family Share Plan : Consumer -0.633*** -0.958*** -0.0682 (0.208) (0.452) (0.141) Wireless Service (0.131) (0.270) (0.0885) Sprint Any Mobile Anytime Plan : -2.233*** -1.989*** -0.730*** Verizon Nationwide Unlimited Talk : -0.873*** -6.013*** 0.452*** Consumer Wireless Service (0.163) (0.337) (0.110) Consumer Wireless Service (0.166) (0.343) (0.113) Sprint Corp : Corporate Promotion -1.618*** -1.661*** -0.363*** (0.175) (0.361) (0.119)

Table 10. Brand-Specific Effects of Prior Traffic and Transactions 2 Hrs. (D.-in-D.)DV: SE POST Dir POST TC POST Predictor: SE PRE Dir PRE TC PRE Est. Cost SE PRE Dir PRE TC PRE Est. Cost SE POST Dir POST TC PRE Est. Cost Dating Chemistry .0289 0.138*** -2.2350 4.63e-05 -.1780 0.243*** -2.3750 -0.000215** -.0020 -.0110 .0550 1.05e-05 (.0672) (.0615) (1.2650) (5.27e-05) (.1380) (.1270) (2.6020) (0.000109) (.0509) (.0504) (.8800) (3.65e-05) eHarmony 0.150*** 0.104*** -.4060** -7.78e-06 0.267*** 0.204*** -.4940 -2.32e-05 .0016 .0060 .1080 -1.64e-06 (.0165) (.0092) (.1580) (1.49e-05) (.0340) (.0189) (.3250) (3.07e-05) (.0124) (.0066) (.1090) (1.04e-05) Match.com 0.388*** .0391*** 0.210*** -2.39e-05 0.156*** 0.263*** 0.547*** -8.58e-05** .0027 .0139*** .0214 -6.84e-06 (.0114) (.0091) (.0984) (1.65e-05) (.0238) (.0190) (.2020) (3.39e-05) (.0086) (.0065) (.0678) (1.14e-05) Retailers Amazon 0.561*** 0.152*** -.0280 -9.26e-05*** 0.539*** 0.487*** -.0410 -0.000161*** 0.107*** .0786*** 0.128*** -6.87e-06 (.0122) (.0078) (.0250) (5.34e-06) (.0256) (.0160) (.0514) (1.10e-05) (.0091) (.0060) (.0147) (3.65e-06) JC Penney's 0.187*** .0851*** -.0500*** 1.62e-06 0.300*** 0.262*** -.1470*** 6.68e-06*** .0797*** 0.172*** .0185** 5.16e-07 (.0081) (.0060) (.0152) (8.60e-07) (.0168) (.0127) (.0314) (1.77e-06) (.0059) (.0041) (.0095) (5.97e-07) Macy's .0508*** .0673*** 0.104*** 5.35e-06*** 0.179*** 0.228*** .0983*** 4.23e-06 .0920*** 0.104*** .0392*** -1.32e-06 (.0082) (.0060) (.0181) (1.11e-06) (.0169) (.0125) (.0373) (2.29e-06) (.0057) (.0041) (.0120) (7.71e-07) Overstock .0681*** .0599*** .0238 5.11e-06 .0538 0.106*** .0719 1.28e-05 .0508*** .0802*** .0110 -1.54e-06 (.0182) (.0166) (.0720) (3.22e-06) (.0378) (.0345) (.1480) (6.62e-06) (.0127) (.0120) (.0488) (2.25e-06) Sears .0174*** .0725*** .0204 1.67e-06 .0368*** 0.312*** .0763** 2.30e-06 .0519*** .0792*** .0253** 7.75e-07 (.0076) (.0054) (.0222) (8.75e-07) (.0158) (.0111) (.0458) (1.80e-06) (.0055) (.0038) (.0149) (6.07e-07) Target 0.289*** 0.253*** -.0920*** -9.81e-06*** 0.250*** 0.441*** -.2180*** -1.12e-05*** -.0040 .0822*** .0763*** 6.57e-07 (.0047) (.0040) (.0213) (8.97e-07) (.0096) (.0083) (.0438) (1.84e-06) (.0035) (.0029) (.0142) (6.20e-07) Victoria's S. 0.104*** .0359*** .0329 1.03e-05*** .0776*** 0.297*** -.1300*** 1.05e-05*** 0.153*** 0.257*** .0015 -1.57e-06 (.0145) (.0121) (.0217) (1.91e-06) (.0299) (.0251) (.0450) (3.93e-06) (.0110) (.0077) (.0134) (1.32e-06) Telecom. AT&T .0739*** .0860*** -.0990*** -7.63e-07** 0.210*** 0.569*** -.3050*** -3.63e-06*** .0973*** .0216*** -.0090*** -4.28e-07 (.0031) (.0010) (.0049) (3.16e-07) (.0064) (.0021) (.0100) (6.49e-07) (.0023) (.0007) (.0033) (2.19e-07) Sprint -.0670*** -.0470*** .0006 1.39e-06*** -.0180 .0484*** -.0240 8.21e-06*** 0.105*** .0791*** .0843*** -1.75e-07 (.0117) (.0058) (.0279) (4.63e-07) (.0243) (.0119) (.0573) (9.52e-07) (.0087) (.0041) (.0189) (3.20e-07) Verizon .0695*** .0098*** -.0130*** 4.79e-07 0.219*** 0.542*** -.4490*** -1.15e-06 0.184*** .0642*** -.0040 1.02e-06*** (.0044) (.0012) (.0037) (3.87e-07) (.0091) (.0025) (.0076) (7.95e-07) (.0034) (.0009) (.0025) (2.68e-07) Travel Expedia .0797*** 0.110*** -.1090** -1.05e-06 0.155*** 0.262*** -.2030** -4.88e-06 .0381*** .0338*** .0746*** -1.87e-06 (.0185) (.0099) (.0486) (9.80e-06) (.0380) (.0209) (.1000) (2.02e-05) (.0132) (.0071) (.0325) (6.80e-06) Orbitz -.0210 -.0140 .0921 -2.51e-05 .0569 .0253 .1210 -1.42e-05 .0358 .0269*** 0.114*** -1.01e-06 (.0300) (.0136) (.0848) (1.85e-05) (.0617) (.0284) (.1740) (3.80e-05) (.0222) (.0101) (.0580) (1.28e-05) Priceline .0570*** .0212** -.0020 -2.63e-07 .0952*** .0027 0.207*** 4.74e-07 .0243 .0425*** .0751*** 2.31e-07 (.0190) (.0120) (.0535) (3.68e-06) (.0392) (.0251) (.1100) (7.57e-06) (.0148) (.0085) (.0363) (2.55e-06) Southwest .0584 .0679** .0248 1.39e-06 -.0800 -.0080 0.589*** 5.93e-06*** .0557 .0645*** -.0590 -6.90e-07 (.0553) (.0389) (.1560) (9.32e-07) (.1140) (.0810) (.3210) (1.92e-06) (.0437) (.0265) (.1050) (6.62e-07)

40

Figure 1. Consumer Decisions Online

Figure 2. Conceptual Distinction among TV Advertisements TV Advertisements

Direct-response Brand-image

Cognitive Affective

Argument-based

Emotion-based Imagery-based

41

Figure 3. Traffic and Transactions by Product Category and Hour

42

Figure 4. Ad Timing and Frequency

Figure 5. Timing Illustration

43

Figure 6. Traffic by Window Length and Ad Content

Note: Increases prior to the start of the ad are an artifact of the discrete time interpolation in the graph.

Figure 7. Traffic within Shorter Window Lengths

Figure 8. Effects of Ad Content on Online Traffic and Transactions (2-hour data) Direct Direct + Traffic  Traffic Direct Response + Emotion + Transactions Transactions n.s.  Search Total Effect is Search Total Effect Engine pos. for all 17 Engine is pos. for Referrals brands Referrals 15 brands; n.s. for 2 + +

Direct Direct  Traffic  Traffic Argument + Imagery + Transactions Transactions  n.s. Search Total Effect Search Total Effect Engine is pos. for 9 Engine is neg. for Referrals brands; Referrals 10 brands; n.s. for 8 n.s. for 7 + n.s. 44