Understanding programmatic TV advertising

ABSTRACT

Television is undergoing tremendous technological developments, which will enable marketers to direct commercial messages to more specific audiences at the individual and/or household level. Traditional ways of buying TV advertising are being challenged by the programmatic approach, which originated with search and display ads, and uses data and technology and real-time auctions to automate transactions between buyers and sellers. We discuss the changes to TV that enable the rise of programmatic models and propose that its future success depends on the coordination and availability of three factors, abbreviated DAD: distribution, ad inventory and data. We also elaborate on the effects of media context and how the TV program environment matters for advertising effectiveness. Finally, we discuss the future of TV advertising and related research needs.

KEYWORDS:

• Programmatic advertising
• TV
• context effects

Introduction

There are two different, and increasingly competing, approaches to buying/selling advertising media. The mass approach is characterized by media companies selling mass audiences through a sales force to advertisers who inform their decisions with probability samples such as Nielsen panel data. The computational advertising approach is marked by purchasing individual exposures informed by whatever data may be available about the device, cookie, household, etc. Often the media sale takes place on an ad exchange in real time through an automated auction, the message may be personalized, and sometimes the behavioural responses to the exposure can be tracked (e.g. clicks, conversions). This is commonly called programmatic advertising.

The mass approach has been dominant for decades and has been the focus of advertising and media buyers and sellers, industry associations and scholarship. The genesis of the computational approach dates back to the 1990s with ventures like Overture and Google AdWords for sponsored search and similar models for display ads. These products were created by the tech community and, on the academic side, have been mostly studied by computer science scholars (e.g. Wang, Zhang, and Yuan 2016; Leskovec, Rajaraman, and Ullman 2012, Ch. 8). The computational approach also has much in common with direct marketing and customer relationship management (CRM), in that it makes decisions at the customer (or household or device) level informed by customer databases and increasingly machine learning (Perlich et al. 2014), and directly observes outcomes such as conversion. Advertisers can still deliver to mass audiences, but they are now doing so by making individual choices on who to address advertising to, when to deliver it, and what commercial message will lead to the best outcomes, all in near real time.

The fundamental goals for the two approaches are different. A high priority and advantage for the mass approach is to reach a large audience with many exposures (GRPs) at low CPMs. This is sometimes called efficiency (Smith and Park 1992). On the other hand, one of the main selling points of the computational approach is the possibility to expose only the most interested consumers with the most relevant messages at the right time (although with click-through-rates commonly on 100ths of a percent, the current practice of programmatic often fails to realize this possibility). Instead of reaching a large audience at low cost, the ultimate goal is more about the return on investment over time (lifetime value), which was also the goal of direct marketing and CRM. This is sometimes called effectiveness.

The two approaches have been on a collision course. In the beginning, the computational approach was confined to isolated areas of advertising such as digital display and search, and only a tiny fraction of advertising budgets went to such ‘digital’ ad channels. Over time, digital advertising has grown quickly and computational approaches are spreading as more digital media channels have been introduced (e.g. social media and mobile apps). We are reaching an inflection point where the last bastion of traditional advertising – TV (the overwhelming majority of mass advertising budgets go to TV) – is experimenting with computational approaches. Furthermore, as more advertisers embrace computational approaches in digital display, mobile and social media advertising, there is increased pressure for TV to offer the same advantages to advertisers.

This article attempts to understand the impending collision and how the computational approach will play out in TV. Applying computational approaches to TV advertising is still nascent and the stakeholders will have to make many decisions and agreements before it will be more widely practiced. Our goal is to give the reader perspective to understand events as they unfold by tracing the history of TV advertising approaches as depicted in Figure 1. This can inform computational approaches for TV so that they improve over what we have seen with other digital media. We introduce a framework with the acronym DAD: delivery, advertising inventory, and data. Each of these three variables has a profound influence on how TV advertising is bought and sold. We elaborate further on data by proposing a framework that systematizes the role of the different data sources available for programmatic TV advertising. This will help in understanding the data ecosystems that are necessary to sustain computational approaches. Furthermore, we discuss whether the surrounding media context should be considered or whether a content-agnostic approach will yield similar predictions. We develop theoretical reasons why media context should be considered. We close with discussion of research questions created by the collision.

Figure 1. The history and evolution of television.

The DAD framework

Television has been the cornerstone of consumers' media consumption for decades. And, correspondingly television advertising has been the centrepiece of the marketing mix as well. Each phase in the history of television has had a profound effect on advertising (see Figure 1). The new era of computational TV advertising (referred to as programmatic) requires all three elements of DAD: delivery, advertising inventory and data to be optimized and work in tandem. Delivery technology must be addressable down to the individual household and/or device level (Kitts, Au, and Ulger 2014). Advertising inventory must be readily available (as opposed to only a few commercial minutes per hour or remnants of unsold inventory from the mass model). Finally, available data must provide a comprehensive view of each individual program audience member in order for marketers to make informed decisions on whom to buy. Furthermore, there must be coordination among all three of these variables in order for the programmatic approach to work. Since different companies control the three aspects, coordination will require agreements, partnerships or acquisitions.

Stakeholders in the television advertising landscape

In order to understand the TV advertising landscape and the potential application of programmatic advertising to TV, we must first identify the various stakeholders involved (Figure 2). In the linear broadcast era the stakeholders were limited to four primary participants: TV networks, audiences who consumed TV content, audience measurement firms who created a common currency view of these audiences and advertisers who entered into negotiations with the networks to purchase advertising in their programs (Napoli, 2003). As TV distribution evolved into a cable/satellite distribution model, the distributor became a new and important stakeholder in the mix. Although the available ad inventory they had to sell directly to advertisers was limited, the distributors had the technology to serve TV ads down to the household. As television migrates over the top (OTT), new stakeholders emerge: both new distributors such as streaming services and smart TV applications, as well as new audience data providers who come from outside the established oligopoly of media measurement firms such as Nielsen and comScore. Furthermore, with programmatic TV an entirely new category of stakeholder has emerged: the exchanges that operate the programmatic marketplaces. Each stakeholder has specific interests in the ecosystem and no one has control over all the variables that are required to optimize programmatic TV. Furthermore, as programmatic technology is fuelled by big data as opposed to probability samples, the government, especially in Europe but not in China, is emerging as a watchdog to protect the audience stakeholder from data mining practices that compromise privacy. This ecosystem is highly complex and warrants discussion of individual stakeholders.

Figure 2. Stakeholders in the programmatic television ecosystem.

First, TV networks make programming decisions to create a content product that will attract desirable audiences for sale to advertisers (Napoli 2003). Networks include traditional broadcast and cable programmers as well as newer content creators. They create ad inventory and advertising is an important revenue source. These networks have advertising sales forces to sell ads in their inventory directly to advertisers. Under the mass model, ads are shown to 100% of the distribution footprint. Unless a network has an OTT streaming option, they do not control distribution to individual audiences or households. Until recently, the networks relied on a broad demographic measure of their program audiences to convince advertisers that their audiences were highly desirable and a good fit with the advertiser's view of their customers.

Next, the distributor is anyone responsible for getting the TV content into the home/onto a device: Sling TV, Roku, Cable/Satellite, Hulu, Netflix, Amazon, Facebook (in the future), YouTube (with YouTube TV), Disney Streaming App, etc. The linear TV distributors, such as cable companies, satellite companies and telephone companies/internet service providers, control content distribution down to a household level. The set-top boxes they install in the household to deliver the signal gather a wealth of viewing information, which is commonly called set-top-box data (Yeh 2014): these distributors know when the set is turned on and what channels are being viewed. Historically, linear TV distributors did not focus their business models on selling advertising to individual households. They made their money mainly from subscription fees. However, as more households began to cut the cord and cancel their subscriptions in favour of OTT delivery options, distributors are becoming increasingly interested in selling advertising to make up for subscriber revenue losses.

Linear TV distributors control the technology that enables improved targeting of ad messages to individual households. Under this model, data-sets that enable marketers to understand the characteristics of the household: who they are/their past behaviour/what they watch are desirable. It is important to note that linear distributors can only sell advertising in the commercial time that is given to them by the networks. Since the networks do not receive any revenue from the commercial inventory they pass along to the distributors by obligation of their carriage agreements, they currently tend to offer remnant inventory to the distributors as opposed to their highest quality fare. As a result, TV distributors tend to sell advertising at the household level as content agnostic (Ferber 2016) meaning as long as the ad is placed in the correct household the obligation is met. There are no guaranteed placements adjacent to specific content.

While linear TV distributors are handicapped by the amount and quality of ad inventory that is made available to them to sell through live TV airings, they have programs in video-on-demand libraries that, through the process of dynamic ad insertion, enable marketers to place ads in specific programs and specific households. Some distributors have joined forces to make this opportunity as convenient as possible for all parties (see, e.g. Canoe Ventures that includes 35MM households from across three major MSOs: Comcast, Cox and Charter Spectrum). In subscriber households, advertisers can insert ads in specific programming from participating networks. This nascent form of TV advertising is growing and in the first quarter of 2017 over 2300 campaigns were run through the platform (MediaPost ‘Television News Daily’ April, 2017). It requires data on both household characteristics to determine the right households in which to serve the ads as well as program audiences to determine which programming in the video-on-demand library attracts audiences that best fit the advertiser's target customer.

Furthermore, as TV content is becoming available OTT, two new sets of distributors with different business models, technology platforms and access to audience data are emerging: subscription streaming services such as Hulu that aggregate content from various television networks into one Internet hub and connected TV devices such as Roku, Chromecast, Amazon Fire TV and Apple TV that function similar to an app store where various TV networks can create their own content ecosystems that audiences can select to view where/when they wish. These services are positioning OTT advertising as the best of both worlds: the best of digital and linear TV advertising all in one. They enable advertisers to select individuals to receive ads based upon a host of characteristics from third-party data sources, including their Internet activity. At the same time, advertisers can select specific TV content they want their ads to run in based upon audience viewing data generated within the platform. Because users have to log into their accounts, these services become smarter about their audiences over time: who they are, what they like to watch and what advertising interests them (Fisher 2017).

Government and industry regulators are another stakeholder, which will affect all aspects of DAD. Perhaps the most famous and important regulation is the European Union's General Data Protection Regulation (GDPR), which takes effect in 2018. GDPR implements strict privacy protections, which will affect how data can be used to make programmatic decisions. The Trump administration's decision not to enforce net neutrality is another example of a recent government action that affects the digital distribution of video content. For an example of recent regulation on ad inventory, recall that programmatic advertising requires integration of the components of DAD. The Trump administration recently blocked the merger between CNN, which creates ad inventory, and AT&T, which provides mobile phone and household Internet service (distribution). This merger would have allowed for easier coordination between the ad inventory and distribution. The merger between NBC (ad inventory) and Comcast (distribution) is discussed below. The investigations around the manipulations of the 2016 US election, Brexit and others through social media (e.g. Cambridge Analytica) will likely also prompt further regulation of social media and other digital advertising platforms around the world.

Data ecosystem used for estimating ad receptivity in programmatic models

Data is at the heart of computational approaches. The more data on individuals that programmatic models have, the better the targeting and personalization. When individual-level (or household/device) data are not available, the advantages of programmatic diminish. Of course, the programmatic ecosystem also requires ad inventory and the ability to deliver ads to individual devices (DAD).

Figure 3 provides a holistic view of the various data sets that can predict outcomes in programmatic advertising. As we discuss below, not all of the data sources shown in the data management platform are always available. Moreover, different data sources are often owned by different stakeholders, implying that the stakeholders must come to an agreement on sharing the information with each other, which may involve trading or paying for data-sets. The fourth phase in Figure 1 of walled gardens is when a single firm attempts to gather all of the data sets in Figure 3, as well as controlling distribution and ad inventory. While some data-sets may be available they are sometimes not used because the cost of obtaining the data may exceed the benefit. In addition, data trading practices raise consumer privacy issues, requiring transparent data processing practices and involving policy makers and regulators such as the U.S. Federal Trade Commission (FTC 2012) or European Data Protection Authorities and consumer organizations (Boerman, Kruikemeier, and Zuiderveen Borgesius 2017).

Figure 3. Framework showing how different data sources are used to make programmatic decisions.

In order to make informed decisions, we need to look at various data-based views of the audience: their past behavioural data, their demographic characteristics, their viewing history and also the current context for the message. Machine learning algorithms link these data-sets to indicators of product interest.

Past behavioural data

This begins with CRM data from the advertiser's (first-party) internal databases recording, for example, previous purchase history such as recency, frequency and monetary value (RFM) (e.g. Roberts and Berger 1999, 95). It is well established in direct marketing that RFM summaries are associated with future purchases from the focal brands (e.g. Zhang, Bradlow, and Small 2015; see Malthouse and Raman 2013 for RFM empirical generalizations). The literature on habitual behaviours suggests that when individuals engage in behaviours that are habitual, they use simplified decision rules, which results in enacting the same behaviours as in the past (Conner and Armitage 1998). Previous behaviours seem to be an important predictor of future behaviour (see Ouellette and Wood 1998 for a review).

Past behaviours could also include search or shopping data, which drive current retargeting efforts. For example, if a customer visits a retailer's website and studies a certain product, the retailer can currently buy display ads on other websites (e.g. news sites) showing that product to the consumer (Lambrecht and Tucker 2013). The record of shopping behaviours on the retailer's website such as what items have been viewed or placed in a shopping cart belongs to the retailer and can be used for targeting and personalization. Such data can help marketers deliver effective messages. For example, Bleier and Eisenbeiss (2015) found that personalizing banner ads based on products consumers placed in their shopping carts during the shopping visit increases click-through rates, and Choi (2013) showed that customers convert more when exposed to retargeting ads because such ads trigger more traffic back to the sites. Such data are often not available when the advertiser's products are mainly sold through third-party channels, for example, consumer packaged goods sold mostly at retail stores.

Past behavioural data might also include second-party data, which can be defined as another organization's first-party data. For example, suppose the focal advertiser is a retailer that sells camping supplies. It may be possible to buy or trade for information from other retailers in the same category, such as whether some household has recently purchased a tent from a different retailer. The tent purchase is RFM information owned by the other retailer and is therefore second-party data. This is a modern incarnation of list rentals, which has been practiced for decades by direct marketers (e.g. Roberts and Berger 1999, Ch. 4).

Household characteristics data

Focusing on past behavioural data is applicable when advertisers are interested in talking to existing customers and/or cross-selling new products and services based upon past purchases. But what about new customer acquisition? In these instances, first-party data is, by definition, not available. Demographics can often be ‘rented’ from a third-party data provider for large numbers of prospective customers. From previous research, we know that basic demographics such as age, gender, income, occupation and race are, overall, poor predictors of behaviour (Haley 1968), but in the absence of stronger predictor variables, demographics will often provide some predictive power.

Household media consumption data

Media consumption data includes which programs a consumer (or household) watches or which pages a person likes on Facebook. The recent coverage of Cambridge Analytica's use of Facebook data in the Trump campaign is now a famous example of using media consumption to infer product (e.g. Trump or Brexit) interest. Likewise, a study of viewing patterns can lend clues as to the types of products and services meet the needs and preferences of viewers within the household that are less obvious. The underlying assumption is that it is useful to know the association between the media consumed and product interest for targeting reasons, even if the causality is suspect because some other variable is causing both media consumption and product interest. For example, a model may show a strong association between viewing cartoons (media consumption) and an interest in juice boxes (advertiser), yet both are consequences of a third variable, namely the presence of small children. As a second example, viewing Fox News or liking certain Facebook pages (media consumption) may be associated with voting for Republicans (advertiser), but both are likely consequences of political beliefs. Hence, the correlations are due to a third variable (political beliefs).

While it is desirable to include measures of omitted variables in the predictive model, often measures are not available and, even when they are, it is better to also include viewing history. In the case of cartoons, young children and juice boxes, the presence of young children is often available from third-party data sources. Even so, consider the possibility of a grandparent who frequently has grandchildren visiting. The HH of the grandparent does not have any children in residence, yet the grandparent may purchase juice boxes for the visits. Thus, the viewing history would reveal a need that the demographics would not.

Current context

The current context refers to situational aspects of the potential exposure, which are summarized as who, what, when, where, why and how. Knowing any combination of these factors can be useful information when making an ad placement decision. By who, we mean who will see the ad? For example, a household may be composed of a mother, father and child, and an advertiser would want to know exactly who (i.e. which combination of household members) is watching when making an exposure decision. When refers to the time of day (Tellis, Chandy, and Thaivanich 2000). For example, consumers may be more responsive to certain types of messages depending on the time of day. When may also be indicative of who is watching. Where refers to the geographic location of the exposure, which becomes especially important with exposures on mobile devices (e.g. Bernritter, Ketelaar, and Sotgiu 2018; Larivière et al. 2013). Why refers to the consumer motivations for watching the program, which could signal an aperture for certain products (e.g. Rubin 1983). How refers to the device type (e.g. live TV, time-shifted TV, desktop computer, tablet or smartphone), and can further inform placement decisions. For example, Wang, Malthouse, and Krishnamurthi (2015) find that consumers are more likely to make simple, habitual decisions on mobile devices and purchases requiring consideration on PCs. Finally, what refers to the program currently being watched. For example, an ad placed in a sad drama may produce a different effect than the same ad placed in football game. The what, i.e. the program the household is currently watching, is easily known. Malthouse and Calder (2010) show that experiences with the media context are of comparable importance to factors traditionally used to price advertising, implying that media could charge higher CPMs for placements in more engaging contexts, because such ads will be more effective.

The ‘what’ (program context) may affect reactions to advertising and should therefore be considered when making programmatic placements. Some programs may be more effective for advertising than others (regardless the nature of the ads). One explanation for this is that some programs are more interactive than others (Cauberghe, Geuens, and De Pelsmacker 2011). Another explanation is the average consumer experience level with the media program, where experiences refer to thoughts and beliefs about the extent that a media product satisfies personal goals (Bronner and Neijens 2006; Malthouse, Calder, and Tamhane 2007, 2010; Calder, Malthouse, and Schaedel 2009). Advertising within or around programs that stimulate the viewer (i.e. are described as exciting or fascinating) is experienced as stimulating (Bronner and Neijens 2006), illustrating a spillover effect of experiences evoked by the program over the ad. Such an effect is in line with the priming effect, where the program in which the ad is placed serves as a priming cue that evokes emotions, stirs the viewers’ attention and activates their associative networks, which guides the interpretation of the ad (Dahlén 2005). Many studies have investigated possible media congruence effects, providing further evidence for the role of the surrounding context (see literature reviews in Dahlén et al. 2008 and De Pelsmacker, Geuens, and Anckaert 2002).

Validation of our model

If our model in Figure 3 has validity, we should expect to find evidence for it in practice. We believe there are many examples of ‘walled gardens’ such as Amazon. Starting with previous behaviours, it knows about previous purchases and shopping actions across hundreds of categories. This information can be joined with records of media consumption including books (Kindle), music, video and news (a trial subscription to the The Washington Post is free to all Amazon Prime users). Amazon Echo hears everything, including what TV programs are being viewed. While there are many factors that explain Amazon's success, the fact that Amazon owns such a complete picture of its individual customers gives it a competitive advantage that is not easily duplicated by its competitors.

As a second example of a walled garden, consider the picture that Google has assembled about its customers. Having a record of general search behaviours gives it unique knowledge about individual customers’ current needs and interests. Google has uniquely strong information about contextual variables because of Android and Google Maps. They know where their customers are at all times and can often infer why they are there because of Google Calendar and Gmail. They also control ad inventory and distribution by owning YouTube, which also gives them a record of media consumption, along with Google News, Chrome (browsing history), Google Play, etc. These products also give Google ad inventory (e.g. the ads that play before and during a YouTube video) and distribution, although its access to more traditional TV programming audiences is currently limited. While we do not know whether Google combines data from all of these apps, the potential exists for Google to have this complete customer view, which would give it a competitive advantage.

A very different example comes from the merger between NBCUniversal (a media and entertainment company) and Comcast Corp (a broadcasting and cable television company, and Internet provider). Comcast has complete viewing histories and the ability to deliver addressable TV. NBC controls a large amount of premium ad inventory. As programmatic TV evolves, Comcast/NBCUniversal has options that other content creators and distributors do not have. Regulators understand the power of controlling DAD and were concerned that an all-powerful Comcast might stifle competition from new online video competitors including Hulu, in which it now owns a stake. Among the conditions to which Comcast agreed: relinquishing management rights of its minority stake in Hulu. Hulu is co-owned by News Corp, Walt Disney Co and NBC Universal. Our framework also reveals a gap in their portfolio (that they will need to fill): product interest.

Similar discussions could be given for Alibaba, Tencent, Facebook and others. In particular, the recent coverage of Cambridge Analytica has raised the public's awareness of Facebook's data assets.

Future research questions

The developments in the TV advertising industry discussed above inspire new research of questions.

• RQ1: When and at what ratio will the two approaches – mass and computational – reach an equilibrium? While digital ad budgets have been growing quickly over the past 20 years, we assume that neither world will vanish anytime soon. When will the proportions stabilize, and at what levels? This may vary by geographic region, as discussed below. A similar question could be asked for other forms of advertising, such as owned media and engagement marketing, discussed in the conclusion.

• RQ2: Will regulation save mass advertising? Privacy and other regulations are becoming increasingly strict. As the costs of data breaches, data security and regulatory compliance increase, many firms may opt out of the one-to-one, database approach, and continue using the mass approach.

• RQ3: Will the asymmetry in data regulation between the EU, Japan, USA and China push firms down one path or the other? The EU recently imposed at $122 million fine on Facebook for joining information between Facebook and WhatsApp, which it also owns. Individual EU countries have also imposed fines on Facebook for violating data protection rules in targeted advertising (Bendix 2017). China, on the other hand, has few privacy or antitrust regulations.

• RQ4: How can data from different sources be combined while complying with privacy regulations. There are opportunities to develop new methods that inform programmatic decisions and comply with regulations. See Berry and Rieder (2017) for an example.

• RQ5: How to measure and optimize exposures for multi-screen consumption. There is already some research on this topic (e.g. Bellman et al. 2013).

• RQ6: Past behavioural and third-party data have been used for years in direct marketing. The other two data types – media consumption history for all households and various contextual variables – are comparatively new. Since they are newer, there are research opportunities to understand why and when they are indicators of product interest. Proposing ways of harnessing contextual variables to improve advertising seems especially fertile.

• RQ7: To what extent does the ‘who’ in the context matter and how can we infer who is in the room? TV has, historically, been a household medium. When we consider TV audiences we must view them both as households and as persons within the household who view particular shows at specific times throughout the day. Knowing who does not end with knowing the household demographics. By the studying content being watched, we can try to infer which person in the household is currently in front of the TV. In addition, people who are in the room can affect viewers’ choices and reactions to advertising. Group watching does not happen offline only. With the development of cooperative media spaces (Gross, Fetter, and Paul-Stueve 2008), it is moving online, which brings new data and research questions. Also potentially relevant is the field of social TV, which describes the simultaneous consumption of TV programming and production of social media by TV viewers (Fossen and Schweidel 2017; Viswanathan et al., forthcoming).

• RQ8: To what extent does the ‘what’ in the context matter? Many studies on context effects have focused on thematic (content) congruence (e.g. Chun et al. 2014; Dahlén et al. 2008; Moorman et al. 2012; Segev, Wang, and Fernandes 2014), for example, placing car ads in programming about cars. Another recurring idea in the literature on congruence effects is that a context that is congruent with the ad can make the comprehension of the ad easier (e.g. Goodstein 1993) and lead to more positive responses (Moorman et al. 2002; Dahlén 2005); however, Dahlén et al. (2008) postulate that incongruence may be more effective. The congruence effect may depend on the medium. Neijens and Bronner (2006) show that the interaction between the program and the ad is the weakest for television and cinema, while the majority of the previous studies focused on magazines. Also, the effects may depend on divided attention (Janssens, De Pelsmacker, and Geuens 2012) and differ for memory, attitude or intention measures due to different mechanisms responsible for the effects. While incongruence may work better for memory, because it makes the ad stand out and hence attracts attention and stimulates elaboration, leading to creating stronger associations (i.e. memories), congruence may be processed more heuristically, for example, via association transfer, leading to more positive attitudes. In addition, we do not know how strong the effect of context is and how long it lasts.

• RQ9: While digital display programmatic media has been typically used as a lower funnel, direct response program, TV advertising has, historically, been used more broadly across the customer journey, from upper funnel branding all the way to direct response. In which parts of the funnel should programmatic advertising be used?

• RQ10: How to create versions of TV ads at scale informed by data? Malthouse and Elsner (2006) discuss possible approaches.

Discussion

This paper contributes to the advertising literature in several ways. It discusses the current state and the future of TV ad models and explores how we can improve programmatic TV advertising outcomes. It proposes and discusses a framework for thinking about the different data-sets used in programmatic TV advertising, which includes four types of data available to build programmatic models: HH past behavioural data, HH characteristics, HH media consumption and context. Finally, it points out needs for future research.

Implications for the advertising field

Programmatic advertising has largely been developed outside of the academic advertising community, leaving this field to tech companies and computer scientists, which created a disciplinary gap between them and advertising and marketing scientists (Yang et al. 2017). The two communities should interact to, on the one hand, leverage prior advertising theories and models, and on the other, improve the analytical models. Advertising associations like the European and American Advertising Academies (EAA and AAA) should promote collaborations with computer scientists and the tech community at their conferences (e.g. plenary session speakers, special sessions) and in their journals (e.g. with special issues, editorials). Advertising curricula should cover both approaches.

Programmatic advertising also brings up some problems to look out for. First, in online display advertising media context is often ignored. In some extreme situations ads are displayed next to aggressive content that lead to scandals and damaged brand reputations. This falls under the heading of ‘brand safety.’ For example, thousands of brands stopped advertising on Breitbart.com after having ads placed on the site by programmatic algorithms. One remedy for this problem is to create private ad exchanges. Second, programmatic online displays led to fake traffic and ad fraud (e.g. Fulgoni and Lipsman 2017). We should investigate such unintended consequences in order to discover if they can also apply to addressable TV advertising and how we can avoid them with, for example, compensation methods that do not create an incentive for fraud.

Finally, we should not forget about ethical and privacy issues. Programmatic advertising requires merging data-sets coming from different sources, which requires transparent procedures. It also gives the advertisers possibilities to target their audiences better. Currently advertisers can version messages at the household level, but in the future they may be able to personalize them on individual basis. This raises serious ethical and privacy issues (see Boerman, Kruikemeier, and Zuiderveen Borgesius 2017 for discussion). Facebook's stock price dropped after the Cambridge Analytica episode, suggesting that firms may suffer a financial penalty for sharing data. The data ecosystem in Figure 3 envisions how a firm can create a 360-degree view of its customers to improve advertising effectiveness. While such a thorough understanding of the customer may produce short-term gains, overly intrusive, creepy or exploitive tactics may backfire in the long run, prompting a consumer backlash.

Practical implications

There are several obstacles on the way to fully programmatic TV advertising. First, as discussed above, there is no one data source, which means that companies need to trade data or work together. Until the industry comes together, the advertisers need to make choices which data to go for. That means that the long-established business models and relationships will need to change. There are already initiatives from TiVo Research and Analytics, Inc. or Nielsen Catalina Solutions that offer advertisers more data types for their models, and our prediction is that this trend will keep on developing.

Programmatic developments also affect TV measurement methods (see Fulgoni and Lipsman 2017 for discussion). As video moves OTT, the way reach is measured also changes in line with what is standard with online measurement, whereas TV measurement focuses more on potential than actual reach. In addition, multi-screening makes measuring reach even more difficult. Hence, advertisers need consider more cross-platform measurement to evaluate their campaigns.

Programmatic advertising will also greatly influence the future direction of media companies who rely heavily on advertising revenue to sustain their enterprises. Media companies must align themselves with appropriate distribution models to be fully accessible where/when/how their audiences wish to engage with them. These distribution models must also provide the technological means to deliver more targeted experiences for advertisers: both in terms of selecting the right audience and the right program environment for the advertiser's message. Media companies will need to understand how their audience is being measured and valued by advertisers.

Media companies invest significant resources to develop and acquire content that will attract the audiences that advertisers want. If the content, itself, is not valued properly in the programmatic advertising equation media companies stand to lose great sums of money and we should then become concerned that we could, potentially, destabilize the production and distribution of high quality TV content. This situation will force media companies to find alternate means of subsidizing high-quality content which could, potentially, pass more of the cost onto consumers.

Final thoughts

Both mass and computational are forms of paid advertising, where the advertiser pays for access to an audience attracted by some third-party media product. The consumer wants to view the media product, and advertising subsidizes the viewing. A critical problem with both mass and computational paid approaches is that consumers, for the most part, do not want to see ads and have new ways to avoid them, for example, ad skipping with DVRs or ad blockers. They can also avoid ads entirely with a growing set of subscription services such as Netflix, HBO Go, Amazon Prime Video and YouTube Red (or Spotify, Pandora, etc. in place of radio). The implications are that (1) advertisers will have to increase their use of other marketing communication channels such as owned media, branded content, product placement and events; and (2) content creators will have to rely more heavily on subscription or micropayment fees of all kinds, and other non-advertising revenue streams in the future.

Acknowledgments

The authors would like to thank professors Jim Lecinski and Tom Collinger of Northwestern University for their feedback and support.

Disclosure statement

No potential conflict of interest was reported by the authors.

Additional information

Notes on contributors

Edward C. Malthouse

Edward C. Malthouse is sills professor, Integrated Marketing Communications, Northwestern University and a research fellow at the Media Management Center, a partnership between Medill and Kellogg. His research interests center on media marketing, database marketing, advertising, new media and integrated marketing communications. He develops statistical models and applies them to large data sets of consumer information to help managers make marketing decisions. Malthouse is also currently the co-editor of ‘Medill on Media Engagement.’ He was the co-editor of the Journal of Interactive Marketing from 2005–2011. His professional experience includes software engineering for AT&T Laboratories, corporate analytics training for Accenture, BNSF, Digitas, Nuoqi and Capital One, and developing segmentations for Cohorts and Financial Cohorts and Motorola.

Ewa Maslowska

Ewa Maslowska is an assistant professor in digital analytics and consumer behavior at the Amsterdam School of Communication Research. Trained as a psychologist and communication scientist, she conducts experimental and data-driven research to understand the effectiveness of marketing communication on consumer behavior in digital environments (e.g., eWOM, branded content diffusion, consumer decision making). Her current research and teaching interests include digital analytics, personalized marketing communication, and brand engagement.

Judy U. Franks

Judy U. Franks joined the Medill IMC faculty in 2008 following a 23 year career in Chicago's leading ad agencies, where she rose to the executive ranks across both the media and creative strategy disciplines. Franks teaches both consumer insight and media courses in the undergraduate and graduate IMC programs at Medill. She has been honored four times by the undergraduate student body at Northwestern for teaching excellence and in 2016 she was named IMC Teacher of the Year by the MSIMC graduating class. Judy is the author of ‘Media: From Chaos to Clarity’ – a text that is used in both industry and academia to provide a framework for changes in the media landscape. She is also a contributing author to the text, ‘The New Advertising’ (Praeger, 2016). Judy's research interests span consumer receptivity to advertising formats in new media contexts, media economics and transmedia storytelling.