Abstract
Theories explaining the impacts of online media often swing between the actions of empowered individuals and the distribution structures put in place by powerful corporations. To explicate how these factors interact, we adapt the concept of audience flow to highlight the temporal dimension of web use and demonstrate how digital architectures subtly nudge masses of people into online attention flows. We identify sequential usage patterns through a network analysis of passively measured clickstreams, combined with data on website ownership and website architectures. Our sample, based on a panel of 1 million users, includes 1761 websites that reached at least 1% of Internet users in the United States. Our findings reveal previously unseen patterns of online audience formation, which have implications for studying media effects and understanding institutional power on the Internet.
Keywords
Media use can be conceptualized as a flow in which people encounter a sequence of offerings. This way of thinking has provided academics and industry professionals with a powerful framework for understanding and controlling audience formation. But with the advent of on-demand digital media, the notion that people’s media use could be directed by media organizations fell into disfavor. In its place arose a rhetoric of user empowerment in which people controlled the time and substance of their media encounters. The vision of newly liberated users has, in turn, contributed to the appeal of theories that see the actions of purposeful individuals as the key to understanding online behavior. We argue that this simple notion of empowerment is an illusion. On the web, media architectures still shape the flow of public attention. This happens in subtle ways that nudge users in particular directions. It often takes advantage of habitual behaviors and is generally difficult for the users themselves to see or understand.
We demonstrate these patterns of sequential Internet use by analyzing passively measured traffic data from a million Internet users on all 1761 popular websites that reached at least 1% of Internet users in the United States. In doing so, we adapt the concept of flow, advanced by researchers of television and radio including Raymond Williams, to digital media studies and develop a new way to identify online attention flows through a network analysis of clickstream data. Drawing also on website-level data about corporate ownership and digital architectures, our research reveals previously unseen structures which have implications for understanding institutional power on the Internet.
How audiences form online
Most theories of media use are tilted in favor of the individual agent (Webster, 2014). Economists and sociologists have long assumed that genre preferences determined people’s media choices (Owen and Wildman, 1992; Peterson, 1992). Uses and gratifications, a dominant paradigm in communication, maintain its commitments to individual gratifications sought as determinants of use (Sundar and Limperos, 2013). Likewise, research on selective exposure (Stroud, 2010), mood management (Zillmann, 2000), and the role of habits (LaRose, 2010) rely almost exclusively on psychologically based theories as antecedents of media choice. These assumptions about media audiences manifest in many subfields of mass communication research. In political communication, for example, the selective exposure thesis posits that with increased media options, people are empowered to choose attitude-consistent information and avoid counter-attitudinal content. Thus, for over a decade, many academics have expected patterns of attention to coalesce into partisan echo chambers (Jamieson and Cappella, 2008; Stroud, 2010). Political communication is not unique. A kind of consumer sovereignty reigns supreme in many theories of media choice.
Political economists have been skeptical of such sovereignty, arguing that consumer choice, and even functioning democracy, were constricted by concentrated media ownership (Downing, 2011). But even analysts sensitive to questions of institutional power were seduced by the newly liberating potential of the Internet and its “nearly free” digital affordances (e.g. Benkler, 2006). Only recently have political economists brought the analysis of distribution to bear on thinking through online audience formation. That research argues that Internet distribution constitutes a costly and powerful structure that corporations deploy to “underpin our communicative encounters” (Murdock, 2018: 365; Sandvig, 2015).
In The Internet Trap, Matthew Hindman (2018) providels a detailed account of how a handful of tech giants have succeeded in dominating Internet traffic. These companies invest massively in hardware and software (i.e. server farms and code) to overtake their lesser competitors in delivery speed and reliability. Based on large-scale field experiments, they are able to optimize user experience through tinkering their websites. Their unmatched marketing resources in the form of paid search and promoted social media feeds contribute to their dominance in online placement. Platform studies have added insights to the political economy of the Internet by examining the interface and entwined functionalities of websites and apps. For example, studies show how platforms such as Facebook and Google manage to trap users through intricate coordination of design and partnerships (e.g. Helmond, 2015; Van Dijck et al., 2018). This focus on the materiality of distribution to think about Internet use effectively brings in more critical purchase than the common wisdom that on the web “the big gets bigger” because of “network effects,” a purported mechanism wherein existing users always bring in more users through their social connections.
In sum, most of the early social science on the uses and effects of online media assumed an empowered, active audience able to choose whatever media they want when they want it. In contrast, political economists assumed that people are being acted upon by powerful, often unseen distribution infrastructures. In the light of this binary, an emerging literature began to theorize that exposure resulted from individuals interacting with institutional structures (Webster, 2014). For example, building on a line of work that stressed hybrid organizational and technological logics as integral to patterns of political communication online (e.g. Chadwick, 2013), Thorson and Wells (2016) laid out the varied influences exerted by actors such legacy media institutions, social networks, and computer algorithms over media exposure and effects we observe today.
This article extends this growing literature in two ways. First, we develop an analytical framework on audience formation that accounts for both individual user agency and structural power. We do this by invoking the concept of flow from television studies, adapting it to online media, and synthesizing it with existing conceptual efforts. Second, by employing novel datasets and analytical methods, our study yields systematic empirical findings that provide some clarity on what people encounter online and how digital distribution actually affects it. The scope of our analysis and conceptualization, importantly, addresses a wide realm of online activities beyond content consumption.
Flows in linear media
Television and radio are experienced as a flow of encounters arranged by a programmer. In the aggregate, these individual flows are seen by institutions as audiences flowing from one program to the next. In this study, “flow” refers to sequential behaviors enacted by groups of people. It is a macro-level phenomenon concerned with how public attention is directed to media content, often planned by media institutions.
This way of understanding media use has been popular since the dawn of broadcasting. Linear television delivered a fixed schedule of offerings in a stream that encouraged audiences to flow from one program to the next. From the earliest days of TV, programmers and marketers realized that there were stable patterns of audience flow, called lead-in or inheritance effects, which could help them to build and manage viewership (Eastman and Ferguson, 2013; Webster, 2006). Empiricists noted these regularities and posited laws about the “duplication of viewing” between pairs of programs (Cooper, 1996; Goodhardt et al., 1987). Cultural theorists also recognized that a “planned flow, is . . . the defining characteristic of broadcasting, simultaneously as a technology and as a cultural form” (Williams, 1975: 86).
Stable patterns of viewing, however, do not mean that audiences are passive dolts. Inheritance effects are typically the strongest when programs scheduled back to back were of the same genre (Webster, 2006). Audiences actively expressed their program preferences, but they did so within the temporal constraints imposed by the fixed schedules of linear television. By the end of the 20th century, however, everything seemed to change (Balnaves et al., 2011). With the advent of increased competition from cable and satellites, and ultimately the explosion of non-linear media on the Internet, the concept of a planned flow lost much of its appeal.
In 2008, The New York Times magazine cited industry experts who claimed that a planned flow was a thing of the past. “Everyone’s composing their own flow. And once you start becoming the composer of your own flow, you can’t go back” (quoted in Webster, 2014: 65). The so-called death of flow reflected the familiar belief that online media empowered users. Media could no longer be pushed at people. Instead, people pulled what they wanted when they wanted it. As Pariser and Helsper (2011) noted, “internet enthusiasts were excited about the shift from push to pull for reasons that are now pretty obvious . . . pull media put users in control. The problem is that pull is actually a lot of work” (p. 67). Indeed, a growing body of evidence suggests that the rumors of flow’s demise are, in the words of Mark Twain, “greatly exaggerated.” And there is reason to believe that the Internet allows institutions to build and manage audiences by structuring patterns of flow. Institutions may shape people’s media encounters by “nudging” them in particular directions, often without their awareness.
Flows in online media
Since its inception, analysts have seen the Internet as “anti-television,” because it provides distribution and transmission technologies that enable individual freedom of choice, lateral connections, and user-generated content. These features seemed to break media monopolies (Sandvig, 2015). Unlike television viewers who were encouraged to follow a preordained sequence, web users were free to explore an endless “cyberspace.” Although some critics argued that there was never any “space in cyberspace” (Manovich, 2001: 253), the concept is largely taken for granted in today’s popular imaginations (Mansell, 2012). The idea of this open, largely unstructured, space has encouraged an inaccurate image of the sovereign user composing her own journeys through a vast online landscape.
In contrast, seeing the Internet as a temporal medium is productive both for enabling closer inquiries into actual web usage and for rethinking the much-touted user agency being enacted in the metaphorical “cyberspace.” Media theorists have noted that, both phenomenologically speaking and in terms of the underlying technical operations, user experiences online are temporally rather than spatially organized (Chun, 2006; McPherson, 2006). McPherson (2006: 243) uses “volitional mobility” to describe the sensation of web surfing, during which the user “mov[es] from link to link with a certain illusion of volition.” In this sense, the Internet “emulat[es] the televisual event,” and the web user’s experiential sequences can be understood as flows.
Just as television audiences flow from program to program, Internet users visit websites sequentially. Each day, millions of people land on NYTimes.com from Facebook, and numerous individuals begin with Yahoo.com, their default homepage, followed by reading pages from Yahoo News. Each such sequence constitutes a flow, which empirically is the volume of people whose attention moves successively across the particular web outlets involved. To adapt to the context of web use, we use the term “attention flow” instead of “audience flow” that television scholars employ.
Importantly, invoking the concept of the flow to study web use redirects scholarly attention toward the role of institutional arrangements in shaping the sequencing of attention (Chun, 2006: 49). For example, remarking on the “increasing popularity of ‘portal’ sites” such as MSN and AOL in the early days of the web, McPherson (2006) cautions against the ascendance of “a Web architecture which works to constrict the surfer’s movement, effectively detouring users along particular paths or containing them within particular sites” (p. 248).
Nudges and online choice architectures
What sets online attention flows apart from their televisual predecessors, however, is the user’s heightened feeling of choice. It is useful to think of the shaping of online flows as “nudges”—that is, “liberty-preserving approaches that steer people in particular directions, but that also allow them to go their own way” (Sunstein, 2014: 358). 1 The literature of behavioral economics has found that “choice architectures,” or the characteristics of the environment in which options are presented, powerfully influence people’s decision-making. For example, people tend to go with whatever is set as the default option. Importantly, nudges are not manipulations imposed on user choice, but part of the context inevitably present when people choose (Thaler and Sunstein, 2009).
We might then think about “online choice architectures” that maintain users’ sense of freedom, while subtly nudging their attention. Within platforms, nudges might emulate linear media by engineering a site’s “stickiness.” For example, YouTube and Netflix have auto-play features and produce behaviors much like television inheritance effects (Sandvig, 2015). But the repertoire of choice architectures online is much larger. Many nudges span outlets. These include various embeddings of hyperlinks, pathways generated based on real-time data (e.g. social media customization, content recommendation, and search results) or on some design decisions (e.g. a gateway to an online form), software default settings (e.g. default homepages and search engines), and technical properties largely imperceptible to the users (e.g. low-browsing speed). In theory, a user is being nudged if she behaves differently in one set of choice architectures versus in another.
In an attempt to bridge traditional media use and effects research and political economy of the Internet, our research program on online audience formation seeks to identify “institutional nudges,” by which we mean nudges performed by institutions through well-thought through, coordinated website designs. These nudges thus manifest in observable patterns of web use that could be systematically explained at scale by institutional intentions.
Notably, the notion of nudge is more expansive than that of curation, a key phrase employed in the burgeoning scholarship on content platforms. As Thorson and Wells (2016) write, to curate is to select, filter, organize, annotate, and frame content, which becomes increasingly essential for people to “efficiently” cope with the information abundance of our age. The notion of nudge that we employ here is broader; it is not limited to content consumption, but extends to other activities people conduct online, such as banking and checking maps.
Three approaches to analyzing flows
How can we study temporal, unfolding experiences of web users in a way that reveals the presence of online choice architectures? Broadly speaking, we can identify three different approaches that researchers have adopted to analyzing flows.
The first focuses on texts, wherein flows are revealed by analyzing the temporal sequence of content delivered by one curating actor (e.g. a channel, news outlet, or social media platform). This was the approach taken by Raymond Williams (1975) in his classic work on television. He demonstrated his method by describing the flow of content on a channel in evermore granular detail, proceeding from programs and narratives to sequences of sounds and images. Williams and others using this approach (e.g. Caputi, 1991) infer how these “flow texts” influence users. This is a micro-level form of analysis often grounded in cultural theory. More recent work theorizing “curated flows” (Thorson and Wells, 2016) seems in keeping with this approach to flows, while featuring the influences of social networks and algorithmic renderings on the user’s exposure to flow texts. All are “user-centric” in that the individual user’s experience of flow is the center of attention. In a world of anywhere-anytime media, it is increasingly difficult to claim which flow texts are idiosyncratic or common, and in turn, extrapolate these micro-level flows to larger social phenomena.
The second approach is concerned with “flow culture” (Bolaño, 2015; Flichy, 1980), and pitched at a higher level of analysis. It focuses on media organizations, which produce and distribute culture and the markets within which they operate. Analyzing these institutional arrangements, researchers have demonstrated how the flow of culture is controlled and made inferences as to its ideological effects (e.g. Kreiss, 2016). Often grounded in the political economy of media, such work can be thought of as a “media-centric” approach since the experience of individuals is of less interest than the work of media organizations and markets. As is the case with user-centric approaches, it is increasingly difficult for media-centric work to make the leap from institutional arrangements to a clear picture of how choice architectures affect the distribution of public attention. Given that platforms now control the flow of audience to and from other media, a static look at organizational affiliations tends to miss the power dynamics at play in digital media.
The third approach focuses on the extent to which audiences flow through sequences of content. This is in keeping with the work of empiricists we noted earlier, who have documented law-like patterns of audience behavior. Academic work of this sort has been done in marketing and communications. Typically analyzing passively observed media use data in order to model macro-level patterns of exposure, it can be thought of as an “audience-centric” approach to analyzing flows (e.g. Webster and Ksiazek, 2012). It is sometimes done for applied purposes (e.g. Rust and Eechambadi, 1989), with few overt theoretical aspirations other than understanding the forces that shape mass behavior. The strength of applying an audience-centric approach to non-linear media is that it can document the actual consumption of online culture. To date, this literature has concentrated on analyzing snapshots of which outlets share audiences. Fortunately, we elucidate next that limitation is not inherent in audience-centric analysis.
Sequential attention flows
Just as studies of linear media identified how audiences flowed from one program to the next, it is possible to capture the dynamics of attention flows across online offerings. Understanding how these flows shape public attention can offer new insights into the origins and likely effects of online media consumption. Consider for instance two partisan outlets MSNBC and Fox News. Studies show high audience duplication between their websites (Gentzkow and Shapiro, 2011). Different pathways may lead people to consume ideologically divergent outlets. Some people may land on the websites through their friends’ recommendation on social media or top-ranked results by search engine, while others, hoping to “check out the other side,” may embark on these outlets using URLs. Discerning these processes requires a temporal dimension that analysis of audience duplication lacks (Möller et al., 2019).
Using clickstream data, we propose a “sequential” audience-centric approach to overcome this methodological limitation. Clickstream data capture user activity as people move from browsing one webpage to another (Wu and Ackland, 2014). These attention flows, as we have conceptualized, are invariably shaped by online choice architectures within which they occur, and our research program aims to identify certain forms of nudges that align with institutional intent. We operationalize the scale of such intents through shared website ownership and traits of website architectures. In this vein, we pose the following research questions:
RQ1: What patterns of attention flows emerge from analyzing people’s sequential browsing activity?
RQ2: In what ways are these patterns of attention flows related to online choice architectures, especially ones that Internet corporations use for audience building?
Method
To address these questions, we obtained passively measured national-level clickstream traffic data, together with ownership information and “website category” of the web outlets from Comscore. We sampled all websites that had at least 1% (2.6 million) of all web users in the United States visiting in October 2015, resulting in a sample of 1761 web outlets. To determine this sample, we first downloaded the entire list of 33,000 web entities measured by Comscore in October 2015. Comscore nests web entities hierarchically based on ownership. Consider Google for example. Comscore aggregates a collection of all Google-owned entities as “Google Sites” (level 1), which can be disaggregated into Google.com and YouTube.com (both at level 2). Google.com can be further expanded to obtain traffic separately for “Google Search,” “Gmail,” and “Google News” (level 3). Based on our framework, Google Search represents a distinct online choice architecture compared to Google News. Thus, to obtain high granularity required to test our conceptual framework, we disaggregated web entities up to the third level and selected all level 3 outlets that met our threshold of 1% unique users. We report this entire sample in a supplemental appendix (Supplemental Appendix 3).
In addition to website ownership data, we conducted at different stages website-level observations both to inform our quantitative analysis of web use data and to interpret its results in the light of our conceptual framework. More detail on this qualitative component is provided below as we delineate our analytical steps.
We extracted the clickstream data for each web outlet. Reported for a specific time period, these reflect the number of total unique users that landed on an outlet immediately after visiting other outlets during this period. In clickstream data, if a user switches from outlet a to b and then from b to c, outlet a is a traffic source for b, and b a traffic source for c. An aggregation of all these clickstreams in a media ecosystem would enable the researcher to infer, through network analysis, attention flows at scale. Since our sample was both large and expansive, we had the complete list of the sources of incoming traffic for most websites. Utilizing these clickstreams from all websites, we created a matrix of websites where each cell indicates the user volume flowing from one site to another. In effect, unlike undirected audience duplication networks (Webster and Ksiazek, 2012), this can be analyzed as a directed network of websites, with edges being user volumes flowing from one website to another.
For this network, we first estimated each node’s weighted outdegrees and in-degrees. The weighted outdegree in a clickstream network represents the total volume of user traffic that flows out of this site to all other sites in the network. The network was highly centralized, with highly skewed distributions of both indegrees and outdegrees but the latter being especially concentrated (Gini coefficient = 0.64). Thus, a handful of outlets with exceedingly large outdegree scores serve as the sources of traffic flows to large portions of the web (see Supplemental Appendix 3). Thus, the graph has an inherent core-periphery structure obviating the need to filter insignificant ties. These are the most popular email providers, portals, search engines, and social networks. Based on their large outdegree scores, we infer that user visits to most other websites originate from these sites.
Results
Patterns of attention flows: clickstream constellations
To address RQ1, we implemented a cluster analysis using a modularity based method for detecting communities with dense interconnections (Blondel et al., 2008). In this context, such a community/cluster represents a socially shared browsing sequence, with several website pairs accessed in succession by users. We refer to these clusters as “constellations.” Other than being mathematically acceptable, a cluster solution coheres with our theoretical expectations if, for most constellations, the constituent sites either exhibit some content similarity (a proxy for user preference) or point to some mechanism explained by online choice architectures. The solution we arrived at has 14 clusters, with a modularity score of 0.25. Since the modularity of a randomly generated network is 0, this score indicates a good fit. Figure 1 shows the 14 constellations of our solution. For further analysis, we left out the three smallest clusters since they had a small number of highly specialized websites.
14 clickstream constellations.
For each constellation/cluster, based on the weighted-outdegree scores of all its constituent websites, we calculated a Gini coefficient (see Table 1). This indicates the extent to which a browsing sequence is “anchored” by a handful of sites that serve as beginning and returning points, preceding visits to most others. Given usual browsing patterns, people more likely move from anchor sites to the subsequent site through built-in pathways such as hyperlinks instead of entering the URLs in their browser. If a small number of sites anchor the browsing sequence, the Gini coefficient tends to be high. The power of anchors is also reflected in the proportion of their cumulative outdegrees in the constellation’s total outdegrees (Table 1).
Clickstream formation: distribution and shared ownerships.
For example, in row 1, the 2 anchors Bing & MSN Content Sites, accounted for 59% of all outdegrees for this constellation.
For example, in row 1, 19 outlets (95%) owned by “Microsoft Sites,” belonged to constellation 1.
Next, to qualitatively examine each constellation’s composition, we first ranked its websites by weighted outdegrees. Then beginning with the anchor(s), we noted functions and brands of all the outlets that account for 80% of the constellations’ total outdegrees. If a constellation had less than 20 outlets meeting this criterion, we further went down the list to examine at least 20 outlets per constellation. In doing so, we examined 328 outlets. We also investigated the existence of historical partnerships between anchors that the ownership data do not directly provide. For each constellation, we combine quantitative results with qualitative examination to discern its likely mechanism.
Google complex
This cluster has the largest number of sites. It mainly contains utilitarian websites such as those of retailers (including eBay and Amazon) along with websites of service providers in domains such as finance, shopping, telecom, travel as well as government. What these sites have in common is that users are and are unlikely to encounter them as part of random browsing, as they visit them to accomplish a specific purpose. With the highest outdegree, Google Search anchors most of these visits. Other prominent anchors in this cluster include YouTube and Gmail; their high outdegrees suggest that users return to them intermittently during typical browsing sessions.
Social network complex
This cluster has major online social networks such as Facebook, Twitter, and LinkedIn. Alongside them are a large number of sites providing content on current affairs, politics, entertainment, and specialized topics such as sport. These sites are run by legacy media organizations (e.g. Washington Post) and digital companies alike (e.g. BuzzFeed). “Socially-driven” outlets such as GoFundMe, Spotify, Fitbit, and Legacy.com constitute the third type of sites in this cluster. Finally, sites of various banks and mobile/ISPs are also present here, which may be because when people go to these long-term services, they need not resort to search engines. Instead, they access them directly by entering the URL or clicking bookmarks, most likely during their social media time.
Yahoo homepages
A number of Yahoo’s flagship online properties (e.g. Yahoo homepage, Yahoo News, Yahoo Sports, and Yahoo Mail) anchor this somewhat smaller constellation. Also present are many outlets specialized in business news (Bloomberg and Business Insider), political news commentaries (e.g. The Atlantic, Dailybest, The Hill, Politico, Slate, Vox.com), and general “soft” news (e.g. HuffPost and USA Today). Classmates.com, which helps search for past high school friends, and online services for investments and mortgages are other notable inclusions in this constellation.
Yahoo search
Along with the anchor Yahoo Search, many utilitarian websites that originated in the 1990s and serve a specific purpose (e.g. Ask.com, eHow, MapQuest, and WebMD) constitute this cluster. This cluster probably captures browsing routines of a segment of older web users that continue to rely on Yahoo Search. Interestingly, included here are also multiple smaller search engines, which when examined all turned out to be powered by Yahoo Search but with different frontends.
Porn constellation
Adult sites primarily constitute this cluster. Among them Pornhub.com has the highest outdegree and thus serves as a gateway to other adult websites. Pornhub’s homepage links to several of these (e.g. Youporn and Xtube), which are part of the “Pornhub Network.” Several other adult sites that are part of this cluster appear to utilize adult advertising networks to feature their content on the network of Pornhub sites.
Bing/Microsoft
With the highest Gini, this constellation consists of several content portals and web services aligned with or owned by Microsoft, including AccuWeather, Office, and Windows. Microsoft’s search engine Bing is the anchor. Relatedly, AccuWeather is Bing’s default weather widget.
User data solicitors
The websites clustering here attempt to extract personal data from users for commercial purposes. Through adverts or pop-ups on other websites, these sites lure users to a variety of online questionnaires with the promise of product discounts and other rewards. Bundled together are also websites run by market research companies, who are in charge of survey design and data analytics of solicited responses.
Job search
Indeed.com, Oracle, and other popular portals for job applications and human resource management constitute this constellation, along with websites of companies that manage databases for job portals.
AOL homepages
Anchored by “AOL Homepages,” this constellation contains a few AOL websites (e.g. AOL email) along with other sites that are part of the AOL brand, including USMagazine, Everyday Health, and ZergNet.
Travel
This cluster has airline and hotel websites, along with third-party sites specialized in booking travel packages, flights, and hotels. Anchors include Tripadvisor, Expedia, and Priceline.
Citibank retailers
K-Mart, Macy, Old Navy, Sears, and other retailers constitute this cluster along with Citibank, which manages the credit card programs for all these stores.
Attention flows, audience building, and online choice architectures
Shared ownerships by constellation
To address RQ2, we first associated shared ownership of web outlets with their shared constellation membership and found a clear pattern. Of the 1761 web outlets, over a third (653, 37%) were controlled by multi-site owners. For example, Google.com and YouTube.com shared an owner (“Google sites”), whereas Hilton.com did not share ownership with any other sites. Notably, 54% of the 5468 website-pairs that shared owners also belonged to the same cluster; by comparison only 20% of the 3 million other pairs (that did not share owners) clustered together. Table 1 further shows that for the majority of the clusters, most web outlets owned by a company clustered together. Importantly, the anchoring websites that we identified earlier tend to be the flagship sites of major Internet companies, and the corresponding clusters often involved other websites owned by the same companies. As discussed earlier, anchors in high-Gini constellations are more influential in shaping the browsing sequences within each constellation.
Our qualitative investigation also revealed the impact of partnerships on online audience building. The presence of AOL-owned sites in the two Yahoo constellations, in particular, can be explained by the two companies’ prevailing content partnership when our data were collected. Using audience duplication data, we conducted a separate analysis (detailed in Supplemental Appendix 1) on how different browsing sequences were enacted by the same or different audiences. The results confirmed our observation.
Website types by constellation
By examining the types of websites that constitute each clickstream constellation, we can shed further light on how various choice architectures are deployed to nudge people online. To infer these forms of institutional nudging, we combined Comscore’s “website categories” into the following five types: search, social media, media content, software/hardware, and e-commerce/service (see Supplemental Appendix 2 for details on this recoding). We found significant associations between these website types and constellation membership in Figure 2 (χ2 = 444.11, df = 52, p-value < .001). Our approach is to view these data as website-level proxies for online choice architectures. A more definitive reckoning of the uses and effects of nudging architectures would require measurement at the level of individual webpages (e.g. Möller et al., 2019) or even code.
Distribution of online choice architectures in each constellation.
Nudging attention flows on the web
In this section, we organize the mechanisms that nudge online attention flows across websites into four distinct forms, each associated with particular website architectures. We arrived at these associations based on the preceding quantitative and qualitative investigations of constellation distribution and anchors, as well as their relations to shared ownership and website types. In Table 2, we arranged these forms of nudging by their visibility to the user, an arrangement that draws inspiration from critical algorithm studies that the power of digital architectures comes much from the lack of user awareness. While this typology of institutional nudges emerges from our empirical study (see Table 2 for their associated website types and the clickstream constellations whose formations appear to be driven by the corresponding nudges), we present this typology also as a heuristic provocation for future empirical research, which we elaborate on in the conclusion.
Forms of institutional nudge.
The first form of nudges occurs through visible features on the web page. This includes ranked recommendations on search engines and content curation on social media. As Figure 2 indicates, different websites are part of social media-anchored clickstream constellations and clusters anchored by search engines. The latter are largely “e-commerce/service” websites for everyday services including banking, health consultations, maps and retail. Contrastingly, “media content” websites where information is an end in itself, providing news, commentaries, and entertainment content, constitute the former. In other words, search anchors users toward utilitarian, functional goods and social media toward symbolic, experiential goods. The utilitarian goods serve the needs of people fulfilling their mundane tasks for which social sharing is likely incidental. The symbolic goods seem more amenable to social sharing. Based on our results, it is reasonable to infer that social networks, including Facebook, largely tend to precede people’s visits on news information sites.
Although web users may not comprehend the algorithmic processes involved, they recognize that search engines and social media direct them to other parts of the web. In comparison, a somewhat less visible form of nudge occurs through “hypertexts” which, integrated within the page content, tend to link to websites with shared economic interests. High volumes of attention flow through “media content” suggest this arrangement for audience building. For example, nudging through hypertexts may explain that the Yahoo homepages constellation contains content offerings such as Yahoo News and Yahoo Finance. These less visible nudges also direct attention to unaffiliated websites with underlying capital connections to the brands such as ATT.net (Yahoo provides a customized portal homepage to Internet subscribers of AT&T). Another example is the “all-media content” Porn Constellation, where a closed maze formed by MindGeek-owned niche adult sites appears to confine sequential users’ flows. Our analysis suggests nudging through hypertexts operate in a no less powerful manner than platform curation. For example, we found Yahoo to remain enormously central, despite the fact that Silicon Valley has predicted its demise for over a decade. The online choice architectures which shepherd Yahoo users into an ecosystem of Yahoo and partner properties play a part in retaining its relevance. 2
A step further along the visibility spectrum, nudges occur through browser software configurations on the user end, such as default toolbars and homepages. For example, many “anchoring” portals serve as default homepages on certain browsers. Furthermore, that major search engines anchor distinct constellations may be due to bundled browser installations. Microsoft Windows browsers (e.g. Internet Explorer and Edge, which default to MSN as a homepage) has Microsoft Bing as the default engine for their search bar; Safari, the default browser for Apple Macs with “livepages.apple.com” as the default homepage, uses Google Search as the default. These ownership and partnership connections are evident in our findings about traffic patterns. In short, corporate software arrangements also aimed at growing user bases, nudge people to incorporate some of the most popular websites—oftentimes, anchoring sites for browsing sequences we observed—into their web use routines.
Many users may lack the initiative or technical know-how to confront and alter default software configurations, and in turn their nudges. A opaquer form of nudge, however, takes place at the “back-end,” which usually reflects partnerships between institutions. For example, the constellation Citibank retailer formed due to all the retailer sites having Citibank as their payment gateway. Likewise, the Job Search Constellation knit together because various job portals manage their data through common vendors such as Salesforce or Oracle.
In summary, while the use of a search engine requires some premeditation and intentionality, other anchoring sites ranging from social media, web portals, and contracted technical platforms on the back-end seem to direct traffic in ways that require little or no conscious planning or choosing on the part of the user. The corresponding online choice architectures involved are difficult for people to discern in increasing degrees. Institutions routinely manage online flows with integrative hypertexts, browser configurations, and back-end bundling, all of which are essentially invisible.
Conclusion
There is a growing consensus that cyberspace—a universe of digital resources that empower people to achieve their own ends—is more complicated than visionaries imagined. Not unlike the linear media of radio and television, the new purveyors of online media strive to manage people’s time and attention to suit themselves. To investigate what those new patterns of public attention look like, we used the concept of flow. Specifically, we investigated how the structural features of the Internet have affected the flow of public attention across and among websites. In doing so, we reclaimed an analytical framework that conceives of media use as a temporal flow and developed new ways of analyzing clickstream data. As a result, we discovered that managing the flow of online audiences goes beyond overt recommendations and website curation to include less visible mechanisms that nudge people. We found these relatively invisible forms of nudges to serve underlying institutional interests in audience building. Specifically, our study informs research on media effects, the nature of audience fragmentation, and the political economy of communication. We consider each of these in turn.
To begin, conceptualizing people’s online media use in this way can enrich our understanding of media exposure and its consequences. As we have noted, the empirical research of audience behavior, grounded in selective exposure, motivated reasoning, and uses and gratifications, has historically presumed an easy congruence of individual attitudes and message characteristics. Following recent efforts at conceptualization that incorporate institutional, technological, and social factors (Thorson and Wells, 2016), our empirical research program on attention flows provides analytical perspectives on environmental factors that shape what people see or hear. Supplementing existing work that combines individual behavioral logs with self-reported characteristics (e.g. Möller et al., 2019), we call for more empirical research on structural features beyond individual awareness and articulation. If people are subtly nudged in directions not dictated by their predispositions, over time it might be possible to cultivate attitudes and behaviors that would not otherwise exist. Similar complications arise in consideration of most media effects research.
Understanding how attention flows across outlets can also shed light on the formative processes underlying patterns of audience fragmentation (Peters and Schrøder, 2018), especially as they pertain to the debate on audience polarization. Drawing into question the extent of partisan polarization, recent studies of audience duplication between outlets have shown a good deal of cross-cutting exposure (Gentzkow and Shapiro, 2011; Taneja et al., 2017; Webster and Ksiazek, 2012). Yet, these barely scratch the surface of online attention flows and cannot explain how the same people might, for example, come across ideologically opposite websites such as Fox News and MSNBC.
Analysis of clickstream data as directed networks can identify what online choice architectures may have led the same user to such cross-cutting exposure. For instance, if these users were largely sent by search engines, it would suggest their conscious choice to check out “both sides.” Yet, as our findings show, social media play a bigger role than search engines in directing people to media content. It is thus possible that social media have some power to promote “mainstreaming” rather than creating “echo chambers.” In this vein, empirical analyses about pathways to content exposure might inform research on confirmation bias. Also, based on what we find, cross-cutting exposure could be channeled by nudges among economically entangled websites, which motivates the need to bring commercial factors into theorizing partisan audience fragmentation. Finally, we would like to emphasize the potential of examining attention flows between news websites and sites embodying choice architectures of various mechanisms. This would reveal formative processes behind fragmentation that remain hidden in studies focusing on audiences of news sites alone.
Furthermore, our work raises fundamental questions about the scope and nature of power relations on the Internet. Analyzing flows enable a fuller investigation of the role of institutional arrangements in shaping the online attention economy, thereby addressing the gap in the political economy literature of how audiences form in relation to the affordances of distribution. This is no small matter. Hindman (2018: 166–167) argues that individual choices online are potentially influenced by “[h]undreds of features on a site,” and that “tiny effects multiply with every user visit,” compounding exponentially to produce the macro-level structure of the web. Our results provide systematic evidence that online choice architectures do, in fact, nudge users. It seems plausible that these “tiny effects” will accumulate over time forming larger structures that channel flows of public attention. Although, subtle as nudges are, users will probably not notice. By calling attention to the economic resources and arrangements entailed in the construction and maintenance of these architectures, our study highlights the advantageous position large corporations have in competing in the online marketplace.
A research program on attention flows provides political economists, as well as those doing platform studies, with new tools to investigate their claims. For example, analysts argue that a handful of platforms have gained disproportionate power through calibrating interfaces for nested-applications and forming strategic partnerships for locking-in users (e.g. Van Dijck et al., 2018). They further argue that these mega-platforms are taking on “infrastructural properties”—that is, they function as shared, indispensable services like public utilities (e.g. Nieborg and Helmond, 2018). The focus on the “infrastructurization” of platforms does stimulate invaluable policy and ethical discussions concerning tech companies and their products (Plantin et al., 2018). However, while examining the business and technical dimensions of the platform illustrates its potential mechanisms for user growth and governance, these methods are invariably platform-centric and unable to address actual user engagement with and beyond the platform (for a review on the platform-centric episteme, see Wu and Taneja, 2020).
In contrast, investigating sequential user behavior across the web re-centers the analysis and may demonstrate more forcefully a platform’s “infrastructural characteristics” and its position in the broader media ecosystem. For instance, our study shows that the power of singular sites such as Facebook to “platformize the web” (e.g. Helmond, 2015) lacks support from the macro patterns of online attention flows. Instead, the political economy of the web in action is better illustrated through a complex cascade of pervasively distributed choice architectures that latently nudge attention across the web. Along the same line, our study showcases how data-extraction technologies (Turow and Couldry, 2018) operate in tandem with structures of digital distribution joined through ownership and partnerships. This perspective complements the existing scholarship that documents how these technologies are employed by mega-platforms to harvest data.
As a first step in this line of inquiry, our empirical study has limitations that future work may overcome. To begin with, as we discussed, our data are on the website-level. One could gain greater resolution on the nudging mechanisms by analyzing more granular, webpage-level traffic data, and digital architectural traits. Our dataset also does not isolate the analytics trackers that commercial actors employ to direct and monitor online user behavior. Incorporating this would significantly advance inquiries into the extractive nature of online architectures. Likewise, to provide more comprehensive empirical analysis of the political economy of online audience building, future research may incorporate rigorous data on partnership, in addition to ownership, as a covariate. Finally, our study does not account for the mobile environment, toward which more platform studies now gravitate. A similar approach, we suggest, can be adapted to examine the sequential usage of mobile apps in relation to institutional factors.
Digital media are colonizing every part of modern life. Although invaluable, they are not benign. The institutions that control them are serving their own agendas. Barring a massive regulatory intervention, the online choice architectures that shape how we use these resources will evolve and operate largely out of sight. These mechanisms subtly direct the flow of public attention to information, entertainment, and commerce. They are invisible arbiters in the marketplace of ideas and beyond. And as such, they can mitigate or exacerbate a variety of media effects, some of which go to the very heart of participatory democracies. Academics are uniquely positioned to reveal these structures and how they intervene in social life, but our theories and methods must better reflect our new realities. Only then can we fully realize the benefits of digital media.
Supplemental Material
Appendices – Supplemental material for Going with the flow: Nudging attention online
Supplemental material, Appendices for Going with the flow: Nudging attention online by Angela Xiao Wu, Harsh Taneja and James G Webster in New Media & Society
Footnotes
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
Supplemental material
Supplemental material for this article is available online.
Notes
Author biographies
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
