Social Media Audience Metrics as a New Form of TV Audience Measurement

Darryl Woodford, Ben Goldsmith, and Axel Bruns

The Uncertain Business of Audience Measurement

Understanding and acting on the behavior of media audiences is a multi-billion dollar

business. Broadcasters and other media providers, advertisers, advertising agencies, media

planners and audience research companies have significant financial stakes in the collection and

analysis of audience data. In addition, policy makers, academics, and audience members

themselves have interests in the technologies and methodologies used to measure audiences, as

well as in the data itself and the uses to which it is put. But the audience rating convention – the

necessary consensus among stakeholders about who and what is counted, how the counting is

done, how the data is interpreted and how it is valued – is under pressure as never before.

Digitization, media convergence and audience fragmentation have dramatically disrupted the

business of audience measurement. New metrics and analytical systems have been developed to

answer some of the questions raised by technological change, but they are also posing challenges

to stakeholders about their capacity to deal with the explosion of raw and customized data on

audience behavior. The volume of information that is available for aggregation and analysis has

grown enormously, but with that growth has come a host of uncertainties about audience

measurement, and in particular about the broadcast ratings system.

Uncertainty has driven an extraordinary research effort, a flight to accountability, in

which a proliferating number of information and research companies have tried to make sense of

the accumulating data about media use, often with conflicting results. This was one of the

reasons behind what Alan Wurtzel, President of Research at American broadcaster NBC, has

Woodford, Goldsmith, & Bruns 1

called the “crisis in measurement” (Wurtzel, 2009), although the wealth of data and the efforts

being made to analyze it may mean that this period could be looked back on as a golden age if

the industry’s ideal scenario – the collection, cross-tabulation and fusing of massive amounts of

data and large datasets – can be realized. This would potentially produce the advertising

industry’s holy grail: single source, or consumer-centric holistic measurement (WFA, 2008),

although serious questions would also arise, not least about privacy and audience members’ and

consumers’ awareness of the data that is being collected (Andrejevic, 2007, 2013).

In some respects, the current state of uncertainty is nothing new. Historically, the

introduction or expansion of commercial broadcasting services, changes to the structure,

economics, technologies, or the policy field of broadcasting, and evolving patterns of audience

behavior have all spurred the development of new technologies, methodologies and rationales for

quantifying television audiences. For various reasons primarily to do with establishing the

parameters for the buying and selling of airtime in predominantly commercial or mixed public

service/commercial broadcasting markets in countries around the world, consensus has tended to

form around the need for an authoritative, simple measure of exposure – who is watching

television, which channel or service are they watching, and for how long. There has long been

great (and recently, increasing) interest in measuring audience members’ engagement with

programming and advertising – how much attention they are paying, what their opinion is about

what they are watching, and what impact the program or commercial has on them – but exposure

has remained the standard for measuring broadcast ratings and the core of the ratings convention

ever since Archibald Crossley’s first survey of American radio listeners in 1929 (Balnaves,

O’Regan, & Goldsmith, 2011). Despite the contemporary crisis, which is multifaceted, ratings

data are still and will continue to be in demand because there will always be the need for

Woodford, Goldsmith, & Bruns 2

common currencies for buying and selling advertising and program content. There will

undoubtedly be changes in the practicalities of audience measurement, particularly given the

challenges presented by the likely spread of broadband-enabled set-top boxes, which have been

described by the Council for Research Excellence as the “wild west” of research (Council for

Research Excellence 2010, p.4).

The availability of multiple channels through subscription or free to air television,

coupled with ever-increasing online video options, amplifies viewer/consumer choice and

consequently distributes the available audience much more widely than earlier broadcasting

systems. Napoli (2003, p. 140) argues that this fragmentation increases the disparity between the

predicted and the measured audience and reduces the reliability of data collected in traditional

sample-based methods. It certainly increases what has been called the “research ask,” and

complicates the carefully calibrated equations that produce the ratings. Although mass audiences

can still be assured for certain major events, often live international sports championships,

audiences in general have dispersed. Content providers, advertisers and research organizations

have had to track not only timeshifting and catch-up TV, but also migration across platforms,

and even beyond the home. Audience fragmentation has precipitated proliferations of data,

methods, metrics and technologies that in turn have allowed samples of a few hundred, in panels

or diaries, to multiply into surveys of millions of subscribers and produce competing currencies.

Opportunities for advertisers to reach consumers through media and other touchpoints have

proliferated, while advertisers’ and content providers’ desire for solid numbers and discontent

with the prevailing currency and methods have opened spaces for research and analysis.

Public service broadcasters have typically been more interested than their commercial

counterparts in qualitative research that provides detailed information about audience enjoyment

Woodford, Goldsmith, & Bruns 3

and engagement with programming. This, for example, was the focus of audience research

conducted by and for the BBC from the mid-1930s (Silvey, 1951). For commercial broadcasters

– as, eventually, for public service broadcasters too – ratings have served a range of purposes,

from measuring the popularity of particular programs, providing guidance in program planning

and scheduling, informing service delivery, keeping abreast of change in audience tastes and

practices, and establishing the value of time sold for advertising. Ratings can also act as a proxy

for the broadcaster’s share price and an indicator of (and influence upon) its overall financial

health (Balnaves et al., 2011). For advertising agencies, media planners and advertisers

themselves, ratings help determine how much will be spent on advertising on a particular

channel or network, as well as where and when advertisements will be placed. But ratings are not

only used within broadcasting. They are also of interest to the mainstream media and the public

at large for what they appear to reveal about the success or otherwise of programs and

broadcasters, as well as to academics, media critics and public authorities who “use, quote,

debate and question” ratings (Bourdon & Méadel, 2014, p. 1). In Canada, for example, the media

regulator uses ratings as one measure to judge the success of CanCon (Canadian Content) drama

policies, and as the basis on which funding for future production is allocated (Savage & Sévigny,

2014). Criticism of the ratings has come from many quarters and taken many forms, from

theoretical and technical questioning of the methodologies and technologies deployed over time

to concerns about the business practices of data suppliers, and the tendency of those who use the

ratings to “endow the audience with a reality and thereness it does not possess” (Balnaves et al.

2011, p. 229). And yet, despite the disruption wrought by digitization, a variety of parties

continue to maintain a variety of interests in the collection of robust, reliable and commonly

agreed upon metrics about audiences, as well as in agreeing on what counts as an audience.

Woodford, Goldsmith, & Bruns 4

 As Alan Wurtzel observed of the situation in the US in 2009,

A couple of years ago, Nielsen delivered a single TV-rating data stream. Today, Nielsen
routinely delivers more than two dozen streams (yes, we counted them) and countless
more are available for any client willing to pull the data. Moreover, set-top boxes (STB),
moving closer and closer to second-by-second data, will produce a staggering amount of
new information. And, with internet and mobile metrics as well, it’s not the amount of
data that is the problem; it’s the quality and utility. (2009, p. 263)

This is the key challenge for ratings providers in the future: providing quality and useful

data. But given that so much is in flux, including common understandings of “quality” and

“utility,” it appears for the moment as though multiplication of research vehicles and

partnerships will inevitably continue as ratings companies jostle over currencies and

simultaneously provide bespoke services to individual clients.

In addition to Nielsen’s multiple streams and the wealth of other services available,

broadcasters, content providers and advertisers must also contend with the power of bottom-up

systems of recommendation and rating that have emerged with the Internet. From Facebook’s

“Like” option which allows readers to signify in a single click their approval or appreciation of

something posted by a friend (importantly, there is no option to “Dislike”) through sharing and

retweeting on Twitter to supporting (or Digg-ing) something posted to Digg.com, the

opportunities for audiences to register opinions or rate all sorts of things on the Internet are many

and varied. To varying degrees, research companies, advertisers and content providers are

realising the importance of social media in gauging audience opinions about the quality of

content. The characteristic online behavior of countless people now routinely involves what

futurist Mark Pesce (2006) calls “the three Fs:” finding, filtering and forwarding information

found online to contacts (or followers in Twitter-speak, friends on Facebook). In contrast to

more restricted media such as free-to-air broadcast television, audiences can now find desired

Woodford, Goldsmith, & Bruns 5

audiovisual content, or close approximations, online. The actions of tagging, rating and

recommending function as forms of feedback, often for the principal benefit of the audience’s

own network. But ever more sophisticated and insistent forms of monitoring behavior and

turning it into useful data are capturing this information, adding it to databases for dissection and

fusion. In terms of quality, audience members who follow or forward content on multiple media

are exactly the audiences that producers of media content are trying to cultivate, in part because

of the ratings they may provide in the future. It is the most committed, the most voracious of the

online explorers or pioneers, the keenest edge of the community that evolves around content,

who can shape the media choices of those around them, who will be most highly valued by

producers, if not always by advertisers. All of these developments point to the likelihood that

measures of popularity in social media will become more extensive in the future. In the

remainder of this chapter, we discuss the ways in which particular forms of social media analysis

can produce useful and actionable data about engagement with television that augment and

extend the ratings’ core focus on exposure.

Toward Social Media-Derived Audience Metrics

Traditional television ratings schemes provide a standardized and broadly reliable, but

ultimately limited and one-sided, measure of audience interest; historically, they provide

information on what audience research could readily and regularly quantify, but fail to offer any

fine-grained, in-depth evaluation of audience activities even from a quantitative perspective,

much less from a qualitative one. In the emerging multi-channel, multi-platform, multi-screen

environment, they become manifestly insufficient.

Audiences for televisual content now access their shows through a range of channels: in

addition to conventional reception of the live television broadcast, they may also utilize

Woodford, Goldsmith, & Bruns 6

streaming cable or broadband catch-up services, time-shifted pause and rewind functionality, or

(official or unauthorized) video downloads. Such services may be offered by a wide range of

providers and platforms, including the original domestic broadcasters, their counterparts in other

geographic regions (where new shows may screen ahead of the domestic broadcast date, and

become accessible to users outside the region through the use of geo-masking VPN services), by

video streaming platforms such as YouTube (where content may have been uploaded by

production companies, one or several regional broadcasters, or fans), and by download services

from the Apple iStore to Bittorrent filesharing sites.

Audience engagement with such content remains identifiable and quantifiable in most of

these cases: on-demand platforms from official catch-up services to unauthorized filesharing

sites each generate their own usage metrics, even if they are not always shared publicly. To date,

however, such metrics have yet to be aggregated and standardized in any reliable form; a number

of scholarly and industry research projects have attempted to do so for individual platforms, but

several such studies, especially by industry-affiliated market research organizations, are also

flawed by an underlying agenda to promote fledgling on-demand services or prove the impact of

content piracy.

Further, significant challenges exist in ascribing meaning to these metrics. Mere figures

describing the number of requests for specific on-demand video streams or downloads may be

highly misleading if they turn out to be inflated by multiple requests from the same user due to

poor server performance or broadband throughput; even unique user figures may be misleading

if there is a significant influx of audiences from outside the intended region of availability

through the use of VPNs and other mechanisms. Recent research suggests, for example, that on-

demand movie and television streaming service Netflix has already gained a 27% share of the

Woodford, Goldsmith, & Bruns 7

overall on-demand market in Australia, even though Netflix does not officially operate in that

country (Ryall, 2014). Australian Netflix users’ activities are therefore likely to inflate the

metrics of the U.S. platform to which they have managed to connect.

Figures for on-demand requests and downloads also fail to accurately capture the quality

of engagement with the televisual content thus accessed: was a downloaded video actually

watched? Did viewers watch the entirety of the program? Here, in spite of their own limitations,

even conventional television ratings provide a more comprehensive picture of audience

engagement, because they are at least able to track audience sizes at regular intervals during a

broadcast, and thus to offer a glimpse or audience attrition or accretion rates. In their use of

demographically representative panels of television households, such conventional ratings also

continue to provide more detailed data on the popularity of specific programming with particular

audience segments; this is likely to be absent from the metrics for alternative channels, where

demographic data are often rudimentary at best.

Such information is especially crucial for broadcasters in the public service media sector,

where an application of conventional ratings to programming which is often deliberately

designed to address specific niche interests and audiences can significantly misjudge the ability

of such programming to meet its intended aims. Here, evaluating the forms and quality of

audience engagement is often more important than simply measuring the total size of the

audience. But for commercial television channels, too, such information provides important clues

which feed back into the design and production of new programming; there is therefore a

significant need to move beyond the limitations of merely quantitative audience measurements.

Media, communication, and cultural studies scholarship has a long history of recognising

the active audience of mass media programming (Fiske, 1992), but has traditionally found it

Woodford, Goldsmith, & Bruns 8

difficult to measure the extent and impact of audience activities or provide comprehensive

qualitative evidence beyond individual small-scale case studies. That is, scholarship in this field

has established the necessary conceptual tools for evaluating and understanding diverse forms of

audience engagement, but has so far lacked access to a substantial base of evidential data on

audience activity to which such tools may be usefully applied in order to determine and

categorize the forms of audience engagement with media content which are prevalent in the

contemporary media ecology, or to evaluate their meaning and relevance in the context of the

specific public service and/or commercial aims pursued by media organizations.

This situation has shifted markedly in recent years, especially due to the emergence of

second-screen engagement through social media as an audience practice that accompanies the

viewing of televisual content. Such engagement has turned the active audience of television into

a measurably active audience that generates a rich trail of publicly available evidence for its

activities, and this trail can be gathered through the Application Programming Interfaces (APIs)

of mainstream social media platforms, or internally from the access logs of the engagement

platforms operated by broadcasters themselves. With the computational turn (Berry, 2011) in

humanities research, such data may now be used to test and verify the conceptual models for

audience engagement that have been developed by media, communication, and cultural studies

disciplines, in order both to quantify the level of such activity for individual broadcasters and

their programming, and to benchmark the quality of this engagement against the aims and

ambitions set by the content producers.

This focus on using social media activities as an indicator of audience engagement is not

without its own limitations, however. In the first place, social media audience metrics require

active television audiences also to be active on social media, and may thus privilege particular

Woodford, Goldsmith, & Bruns 9

audience demographics that are especially likely to be using platforms such as Facebook and

Twitter to discuss their television viewing. Further, social media-based engagement with

televisual content is likely to be greatest when individual users are able to engage with other

viewers of the same programming in close to real time; such metrics continue to privilege live or

close to live viewing (through conventional broadcast or streaming services) rather than

significantly time-shifted access. For major television events, a considerable social media

audience around a shared televisual text is likely to persist at least for several hours, perhaps

even days, before and especially after the live broadcast, so that the measurement of social media

audience activities need not necessarily require exactly simultaneous engagement with the same

text. This is demonstrated for example by the global social media response to television events

such as new episodes of popular series from Doctor Who to Game of Thrones, which are

typically screened in different timeslots but in close temporal proximity to each other in different

broadcast regions around the world.

If such limitations inherent in the data derived from television-related social media

activities can be successfully negotiated, then a range of new opportunities for quantifying as

well as qualifying audience engagement with televisual content emerge. First, a number of

comparatively simple audience metrics may be established, including the volume of postings that

relate to specific programming, and the number of unique users generating such audience

responses. Here, the substantially improved precision of public social media data compared to

conventional ratings data makes it possible to identify almost to the second which moments in a

particular broadcast generated the greatest audience response, and thus how such activity ebbed

and flowed with the progress of the show; a measurement of unique active users over the course

of the broadcast also offers first insights into the influx or exodus of viewers. Various contextual

Woodford, Goldsmith, & Bruns 10

factors must also be considered in such analysis, however – different program types and formats

may lend themselves more or less well to continued social media activity, for example:

audiences may be glued to the screen during drama programming, and post social media updates

only during commercial breaks, while during political talk shows they may be more prepared to

respond to the panelists’ statements on a continuous basis.

Additionally, publicly available background data derived from the social media platforms

themselves may also be brought to bear on the analysis: for example, in addition to measuring

the total number of users participating in a social media conversation about a given show, it

would also be possible to determine the number of social media friends or followers for each

user’s account, and thus to evaluate the extent to which the broadcast has been able to attract

highly networked (which may be read as “influential”) participants. Similarly, if background data

exist not just about the size of such friendship networks, but also about their structure (as Bruns,

Burgess, & Highfield, 2014, have developed it for the Australian Twittersphere, for example), it

becomes possible both to pinpoint the location of individual users within that network, and to

determine the total footprint of a particular programme within the overall social media platform.

Such indicators begin not just to quantify total engagement, but also to provide a post-

demographic alternative to the audience segmentation models of conventional ratings: as social

media networks are often structured not primarily according to geographic or sociodemographic

factors, but by similarities in interests, this approach to analysing social media-based audience

activities offers insights into whether a specific broadcast was able to achieve deep engagement

with those sections of the overall network which are particularly concerned with the broadcast’s

topics, and/or whether it managed to generate broad engagement irrespective of users’ day-to-

day interests and preferences (cf. fig. 1). Depending on broadcaster and program type, either or

Woodford, Goldsmith, & Bruns 11

both of these objectives may be desirable: a political talk show on a niche public broadcast

channel may seek deep engagement with a narrowly defined group of so-called “political

junkies” (Coleman, 2003), for example, while a broad-based entertainment show on a major

commercial station would aim for responses from as broad a public as possible. Again, it should

be noted that such analyses assume that engagement by the social media audience either provides

a reasonable approximation of engagement by the wider television audience beyond specific

social media platforms, or that it is possible to correct for the demographic and post-

demographic skews in the measurement of audience interests and activities that such a focus on

social media-based engagement activities produces.

Fig. 1: Social media footprints of different TV programming in Australia – Twitter-based

engagement with political talkshow Q&A (left) and the 2014 Grand Final of the Australian
Football League (right). Against the backdrop of a follower network map for the 140,000 most
connected Twitter accounts in Australia (in gray), actively tweeting accounts for either
broadcast are shown in blue. Q&A tweeters are recruited predominantly from a network cluster
focusing on politics (top left); Grand Final tweeters from a cluster focusing on sports (top
center), but with much wider take-up across the Australian Twittersphere.

Finally, the immediate availability of audience members’ social media responses to

specific televisual programming also enables a qualitative analysis of their reactions beyond

mere engagement metrics. It becomes possible, for example, to extract from the content of

audience posts the key themes and topics of their responses, which may highlight the names of

Woodford, Goldsmith, & Bruns 12

popular (or at least controversial) public figures, organizations, and actors, and to chart their

relative centrality to the programming over the course of individual episodes or entire seasons.

This can also feed back into programming choices, from featuring popular journalists and

presenters in current affairs programming to enhancing storylines for favorite characters in

drama series. Such approaches may also seek to explore the use of sentiment analysis, in order

not only to determine the volume of mentions for specific themes or persons, but also to identify

the tone and context in which they are mentioned (Is a reality TV contestant controversial or

popular? Is the coverage of a topic appreciated or criticized?); it should be noted in this context,

however, that the effectiveness of current sentiment analysis techniques in processing the very

short texts of social media posts remains disputed (Liu, 2012; Thelwall, 2014).

Context-Sensitive Approaches to Measuring Audience Engagement

To a great extent, and perhaps as a reflection of the persistence of ratings thinking in

social media audience measurement, existing approaches by commercial research enterprises to

the analysis of social media data around television are largely based on relatively simplistic

volumetric measurements. Nielsen, for example, uses the SocialGuide platform to rank shows

according to what the company terms the “Unique Audience” of a show; that is, the estimated

number of Twitter users who could have seen a tweet about a show (Nielsen Social, 2014). But

this measurement fails to account for the different contexts in which shows air: for example, in

the US it compares shows screening on the less subscribed USA Network to those broadcast by

the mainstream national network ABC, and places moderately popular FOX afternoon sporting

events on an equal footing with primetime pay-per-view wrestling broadcasts. UK operator

SecondSync, which has now been purchased by Twitter, Inc., similarly ranks social media

activity by two volume-based metrics: total tweets, and tweets per minute; in both cases, it also

Woodford, Goldsmith, & Bruns 13

compares shows on different types of network without accounting for their underlying

differences (SecondSync, 2014).

Such approaches to social media audience metrics are clearly and significantly limited in

their ability to measure engagement effectively. For instance, a simple ranking of shows by the

total number of tweets they have received ignores the number of tweets posted per user, and thus

fails to differentiate between, on the one hand, broad but shallow engagement by a large number

of moderately committed viewers and, on the other, deep but narrow engagement by a dedicated

niche audience of fans. These generic metrics also implicitly assume that the mode of

engagement with a show is the same for viewers of all formats; that is to say, they assume that

audiences engage in the same way with a reality TV show as they do with a drama, for instance.

But this is disproved by SecondSync’s own data, which show that the peak of audience activity

for drama broadcasts often occurs after the conclusion of an episode, whereas for reality TV

viewers are more likely to tweet during a show (Dekker, 2014). Although a ranking of shows by

their tweets-per-minute average may allow for such genre-specific variations in audience

engagement, it does not incorporate any evidence of sustained engagement with a show; a show

that flat-lines except for a moment of major social media controversy would rank highly by this

metric, compared to a broadcast which receives solid and steady engagement throughout.

Metrics that seek to quantify sustained audience engagement, and do so with regard to the

specific characteristics of that engagement, would then already be a significant improvement

over currently available measurements. When seeking to understand the social media footprint of

television shows, it is important that contextual factors that affect social media users’

engagement with television content are accounted for. In particular, it would be desirable to

normalize available measures of the volume and dynamics of content posted through social

Woodford, Goldsmith, & Bruns 14

media, and thus of social media engagement with a show, by accounting for underlying systemic

factors such as the geographic reach of a broadcast network, the weekday and month of a

broadcast, the broadcasting genre, or the show’s time slot. In this way, viewer engagement with a

high-budget primetime drama on a major television network could be benchmarked more

meaningfully against the social media activities around a reality TV show airing on cable

television. Rather than simply comparing raw volume figures, which will always favor major

channels and primetime broadcasts, comparisons could thus be based on measurements of a

show’s social media performance relative to the long-term average for engagement broadcasts on

the same channel, in the same time slot, and/or of the same genre.

Given that this critically depends on accounting more comprehensively for the broadcast

context of a given show, it is logical to consider other fields in which contextualizing statistics is

significant. Noteworthy new impulses for the further development of social media engagement

analytics come from the field of sports metrics, where data analysts have long faced a very

similar challenge to that which underlies audience measurement: separating the signal from the

noise (Silver, 2012). Sporting analytics has addressed this challenge by seeking to account for

the fact that traditional measurements of team performance (wins and losses) and players

(individual statistics) can be influenced by a wide range of factors beyond the skill level and

performance of a player on the field, including the skill of other players on a team’s roster, the

standard of the opposition, and the playing conditions of a specific match.

Despite recent developments in ice hockey, basketball and American football

(Moskowitz & Wertheim, 2012), as well as soccer (Anderson & Sally, 2013), baseball analytics

remains the most developed of these fields, through the work of researchers such as James

(1982), Silver (2003-2009), and Tango, Lichtman, & Dolphin (2007). The field of baseball

Woodford, Goldsmith, & Bruns 15

analytics that has emerged from their efforts is called Sabermetrics (named after the Society for

American Baseball Research, SABR); we therefore refer to our adaptation of these methods to

the study of television audience engagement on social media as Telemetrics.

For the purpose of interpreting and improving contemporary audience engagement

metrics, the most useful sporting analytics we may draw on are those that seek to separate a

player’s actual performance from the contextual factors outside the player’s control that may

have affected it. In baseball, pitchers have historically been evaluated through a statistic called

ERA, or Earned Run Average, which is calculated by dividing the number of Earned Runsi

conceded by the number of innings pitched. However, this metric has been shown to be inferior

to contemporary, context-based metrics. One measure of the validity of a statistic that evaluates

performance is the extent to which it is predictive of future performance. However, research has

shown that ERA (Swartz, 2012), as a measure of pitching ability, is not as predictive of the

pitcher’s future performance as those metrics which account for context. A number of competing

statistics have been developed which account for particular elements of the pitcher’s context,

such as the quality of the fielders, the random distribution of errors, and the performance of the

opposition batters whom the pitcher faced on a given day. These alternative metrics include

measures such as xERA (expected ERA), FIP (Fielder Independent Pitching) and xFIP (expected

Fielder Independent Pitching). The statistical measure that is most commonly used in

contemporary Sabermetrics is SIERA (Skill-Interactive ERA), which measures pitching ability

by taking into account only those metrics which are solely under a pitcher’s control.

In measuring television audience engagement through social media, it is vital to control

for the systemic boost in social media activity caused by a broadcast’s time slot, network, and

other factors. To do so, we can draw on a range of sporting metrics that account for such factors

Woodford, Goldsmith, & Bruns 16

by weighting the standard measures accordingly. PERA, or Peripheral ERA (Baseball Prospectus

Team of Experts, 2004), is one example of this: it recognizes the inherent “park factors” of each

stadium where baseball is played. Essentially, this is calculated by benchmarking each playing

statistic for the home and away teams in a given stadium against their overall performance away

from that stadium: for example, a stadium with a Home Runs park factor of 112 sees 12% more

home runs than the average stadium. Each pitcher’s PERA can then be calculated by adjusting

the counts of hits, walks, strikeouts and home runs that underlie the standard ERA measure by

the park factors of the stadium where the game was played, thus eliminating any such location-

specific contextual factors.

Translating this methodology to the measurement of social media activities relating to

broadcast content, we have developed a similarly context-independent metric to quantify

Twitter-based audience engagement, the Weighted Tweet Index (Woodford & Prowd, 2014).

Using this approach, we have been able to identify a number of the contextual factors that

influence social media activity levels, including the multiplier effects resulting from the specific

television network, the genre, the time of day and year, and the location of a specific episode

within the seasonal cycle of a show. The Weighted Tweet Index builds on large longitudinal

datasets for a wide range of US television series during the 2012-13 broadcast seasons, including

Twitter activity metrics published by Nielsen SocialGuide and data collected directly from the

Twitter API. Drawing on data for 9,082 individual episodes over 21 months (April 2012 -

January 2014), we calculated a range of contextual broadcast factors analogous to the park

factors described for PERA, allowing us to understand the influence of these factors on the

volume of social media audience engagement. These factors are normalized to an index value of

1: thus, a factor of 1.28 represents overall engagement 28% above average.

Woodford, Goldsmith, & Bruns 17

 Unsurprisingly, the largest influence on social media engagement observed in our dataset

was the broadcast channel itself: we identified a significant difference in the baseline social

media activity levels for shows aired on major networks (e.g. CBS), and those shown on cable

channels such as MTV. Our data contained shows on 161 US television channels, with major

networks such as ABC (1.15), CBS (1.09) and NBC (0.80) differing substantially from cable

channels such as BBC America (0.09), Nickelodeon (0.12) and VH1 (0.47). A second key factor

that influences engagement with shows on social media is the time at which an episode airs. This

affects the size of television audiences more generally: networks have defined seasons for new

shows, pause shows over Christmas and New Year when audiences traditionally fall, and rarely

air prime shows on Fridays. Quantifying the differences between these times is key to both

evaluating historic engagement values and predicting future activities; in our analysis (Table 1),

the factors for monthly engagement varied from 0.587 (April) to 1.286 (January), and daily

variation ranged between 0.24 (Friday) and 1.51 (Tuesday).

Table 1: Selected underlying broadcast factors affecting social media

audience engagement in the US

Index Index Index

Month Factor Day Factor Network Factor
January 1.29 Monday 1.45 ABC 1.15
February 1.1 Tuesday 1.51 CBS 1.09
March 1.2 Wednesday 1.18 ESPN 1.05
April 0.59 Thursday 1.16 FOX 0.96
May 0.62 Friday 0.24 NBC 0.8
June 0.87 Saturday 0.58 TNT 0.69
July 0.86 Sunday 0.88 MTV 0.56
August 1.03 VH1 0.48
September 0.97 ABCF 0.36
October 1.27 MTV2 0.3
November 1.1
December 1.26

Woodford, Goldsmith, & Bruns 18

 These long-term factors are exceptionally valuable for any attempts to move beyond a

simplistic ranking of shows based on their raw social media activity metrics: for the first time,

they enable a benchmarking of the social media-based audience engagement with television

content that is able to compare prime time and daytime broadcasts, mainstream and cable

content, drama and reality TV genres without merely coming to the obvious conclusion that

mainstream content generates more tweets, likes, and comments. The Weighted Tweet Index

provides a valuable starting-point for advancing beyond the basic metrics generated by

commercial analysts such as Nielsen SocialGuide and SecondSync, and constitutes a key tool for

the evaluation of shows on a like-for-like basis, and for predictions of how a successful cable

show might fare if aired on a mainstream network. Its weighted metrics allow networks and

producers to benchmark their shows against others, not just on raw numbers, but by controlling

for the other factors which influence audience engagement. However, it is important to note the

limitations of this approach. Key among these is that such weightings can never account for the

content of a specific episode. For example, in the 2013 season of Big Brother (US) we saw a

large spike in social media activity that was attributable to a controversy over racism; such acute

events are impossible to account for through purely quantitative approaches. Necessarily, the

existing weightings can also be further refined, just as the sporting analytics frameworks we have

drawn from were developed over a number of years.

A particular focus of related sporting analytics has been the prediction of future

performance, on both the team and the player level, for a variety of purposes. Team executives

need to make decisions on roster composition, contract values, and other issues; sporting media

and fan sites are tracking the performance of teams and seek to contribute insightful

commentary; participants in fantasy sports and gambling markets may have significant financial

Woodford, Goldsmith, & Bruns 19

investment in players’ performances – these all subscribe to data sites that offer performance

predictions based on cutting-edge data analytics approaches. One example of this is Baseball

Prospectus’s PECOTA (Player Empirical Comparison and Optimization Test Algorithm), which

uses advanced Sabermetric statistics to predict players’ performances several seasons into the

future. These player-level statistics can then be used with the Pythagorean expectation formula,

developed by Bill James (1980), to estimate the games a team should have won, in order to

calculate expected wins and losses for teams over the course of a season.

By determining the contextual broadcast factors that influence social media engagement

and applying them to the long-term social media engagement averages for a show once the

scheduling of upcoming episodes is known, it is similarly possible to generate predictive

measures of the expected social media volume for these episodes. Predictive measures can serve

a number of purposes: for the viewer, they enable the selection of shows that are likely to have

an active social media audience to engage with; for broadcasters, television producers, and social

media strategists, they provide a benchmark to measure whether a show has been as successful

on social media as it should have been; and for advertisers, they offer a tool for more targeted

promotions, both through traditional commercials and directly through social media-based

advertising that reaches a specific social media demographic. Although current social media

audience measurement systems remain imperfect and are as yet unable to meet all of the

demands of all of the various stakeholders and interested parties – producers, broadcasters,

advertisers, advertising sales agents, media buyers and planners, audience research agencies,

academics and audiences themselves – they can nonetheless already illuminate new forms of

audience behavior and provide insights into particular audiences’ levels of engagement with

screen content.

Woodford, Goldsmith, & Bruns 20

 Our new approach draws on developments in sports metrics to develop a method for

comparing both the performance of particular television content and measuring audience

engagement through computational analysis of social media data. Our findings to date indicate

that, for the moment at least, social media derived television metrics are no cure-all for the

current shortcomings of traditional television audience metrics. Ratings systems for commercial

television will continue to be used for as long as the various stakeholders are able to extract

value from them. New measurement systems such as Telemetrics that are based on social media

analysis are unlikely to replace the ratings; rather, such systems will co-exist with and

complement each other as the media industries’ long quest to understand their audiences

continues.

References

Anderson, C., & Sally, D. (2013). The numbers game: Why everything you know about soccer is
wrong. London, UK: Penguin Press.

Andrejevic, M. (2007). iSpy: Surveillance and power in the interactive era. Lawrence, KS:
University Press of Kansas.

Andrejevic, M. (2013). Infoglut: How too much information is changing the way we think and
know. New York, NY: Routledge.

Balnaves, M., O’Regan, T., & Goldsmith, B. (2011). Rating the audience: The business of
media. London, UK: Bloomsbury Academic.

Baseball Prospectus Team of Experts. (2004). Baseball prospectus. New York, NY: Workman
Publishing Company.

Berry, D. (2011). The computational turn: Thinking about the digital humanities. Culture
Machine, 12. Retrieved February 25, 2015 from
http://www.culturemachine.net/index.php/cm/article/view/440/470

Bourdon, J., & Méadel, C. (2014). Introduction. In J. Bourdon & C. Méadel (Eds.), Television
audiences across the world: Deconstructing the ratings machine (1-30). Basingstoke,
UK: Palgrave Macmillan.

Woodford, Goldsmith, & Bruns 21

Bruns, A., Burgess, J., & Highfield, T. (2014). A ‘big data’ approach to mapping the Australian
Twittersphere. In K. Bode & P. Arthur (Eds.), Advancing digital humanities.
Basingstoke: Palgrave Macmillan.

Coleman, S. (2003). A tale of two houses: The House of Commons, the Big Brother house and
the people at home. Parliamentary Affairs, 56(4), 733-758.

Council for Research Excellence. (2010). The state of set-top box viewing data as of December
2009. Retrieved February 24, 2015 from http://www.researchexcellence.com/files/
pdf/2015-02/id90_set_top_box_study_ 2_24_10.pdf

Dekker, K. (2014). 2014 – The top tweeted shows so far. SecondSync Tumblr. Retrieved January
8, 2015, from http://secondsync.tumblr.com/post/80172270097/2014-the-top-tweeted-
shows-so-far

Fiske, J. (1992). Audiencing: A cultural studies approach to watching television. Poetics, 21(4),
345-359.

James, B. (1980). The Bill James abstract. Self-published.

James, B. (1982). The Bill James baseball abstract. New York, NY: Ballantine Books.

Liu, B. (2012). Sentiment analysis and opinion mining. San Rafael, CA: Morgan & Claypool.

Moskowitz, T., & Wertheim, L. J. (2012). Scorecasting: The hidden influences behind how
sports are played and games are won. New York, NY: Three Rivers Press.

Napoli, P. (2003). Audience economics: Media institutions and the audience marketplace. New
York, NY: Columbia University Press.

Nielsen Social. (2014). Nielsen SocialGuide Intelligence. Retrieved January 8, 2015, from
http://www.nielsensocial.com/product/social-guide-intelligence/

Pesce, M. (2006). You biquity. Keynote address from the Webdirections South Conference,
Sydney, AU.

Ryall, J. (2014, 16 July). How Netflix is quietly thriving in Australia. Sydney Morning Herald.
Retrieved January 8, 2015, from http://www.smh.com.au/entertainment/tv-and-
radio/how-netflix-is-quietly-thriving-in-australia-20140716-ztirm.html

Savage, P., & Sévigny, A. (2014). Canada’s audience massage: Audience research and TV
policy development, 1980-2010. In J. Bourdon & C. Méadel (Eds.), Television audiences
across the world: Deconstructing the ratings machine (69-87). Basingstoke, UK:
Palgrave Macmillan.

Woodford, Goldsmith, & Bruns 22

SecondSync. (2014). Twitter leaderboard. Retrieved January 8, 2015, from
https://secondsync.com/leaderboard.html

Silver, N. (2003-2009). Nate Silver author archives. Baseball Prospectus. Retrieved January 8,
2015, from http://www.baseballprospectus.com/author/nate_silver

Silver, N. (2012). The signal and the noise: Why so many predictions fail – but some don’t.
London, UK: Penguin Press.

Silvey, R. (1951). Methods of viewer research employed by the British Broadcasting

Corporation. Public Opinion Quarterly, 15(1), 89-94.

Swartz, M. (2012). Are pitching projections better than ERA estimators? Fangraphs. Retrieved
January 8, 2015, from http://www.fangraphs.com/blogs/are-pitching-projections-better-
than-era-estimators/

Tango, T. M., Lichtman, M. G., & Dolphin, A. E. (2007). The book: Playing the percentages in
baseball. Lincoln, NE: Potomac Books.

Thelwall, M. (2014). Sentiment analysis and time series with Twitter. In K. Weller, A. Bruns, J.
Burgess, M. Mahrt, & C. Puschmann (Eds.), Twitter and Society (83-96). New York, NY:
Peter Lang.

Woodford, D., & Prowd, K. (2014). Everyone’s watching it: The role of hype in television
engagement through social media. Paper presented at the Social Media and the
Transformation of Public Space conference, Amsterdam, NL.

World Federation of Advertisers (WFA). (2008, June). Blueprint for consumer-centric holistic
measurement. Brussels, BE: World Federation of Advertisers. Retrieved January 8, 2015,
from http://www.wfanet.org/media/pdf/Blueprint_English_June_2008.pdf

Wurtzel, A. (2009). Now. Or never. An urgent call to action for consensus on new media
metrics. Journal of Advertising Research, 49(3), 263-265.

i Earned runs differ from total runs in that they exclude any runs given up after a fielding error prevented the third out of an inning.

Woodford, Goldsmith, & Bruns 23