Social advertising

Social Advertising: How Advertising that Explicitly
Promotes Social Influence Can Backfire
Catherine Tucker∗
June 3, 2016

Abstract
In social advertising, ads are targeted based on underlying social networks and
highlight when a friend has ‘liked’ a product or organization. This paper explores the
effectiveness of social advertising using data from field tests of different ads on Facebook by a nonprofit. We find evidence that social advertising is somewhat effective, but
that social advertising is less effective if the advertiser explicitly states they are trying
to promote social influence in the text of their ad. Indeed, automated endorsements
appear to backfire in general unless the advertiser refrains completely from promoting
social influence in their ad content. We exploit variation in the appearance of endorsements due to differences in privacy settings, and find that the effectiveness of social
advertising is due to the ability of targeting based on social networks to uncover similarly responsive consumers, especially for consumers in non-traditional target markets.
Our results suggests that advertisers must avoid being overt in their attempts to use
automated social endorsements in their advertising.

∗

Catherine Tucker is the Sloan Distinguished Professor of Marketing at MIT Sloan School of Management,
Cambridge, MA and Research Associate at the NBER. Thank you to Google for financial support and to an
anonymous nonprofit for their cooperation. Thank you also to Jon Baker, Ann Kronrod, Preston Mcafee,
and seminar participants at the George Mason University Roundtable on the Law and Economics of Internet
Search, Carnegie Mellon, the University of Rochester, UCLA and Wharton for valuable comments. All errors
are my own.

1

Electronic copy available at: />

1

Introduction

Recent advances on the internet have allowed consumers to interact across digital social networks. This is taking place at unprecedented levels: Facebook was the most visited website
in the US in 2010, accounting for 20% of all time spent on the internet, a higher proportion
than Google or Yahoo! (ComScore, 2011).1 However, it is striking that traditional paid marketing communications have been at the periphery of this explosion of social data despite
the documented power of social influence on purchasing behavior (Algesheimer et al., 2005).
To address this lack, Facebook2 and LinkedIn3 have introduced a new form of advertising
called ‘social advertising.’ A Social Ad is an online ad that ‘incorporates user interactions
that the consumer has agreed to display and be shared. The resulting ad displays these interactions along with the user’s persona (picture and/or name) within the ad content’ (IAB,
2009). This represents a radical technological development for advertisers, because it means
that potentially they can co-opt the power of an individual’s relationships online to target
advertising and engage their audience.
This paper asks whether such automated ad units designed to promote social influence
are effective, and what messaging advertisers should use around them. We explore these
questions using data from a field experiment conducted on Facebook by a nonprofit. This
field experiment compared the performance of social ads with conventionally targeted and
untargeted ads, and examined how the performance varied with different message combinations. The social ads were targeted to the friends of ‘fans’ of the nonprofit on Facebook. The
1

This has increased in 2016 to 50 minutes each day. See />business/facebook-bends-the-rules-of-audience-engagement-to-its-advantage.html
2
Many other social media platforms have embraced such ads. See />uploads/2015/09/Social-Advertising-Best-Practices-0509.pdf for examples.
3
Linkedin appears to have started this practice in 2011. See - />
2

Electronic copy available at: />

ads featured that fan’s name and the fact that they had become a fan of this nonprofit.4 We

find that on average these social ads were effective and that this technique is particularly
useful for improving the performance of untargeted campaigns.
Through randomized field tests, we investigate the effectiveness of advertisers deliberately promoting social influence in their advertising copy through including a statement
that encourages the viewer to, for example, ‘be like their friend.’ We find that consumers
reject attempts by advertisers to explicitly harness or refer to a friend’s actions in their ad
copy. This result contrasts with previous empirical research that finds consistent benefits
to firms from highlighting previous consumer actions to positively influence the consumers’
response (Algesheimer et al., 2010; Tucker and Zhang, 2011). This rejection is reasonably
uniform across different wording, though slightly less severe for ads that make a less explicit
reference to friendship. We present evidence that this happens both when we consider actual
subscriptions and when we control for variation in impressions.
We then use non-experimental variation due to differences in friends’ privacy settings to
explore whether the inclusion of an endorsement or the use of an individual’s online social
network explain our findings. Comparing the performance of these ads that contained the
name of the fan and were targeted towards the fan’s friends with those that were simply
targeted to that fan’s friends suggests that their effectiveness stems from the ability of
social targeting to uncover similarly responsive consumers. Indeed, the fact that including
endorsements appear to backfire on average in our data can be explained by a negative
reaction of Facebook users when that automated endorsement is coupled with an advertiser
explicitly trying to promote social influence.
We then present additional evidence to rule out two potential alternative explanations
for our findings. First, we rule out that the overt mention of social influence simply made
4

While Facebook has now abandoned the ‘fans’ terminology, they still offer this ad unit in a slightly different form - see />
3

Electronic copy available at: />

people aware they were seeing an ad rather than something organic to the site. We do this by

comparing an ad that states it is an ad with an ad that does not, and finding no difference.
Second, we investigate whether it was simply that messages referring to friendship were
bad ad-copy, by examining how the ads perform for a group of Facebook users who have
shown a visible propensity for social influence. We identify such users by whether or not
they have a stated attachment to a ‘Fashion Brand’ on their Facebook profile. These users,
in contrast to our earlier results, react more positively to the advertiser explicitly co-opting
social influence than to a message that did not. This suggests that it was not simply that the
message was badly communicated, but instead that our results reflect distaste for explicit
references to social influence accompanied by automated endorsements among most, though
not all, of consumers we targeted.
This paper contributes to three main literatures. The first literature studies how social
networks affect adoption of new products and services and how such social networks can be
used to target consumers.5 Provost et al. (2009) show how to use browsing data to match
groups of users who are socially similar. Similarly, Oestreicher-Singer and Sundararajan
(2012) show the importance of such connections in recommendations systems. Aral and
Walker (2012) discuss how to use social networks to target potential users, and Aral and
Walker (2011) discuss viral marketing strategies on such networks. Hill et al. (2006) show
that ‘Network neighbors’ - those consumers linked to a prior customer - adopt a service at a
rate 35 times greater than baseline groups selected by the best practices of a firm’s marketing
team. Our paper builds on this result and shows that in particular in untargeted populations,
social networks can be used to predict response to advertising as well as adoption. We extend
Hill et al. (2006), by emphasizing that though targeting social networks with advertising to
5

Zubcsek and Sarvary (2011) present a theoretical model that examines the effects of advertising on
a social network, but assume that a firm cannot directly use the social network for marketing purposes.
Instead, firms have to rely on consumers to organically pass their advertising message within the social
networks.

4

Electronic copy available at: />

promote adoption may seem attractive, firms themselves should be wary of explicitly trying
to push social influence in the wording of their marketing communications.
Bakshy et al. (2012) is unusual in this literature, since it focuses on advertising and evaluates the effectiveness of increasing the number of social cues explicitly mentioning friends
present in a Social Ad unit on Facebook. They randomly vary whether a Social Ad displays
endorsements from either none, one, two or three friends and find that click rates increase for
one or two friends, but only marginally increase for three friends. This is an important finding, given work by Centola (2010) that shows that adoption is more likely when participants
receive social reinforcement from multiple neighbors in their social network. Bakshy et al.
(2012) also find around a 5% increase in clicks from naming an individual relative to the
total number of people who like a product. Our paper provides two contributions that build
on this work. First, unlike Bakshy et al. (2012) who study ad units that did not allow variation of messaging, we show that there can be a negative effect of ad units displaying social
endorsements if they are accompanied by an ad campaign that explicitly refers to and tries
to embroider upon them, unless the ad is targeted at people who have already embraced
commercialization of their social networks. Second, using non-experimental variation, we
highlight that if an advertiser does not account for the underlying like-mindedness of those
who are socially connected, they are likely to vastly overestimate the effectiveness of social
endorsements appearing in ads - this is something that Bakshy et al. (2012) control for but
do not measure.
The second literature this paper contributes to is a small literature that highlights the
difficulty for advertisers of explicitly harnessing social influence in marketing communications. Lambrecht et al. (2014) shows that it is difficult to engage the kind of people who
propagate trends on Twitter. Tucker (2015) shows that advertising content that is naturally
viral is also often less persuasive. Bakshy et al. (2011) uses Twitter user data to model
the effects of potential influencers in spreading a message, and find that the size of an influ5

Electronic copy available at: />

encer’s network does not provide clear guidance on whom to compensate. The finding of this
paper, that attempts to explicitly coopt social influence in advertising wording can backfire,

provides further and concrete evidence that advertisers need to be careful in how they shape
marketing communications designed to take advantage of online social networks.
Third, this paper contributes to a literature that evaluates the effectiveness of personalization in advertising content. Some of this research highlights positive effects of personalization. For example, Sahni et al. (2016) show that even adding a name to email headers
can increase effectiveness. However, Lambrecht and Tucker (2013) suggests that too much
specificity, unless it is timed right, can backfire, and Tucker (2014) suggests that similarly
without appropriate privacy controls personalization in messaging can backfire. By contrast,
this paper studies a new and potentially sensitive form of personalized content, which is the
automated inclusion of endorsements from an online social network. In line with the general
themes of this literature, our research suggests that such messaging can be effective but that
the advertiser has to be careful to not directly refer to such endorsements in their messaging.
Managerially, our results have important implications. Social advertising is effective.
However, when advertisers attempt to reinforce this social influence in ad copy, consumers
are less likely to respond positively to the ad. This is, to our knowledge, the first piece
of empirical support for emerging managerial theories that emphasize the need for firms to
not appear too obviously commercial when exploiting social media (Gossieaux and Moran,
2010).

6

Electronic copy available at: />

2

Field Experiment

The field experiment was run by a small nonprofit that provides educational scholarships for
girls to attend high school in East Africa. Without the intervention of this nonprofit, and
other nonprofits like them, many girls do not attend secondary school because their families
often prioritize the education of sons. Though the nonprofit’s main mission is funding these
educational scholarships, the nonprofit has a secondary mission which is to inform young

people in the US about the state of education for African girls. It was in aid of this secondary
mission that the nonprofit set up a Facebook page. This page serves as a repository of
interviews with girls where they describe the challenges they have faced. In general, their
fans on Facebook do not have direct experience with the African education system, ruling
out some of the prosocial behavior studied in Small and Simonsohn (2008).
To launch the field experiment, the nonprofit set up 18 simultaneous campaigns on Facebook.6 Table 1 summarizes the 18 conditions that were run. Since legal restrictions shaped
the design of this field test in a manner which prevented a full factorial design, this section
lays out in detail the different variations of ads tested, and how they were decided upon.
One focus of the experiment was that 15 of the ad campaigns employed the Facebook ad
option which meant that they were targeted only to users who were friends of existing fans
of the nonprofit. Assuming their friend had not opted out of having their name displayed
on Facebook, such ads also displayed a ‘social endorsement’ where the name of the friend
who had liked the nonprofit was shown at the bottom of the ad as shown in Figure 1. The
Facebook ad platform does not allow advertisers to separately test the effects of a social
6

This kind of horserace between campaigns was described by Facebook as being the best way for
an advertiser to set up ‘A/B’ testing or randomized testing between different ad campaigns on their
platform. Such an approach is described in ‘A/B Testing your Facebook Ads: Getting better results through experimentation’ (Facebook, 2010) Since this paper was written, the general structure and
set up of campaigns at Facebook has changed in a manner which makes experimentation easier - see
for an example.

7

Electronic copy available at: />

endorsement and social targeting.7 Instead, these are combined in a single Social Ad unit
that the advertiser can compare against not using a Social Ad unit. Later in this paper, we
use non-experimental variation to tease apart the difference, but this is obviously less clean
than being able to directly randomize the two. Therefore, the experimental variation we

study in the paper is the difference between the Social Ad unit and the non-Social Ad unit.
Figure 1 displays as an example of a Social Ad unit an anonymized sample ad for Campaign 2. The blacked-out top of the ad contained the nonprofit’s name. The grayed-out
bottom of the ad contained a supporter’s name, who had ‘liked’ the nonprofit and was a
Facebook friend of the person who was being advertised to. It is only with developments in
technology and the development of automated algorithms that such individualized display
of the friend’s name when pertinent is possible. At the time the experiment was run, these
ads appeared on the right hand of users’ Facebook pages. At the time, Facebook did not
share data with its advertisers on whether these ads were seen on a ticker or on other pages
that a user may be browsing. As can be seen in Figure 1, each different ad was accompanied by the same picture of an appealing secondary-school student who had benefited from
their program. Based on the work of Small and Verrochi (2009), this girl had a unsmiling
expression.
The nonprofit also explored different ad-text conditions. The different ad texts were
broadly designed to cover the kinds of normative and informational social influence described
by Deutsch and Gerard (1955); Burnkrant and Cousineau (1975).8
These differences in ad text are intended to represent a traditional taxonomy of social
7

One advantage enjoyed by the researchers in Bakshy et al. (2012) is that because they worked inside
Facebook, they were able to perform randomizations that are not available to advertisers which allowed them
to test different versions of the Social Ad unit. In this paper, we focus instead on measuring randomized
variation which is accessible to outside advertisers, such as message content and demographic targeting, and
understanding its interaction with the Social Ad unit.
8
Other forms of social influence studied in the literature involve network externalities where there is a
performance benefit to multiple people adopting (Tucker, 2008), but it is unlikely that there is a performance
benefit here.

8

Electronic copy available at: />

influence rather than exploring more subtle categories developed by state-of-the-art social
influence research. Ultimately, any message that refers to any form of social influence in
the uncontrolled context of an ad unit can potentially evoke a multitude of other forms of
social influence, and consequently cannot distinguish cleanly between different types of social
influence. We do not claim that the variation in message conditions captures the frontier of
our understanding of social influence. The recent literature on social influence emphasizes
that the underlying mechanism is nuanced and more complex than traditional taxonomies
might suggest. For example, Cialdini and Goldstein (2004) suggest that social influence may
be driven at the subconscious rather than the conscious level. However, such advances are
even harder to capture in such a constrained setting where there is no individual data, data
on the user’s state of mind, or an ability to manipulate the external environment in which a
user sees the ad. Similarly, as outlined by Kallgren et al. (2000), newer work emphasizes that
social influence is mediated by focus, and given the lack of attention most users give to ads,
again this is difficult to manipulate in such an uncontrolled setting. However, the variation
in messages does allow us to study whether explicit advertising messages that attempt to use
different types of wording to evoke social influence are effective in general, and the extent
to which advertisers should try and follow various implications of these more traditional
taxonomies in their advertising wording.
As shown in Table 1, the Social Ad units displayed all five variants of potential messages.
For each of the non-Social Ad campaigns, we ran the baseline variant of the ad text, which
simply says ‘Help girls in East Africa change their lives through education.’ The nonprofit
could not run the other four message conditions that refer to others’ actions, because federal
regulations require ads to be truthful and they did not want to mislead potential supporters.
There was also a demographic layer to the experiment. The ad campaigns were targeted
to three different groups. The first group was a broad untargeted campaign for all Facebook
users aged 18 and older in the US who were not already affiliated with the nonprofit. The
9

Electronic copy available at: />

second group were people who had already expressed interest in other charities. These people
were identified using Facebook’s ‘broad category targeting’ of ‘Charity + Causes.’ The third
group were people who had already expressed an interest in ‘Education + Teaching.’ These
latter two categories were chosen by the nonprofit as they wanted to have a benchmark
to compare the performance of social advertising that was more nuanced and likely to be
employed than a simple untargeted campaign. Previously, the nonprofit had tried such
reasonably broad targeting with little success and was hopeful that social advertising would
improve the ads’ performance (Tucker, 2014). In all cases, the nonprofit explicitly excluded
current fans from seeing its ads.
Table A3 in the appendix describes the demographics of the roughly 1,500 fans at the
beginning of the campaign. Though the initial fans were reasonably spread out across different age cohorts, they were more female than the average population, which makes sense
given the nature of the nonprofit. At the end of the experiment, the fans were slightly more
likely to be male than before.
The way that Facebook deploys its ‘A/B’ does not exclusively allocate users to a condition. This means that a Facebook user could see a socially enabled ad one day and a
non-socially enabled ad the next day. The randomization in the ordering of which ads are
displayed means that the average effect of exposure to the different ad-types should hold.
However, the lack of individual-level data provided by Facebook to external researchers
means that we cannot study whether there are explicit complementarities or substitution
effects between the two different types of advertising.
One issue, of course, with an uncontrolled setting such as Facebook, is what would have
happened in the absence of advertising. Data from the nonprofit that covers the months
prior to the campaign suggest that it attracted roughly five new subscribers to its Facebook
Newsfeed each month, or on average one new supporter a week. Analysis of aggregate data
suggested that this pattern continued during the five weeks that the experiment ran. There
10

Electronic copy available at: />

appeared to be six new subscribers over that period who could not be attributed directly

to the Facebook advertising campaign. This is too small a number to perform statistical
analysis on, but is suggestive that the non-advertising baseline of attracting users to the
website organically is small.9

3

Data

The experiment was run on Facebook for five weeks over the Fall of 2011.
The data that Facebook shares with advertisers and that the nonprofit has access to is
both anonymous and aggregated, so that firms and researchers can only evaluate the performance of a campaign on a single day. This lack of individual data means that we cannot
trace the effects of social advertising on the friends of any one individual. It also means
that we cannot examine heterogeneity in the degrees of influence across individuals, as is
studied, for example, by Godes and Mayzlin (2009) in their study of offline firm-sponsored
communications. However, given that the central research question of the study is whether,
on average, different types of social advertising are more effective from an advertiser perspective, the aggregate nature of the data is sufficient. The aggregate nature of the data means
we cannot explore the underlying network structure, nor use any of the recently developed
techniques to estimate peer effects in networks experimentally such as those suggested by
Walker and Muchnik (2014); Athey et al. (2015).
Table 2 shows the total campaign statistics over the course of the five-week period. The
campaign was seen by three million unique people who use Facebook and produced over
3,000 clicks and 1,700 new connections or subscriptions to the nonprofit’s Newsfeed.
Table 3 reports daily summary statistics for each of the 18 campaigns in our data. Over
9

This also addresses a related concern which is what effect there was from the fact that news of users’
likes could be rebroadcast in a variety of ways on a Facebook profile. In this period, the majority of users’
likes of the nonprofit would have been reported in the ‘News ticker’ which is to the right-hand side of the
ads on Facebook. These likes are broadcast instantaneously at the time they occurred. However, it is not
clear the extent to which these rebroadcasts were observed by friends.

11

Electronic copy available at: />

a five-week period, there were 630 observations of a campaign on each day. There are two
click-through rates reported in Table 3. The first click-through rate is the proportion of
people who clicked on an ad that day. The denominator here is the Ad-Reach measure that
captures the number of people exposed to an ad each day. The second click-through rate is
per ad impression. We focus on the former in our econometric analysis, because impressions
can be a function of person refreshing their page or using the back button on the browser
or other actions which do not necessarily lead to increased exposure to the ad. We show
robustness subsequently to using this click-through rate per impression measure. Due to
the relatively small number of clicks, these click-through rates are expressed as percentage
points or sometimes as fractions of a percentage point.10
The data also contains an alternate means of measuring advertising success. The connection rate measures the number of people who liked a Facebook page within 24 hours of
seeing a sponsored ad, where the denominator is the ad’s reach that day. We compare this
measure to clicks in subsequent analysis to check that the click-through rate is capturing
something meaningful. We also use the cost data about how much the advertiser paid for
clicks for each of the campaigns in a robustness check.
Table 4 reports summary statistics at the daily level for each of the 18 different campaigns.
It is immediately apparent that there was large variation in the number of impressions for
each campaign - this is one of the considerations which motivates us to use an aggregate
logit specification. This variation in impressions based on ad performance is a challenge
for all empirical research not conducted by advertising platforms themselves. Indeed, these
challenges are documented in Johnson et al. (2015). As well as using the aggregate logit
specification, we also conducted a variety of robustness checks to reassure ourselves and
readers that this variation does not drive our results which are discussed in section 4.4.

10

The data reassuringly suggests that there were only five occasions where someone clicked twice on the
ads. Therefore, 99.8% of the click-through rate we measure captures a single individual clicking on the ad.

12

Electronic copy available at: />

4

Results

4.1

Does Social Advertising Work?

First, we present some simple evidence about whether social advertising is more effective
than regular display advertising. Figure 2 displays the basic comparison of aggregate (that
is, across the whole five-week period) click-through rates between non-socially-targeted ads
and ads that were socially targeted. Since these are aggregate click-through rates, they differ
from the daily click-through rates reported in Table 3. These are expressed as fractions of a
percentage point. It is clear that social advertising earned far larger click-through rates.
However, as is obvious in Table 4, there was variation in the number of impressions
across the different campaigns, which may be either a function of the Facebook underlying
advertising algorithm or availability of appropriate Facebook users to show ads to. To help
control for this, we turn to an aggregate logit model.11 Since our data is at the campaignday level, we build our main empirical specification at the individual level and then use
aggregated estimation techniques to reflect the fact that we only have campaign-day level
data.
For individual i who sees message content j and targeted based on k on day t, the
likelihood of clicking on a Facebook advertising technique is:

Clickijkt =
+ β1 SocialT argetingAndEndorsementj
+ β2 SocialT argetingAndEndorsementj × M essagej
+ β4 SocialT argetingAndEndorsementj × U ntargetedk
+ β3 M essagej + β5 U ntargetedk + δt +

jk

11

(1)

In earlier versions of this paper we used an ordinary least square specification at the campaign level,
complemented by a poisson and negative-binomial robustness check, with similar results.

13

Electronic copy available at: />

SocialT argetingAndEndorsementj is an indicator for whether or not this campaign
variance was socially targeted and displayed the endorsement. Since Facebook does not allow
the testing of these different features separately, this is a combined (rather than separable)
indicator. M essagej is an indicator for which of the five message variants was displayed.
U ntargetedk is a fixed effect that captures whether this was the untargeted variant of the ad.
This controls for underlying systematic differences in how likely people within that target
and untargeted segment were to respond to this nonprofit. We include a vector of date
dummies δt . Because the ads are randomized, δt should primarily improve efficiency.12
Facebook reports data by grouping all connections, unique clicks, clicks, reach and impressions. This means that while a Facebook users makes a binary choice whether to click,
our data is aggregated across consumers, and we observe a number of successes (unique

clicks) out of a number of trials (reach) for each campaign-day. As discussed, it is desirable to account for differences in daily impressions across campaigns. To see why this is
important, imagine two campaigns, one which received 100 impressions and the other which
received 10,000 impressions, where both received zero clicks. Simply using the click-through
rate as a dependent variable would effectively treat these instances as the same, though
they convey different information. As a result, we estimate an aggregate logit model using
maximum likelihood (Flath and Leonard, 1979).
Let F denote the logistic likelihood function. Due to the aggregate nature of Facebook
data, which does not have user-level variables, all individuals i exposed to a particular
campaign j with message m on day t have the same vector of x control variables. The
likelihood of observing each observation of the sum of positive unique clicks as a function of
the sum of reach for that campaign that day is then:

F (βx)s {1 − F (βx)}r−s
12

Indeed, specifications without δt are similar.

14

Electronic copy available at: />
(2)


where s is the number of unique clicks and r is the population of Facebook users exposed
to the messages.
Table 5 reports our initial results. Column (1) presents results of a simple specification
that focuses only on the Social Ad unit and essentially replicates the insights of Figure 2. It
suggest that social targeting and a friend’s endorsement increased click-through propensity.
Column (2) adds an extra coefficient that indicates whether that campaign was untargeted
rather than being targeted to one of the customer groups identified as being likely ‘targets’

by the nonprofit - Educational and Charity supporters. It suggests that indeed, as expected,
an untargeted campaign was ineffective, but that its performance was enhanced by the use of
social advertising. Given the widely reported lack of efficacy of untargeted campaigns (Lewis
and Reiley, 2014), the increase in effectiveness allowed by social advertising appears large
for untargeted campaigns and potentially emphasizes that one managerially useful factor in
social advertising is its ability to identify potential targets among an otherwise untargeted
population.
Column (3) then explores how the particular message used interacts with the performance
of social advertising. In the initial specification, we compare the performance of all the ad
units that had wordings that attempted to touch upon social influence relative to the baseline. We use the additional binary indicator variable Explicitj to indicate when the advertiser
uses a message that evokes social influence explicitly in their ad copy, in addition to the social endorsement automated by the Facebook algorithm. This covers all the non-baseline
conditions described in Table 1. We interact this with SocialT argetingAndEndorsementj ,
meaning that SocialT argetingAndEndorsementj now measures the effect of the baseline
message effect, and the interacted variable measures the incremental effect of explicitly mentioning the potential for social influence in the ad.
The negative coefficient on the interaction between SocialT argetingAndEndorsementj
and Explicit in Column (3) suggests that explicit reference to a social influence mechanism
15

Electronic copy available at: />

in the ad affected the performance of the ad negatively. That is, when the advertiser was
explicit about their intention to harness social influence, it backfired. Further, the larger
point estimate for SocialT argetingAndEndorsementj suggests that the baseline message is
even more effective than the estimates in Columns (1) and (2) of Table 5 suggests.
Column (4) in Table 5 reports the results of a specification where we break up Explicit
by the different types of ‘social influence’-focused advertising messages featured in Table 1.
It is striking that all measures are negative. It is also suggestive that the one message that
had a smaller point estimate than the others did not refer to the friend explicitly but instead
referred obliquely to the friend’s action. This is speculative, however.
One explanation of these results is that implementations of common influence strategies

that make the fact that they are implementing an influence strategy explicit are less effective
(Kaptein et al., 2011). This reflects the potential for negative consequences for marketing
communications of perceptions of persuasive intent and means of persuasion in Boush et al.
(2009). We go on to present evidence that such negative effects of ad messaging that explicitly
promotes social influence are likely driven by the presence of an automated endorsement.
4.2

Using Non-experimental Variation to Tease Apart the Effect of Endorsements and Targeting

Obvious next questions are what explains the success of social advertising, and the extent
to which the pairing of a Social Ad unit with explicit attempts to coopt social influence in
advertising wording appears to backfire.
One set of explanations lies in the idea that the endorsement of a friend is informative.
This was the explanation documented in Bakshy et al. (2012). Another explanation is that
social targeting uncovers people who will be more likely to be interested in their nonprofit as
they are similar, in unobserved ways, to their friends who are already fans of the nonprofit.
Manski (1993) pointed out that this particular issue of distinguishing homophily (unobserved

16

Electronic copy available at: />

characteristics that make friends behave in a similar way) from the explicit influence of friends
on each other is empirically problematic. This distinction was also documented in Aral
et al. (2009); Shalizi and Thomas (2011) who suggest that contagion and homophily will be
confounded in all observational studies. The empirical literature in economics has similarly
addressed the obvious identification problem outlined by Manski (1993), that arises from
attributing causal behavior when socially-connected people behave in a similar way (Bayer
et al., 2008; Cai et al., 2009; Conley and Udry, 2010).
Ideally, to address this we would simply randomize whether users saw the endorsement

or not. However, Facebook’s advertiser interface does not allow external parties to do that.
What we can do is take advantage of the fact that sometimes ads were shown to people
without the endorsement if that fan has selected a privacy setting which restricts the use of
their image and name. The interface which users used to do this is displayed in Figure A1;
users simply selected the ‘No One’ rather than the ‘Only my friends’ option.
Of course, this will not approximate true randomization. It is likely that the fans who
select stricter privacy settings differ in unobserved ways from those who do not, and that
therefore their social networks may differ as well. This leads to a potential for unobserved
bias. However, it is possible to conjecture in which direction the bias is likely to go. This
is because it seems likely that people who have stricter privacy settings are more likely to
distrust or not click on online advertising (Hoofnagle et al., 2012). There is also evidence
that within a social network friends may share such attitudes (Shalizi and Thomas, 2011).
As a consequence, it seems likely that if people who are friends with people who are privacysensitive on Facebook are also privacy sensitive, they may share their distaste for online
advertising and they are less likely to click on ads. It seems likely, though of course not
provable, that the estimate for any Social Ad unit’s effectiveness for that group of users will
be biased downwards by their distaste for clicking on ads. Even if this is not correct, the
comparison of the effectiveness of different randomized message wordings for the subset of
17

Electronic copy available at: />

Facebook users who have privacy-protective friends should still hold, as this was randomized.
Table 6 displays the results of repeating the analysis done for Table 5 but this time
distinguishing whether or not a social endorsement was present for that individual i, which is
a function of whether their friend had precluded such endorsements in their privacy settings.
A comparison of Columns (1) and (2) in Table 5 with Columns (1) and (2) in Table 6 makes
it clear the ads that were displayed to friends of fans but lacked a clear endorsement were
if anything more effective than those that had a clear endorsement. This is an unexpected
result, given the huge body of work and also the results presented in Bakshy et al. (2012) that
suggest an endorsement increases click-through rates by 5%. The rest of Table 6 presents

results which potentially explain this unexpected finding.
Column (3) distinguishes between the interaction between different message conditions
and their effectiveness given whether a message explicitly trying to promote social influence
was present. Since these ads did not display the friend’s name at the bottom, it should
not be so obvious to a viewer that the firm is explicitly trying to harness the social influence that results from the friend being a fan of the nonprofit. We recognize that there
may of course be some confusion at the mention of a friend when no name is displayed,
but this confusion should work against us rather than for us. The results suggest that
the negative reaction to the explicit mention of social influence is driven by the conditions
which had an endorsement present. Indeed, when the messages with exhortations about
social influence are controlled for, the baseline message which is captured in Column (3)
by SocialT argetingAndEndorsement produces a small but imprecisely measured positive
effect much in line with Bakshy et al. (2012). Column (4) presents the results where we
separate out the different message conditions by wording. Though we lose some precision,
the results do support the evidence in Column (3), that attempts to co-opt social influence
in the wording of ads backfires in particular for ads that couple this messaging with an
automated social endorsement.
18

Electronic copy available at: />

This novel finding that automated endorsements can backfire when coupled with wording
that explicitly tries to promote social influence is especially surprising, given papers such
as Salganik et al. (2006) which suggest that social influence is a central part of the online
economy. The results in Column (3) and (4) suggest that it was the combination of the
friend’s name and the mention of social influence which was particularly off-putting. What
is damaging is the combination of an advertiser making it explicit they are trying to harness social influence and the algorithmic social advertising message. We explore boundary
conditions for this effect in section 4.5.
4.3

Alternative Measures of Advertising Effectiveness

Table 7 checks the robustness of the finding in Table 5 that explicit messaging about social
influence in conjunction with a Social Ad unit backfires. There is always the possibility of
course that people clicked on the ads because they were annoyed or wanted to understand
more the extent of privacy intrusion rather than because the ads were actually effective. To
explore this, we estimate a specification where the dependent measure was the proportion
of clicks that became subscribers of the nonprofit’s Newsfeed. The results are reported
in Columns (1)-(3). We see that again social advertising appears to be more effective at
encouraging Facebook users to take the intended action as well as simply clicking. This is
evidence that people are not clicking on social ads due to annoyance at their intrusiveness
but instead are clicking on them and taking the action the ads intend to encourage them to
take. If anything the result suggests that explicit wording promoting social influence is more
pronounced in these point estimates.
Table 8 reports the results of using a dependent measure which, rather than looking at the
likelihood of a unique click given a certain number of Facebook users who saw an ad, instead
looks at the likelihood of any click, which is necessarily unique as a function of the total
impressions. In other words, each individual can potentially now see multiple impressions of

19

Electronic copy available at: />

an ad on a day. As might be expected, point estimates are lower given the lower baseline,
but in general the pattern is similar, which is reassuring that multiple impressions (or lack
of them) do not drive our results.
A final question is whether ads that are socially targeted and display endorsements are
more expensive for advertisers, thereby wiping out their relative effectiveness in terms of
return on advertising investment. We explore this in Table 9. There are several missing observations where there were no clicks that day and consequently there was no price recorded.
The results suggest that advertisers pay less for these clicks that are socially targeted. This
suggests that Facebook is not charging a premium for this kind of advertising. Though

Facebook shrouds in secrecy the precise pricing and auction mechanism underlying their
advertising pricing, this result would be consistent with a mechanism whereby advertisers
pay less for clicks if they have higher click-through rates. In other words, prices paid benefit
from an improved ‘quality-score’ (Athey and Nekipelov, 2011). The results also suggest that
advertisers pay less for demographically untargeted clicks, which is in line with previous
studies such as Beales (2010), and that they also pay less if they combine Social Ad units
with an untargeted demographic. Columns (2) and (3) shows that using explicit wording
that tries to promote social influence leads to higher cost per click. This is to be expected as
such campaigns underperform relatively and consequently the Facebook pricing algorithm
charges more for them.
We are not able to repeat the analysis done in Table 6 which used non-experimental
variation to look at the effects of endorsements being present or not for either subscriptions
or cost per click, as Facebook does not report the breakdown in terms of whether these were
derived from ad units that had endorsements or not. This is why our primary dependent
variable in this study is unique clicks as a function of the ad’s daily reach rather than these
other important ways of measuring ad effectiveness.

20

Electronic copy available at: />

4.4

Checking for Variation Introduced by the Ad Algorithm

As discussed, one issue we faced is that we do not observe how the Facebook ad algorithm
or external events affected the validity of our experiment. To address this, we performed
checks to investigate the validity of our results not from the point of view of whether we
were measuring the right dependent variable but instead to check whether there may be
challenges to the validity of our experiment.

First, we checked whether our ability to measure the relative effectiveness of the Social Ad
unit was affected by time variation in the ad-serving algorithm. As discussed in section 3, the
Facebook ad-serving algorithm itself may interfere with the running of a clean experiment,
as it automatically favors more successful campaigns. Table A2 in the appendix reports the
results of this investigation.
We investigated how the number of impressions for any one campaign, the general time
trend and whether it was the first day of the campaign affected our measurement of the
effectiveness of the Social Ad unit. The rationale for studying the moderating effect of impressions is that the number of impressions served across the campaigns varied across days,
potentially based on ad performance, and this may have distorted our results. The rationale for studying the moderating effect of time was to see whether the Facebook algorithm
absorbed more information over time about advertising performance and if so whether this
machine-learning process influenced our results. The rationale for studying the moderating
effect of the first day is that typically that is the time when the most information about ad
performance is collected by the ad algorithm, potentially introducing a non-linearity. Reassuringly, the results suggest that indeed our measurement of Social Ad unit effectiveness
is similar when we introduce moderating controls for time-varying influences that might be
introduced by differences in how the ad was served.
Second, we also checked whether there were any external influences in the outside world

21

Electronic copy available at: />

that could have influenced ad performance. For example, changes in awareness or interest in
either the general topic of donations or the specific topic of African education could provide
a challenge to the external validity of our results. To measure this, we downloaded data from
Google Trends that measured the relative popularity of searches for ‘African Education’ or
‘Donations’ in the period of our study.13 Table A2 in the appendix reports the results of
introducing this potential moderator, providing reassuring evidence that our measurement
of the effectiveness of the Social Ad was not influenced by changes in searches for either
donations or African education over time.
4.5

Ruling out Alternative Explanations

We then collected additional data to help rule out alternative explanations of our finding
that the explicit mention of social influence was undesirable in social ads.
One obvious potential explanation is that what we are measuring is simply that people
are unaware that what they are seeing is actually an ad, rather than part of Facebook. When
a nonprofit uses a message such as ‘Be like your friend’ then it becomes obvious that this
is an ad, and people respond differently. To test this, we persuaded the nonprofit to run a
subsequent experiment that allowed us to explicitly tease this apart. In this experiment, we
compared the performance of ads that said ‘Please read this ad. Help girls in East Africa
change their lives through education.’, and ads that simply said ‘Help girls in East Africa
change their lives through education.’14 This also allows us to test for a simple version of
a persuasive intent explanation for our result, which is that Facebook users would react
negatively to anything which revealed advertisers’ persuasive intent (Kaptein et al., 2011;
Boush et al., 2009).
If it were the case that Facebook users were simply mistaking socially targeted ads for
13

Google Trends data has been used in a variety of academic studies to measure how many people are
searching for specific items in order to better inform economic and even health forecasting (Choi and Varian,
2012; Carneiro and Mylonakis, 2009).
14
Recent research has questioned the use of the imperative in advertising copy, which is why we used
‘please’ (Kronrod et al., 2012)

22

Electronic copy available at: />

regular content and the explicit appeals to social influence stopped them making this mistake,
we would expect to also see a negative effect of wording that made it clear that the message
was an ad.
Column (1) indicates that this test was similar (despite the shortened time frame and
later date) to our earlier testing. Column (2) of Table 10 shows the results of separating
out the difference messages. It appears that adding ‘Please read this ad’ if anything helped
ad performance, which suggests that it was not the case that Facebook users were simply
mistaking socially targeted ads for content if there was no explicit message. Obviously,
though, the sample size here is small, making more definitive pronouncements unwise.
Another alternative explanation for our findings is that the messages referring to the
friend were poorly-written or unappealing. To test whether this was the case, we selected
an alternative set of users who might be expected to react in an opposite way to potential
presumptions of social influence in advertising messaging. The underlying idea here is that if
indeed the negative reaction to this emphasis on social influence in advertising messaging in
conjunction with an automated friend endorsement reflected some notion of friendship being
something that advertising wording should not try and coopt, then, if the person had already
themselves coopted brands into their social media persona, we would not see such a negative
reaction. Specifically, the nonprofit agreed to run test conditions identical to those in Table
5 for the people who expressed affinity with ‘Fashion’ goods on their Facebook profiles. The
Fashion category of users were chosen because typical models of social influence have focused
on fashion cycles (Bikhchandani et al., 1992). These models emphasize the extent to which
people who participate in fashion cycles receive explicit utility from conformity, even when
this conformity is provoked by a firm. In other words, they may find advertiser-endorsed
social influence more persuasive and advertiser attempts at emphasizing the power of social
influence more acceptable than the general population does.
As shown in Table 11, this group of users exhibits a different pattern to that exhibited by
23

Electronic copy available at: />

the general population. They appear to respond positively to social advertising. However,
strikingly, they reacted neutrally (or even positively - though this is imprecisely measured)
to advertising messages that emphasized social influence and the actions of the friend in the
ad copy. In other words, social advertising for this group worked even when the advertiser
explicitly embraced the potential for social influence. This result suggests that there may be
heterogeneity in consumer responses to the wording of social advertising messages depending
on their social media behavior. This is evidence against an alternative explanation for our
results in Table 5 based on these advertising messages which explicitly refer to the potential
for social influence being confusing or overly wordy, since they were effective for this subset of
fashion fans. In general, the results of Tables and 5 and 11 suggest that there is heterogeneity
in distaste for advertiser attempts to harness social influence given previous consumption
patterns, but that the average Facebook user who could be found in the broad swathe of
‘untargeted’ Facebook users in the US over 18, finds such attempts off-putting.

5

Implications

How helpful is data on social relationships when it comes to targeting and delivering advertising content and what messaging approach should firms take when trying to use social
relationships in advertising? This paper answers these questions using data on different ads
run by a non-profit on the large social networking site Facebook. We find evidence that
social advertising is indeed effective, but that it works less well if an advertiser attempts to
couple an automated endorsement from a fan with an advertising message which explicitly
promotes social influence.
When advertisers attempt to emulate or reinforce this social influence, consumers appear
less likely to respond positively to the ad. Speculatively, the results suggest that intrusive
or highly personal advertising is more acceptable if done algorithmically by a faceless entity
such as a computer, than when it is the result of evident human agency and coupled with

24

Electronic copy available at: />

explicit intent in marketing communications. Very speculatively, there is perhaps a parallel
with users of web-based email programs accepting an algorithm scanning their emails to
serve them relevant ads when the interception of emails by a human agent would not be
acceptable.
Our results suggest that social advertising works particularly well for untargeted populations, which may mean that social advertising is a particularly useful technique when
advertising to consumers outside the product’s natural or obvious market segment, since
there are less obvious ways of targeting in these settings. Using non-experimental variation
due to differences in privacy settings, we find that the majority of social advertising’s efficacy
appears to be because social targeting uncovers unobserved homophily between users of a
website and their underlying receptiveness to an advertising message.
There are of course limitations to our study. First, the nonprofit setting may bias our
results in ways that we cannot predict. Second, the aims of the nonprofit mean the outcome
measure we study is whether or not people sign up to hear more about the nonprofit, rather
than studying the direct effect of advertising on for-profit outcomes such as customers making purchases. Last, we rely on non-experimental variation to tease apart the relationship
between explicit messaging and the presence or absence of an endorsement. Notwithstanding these limitations, we believe that this paper makes a useful contribution in terms of
documenting when social advertising is effective and when it is not.

25

Electronic copy available at: />