Content uploaded by Marcus Schoegel
Author content
All content in this area was uploaded by Marcus Schoegel on Jan 15, 2019
Content may be subject to copyright.
Article
Detecting, Preventing, and Mitigating
Online Firestorms in Brand Communities
Dennis Herhausen, Stephan Ludwig, Dhruv Grewal,
Jochen Wulf, and Marcus Schoegel
Abstract
Online firestorms pose severe threats to online brand communities. Any negative electronic word of mouth (eWOM) has the
potential to become an online firestorm, yet not every post does, so finding ways to detect and respond to negative eWOM
constitutes a critical managerial priority. The authors develop a comprehensive framework that integrates different drivers of
negative eWOM and the response approaches that firms use to engage in and disengage from online conversations with com-
plaining customers. A text-mining study of negative eWOM demonstrates distinct impacts of high- and low-arousal emotions,
structural tie strength, and linguistic style match (between sender and brand community) on firestorm potential. The firm’s
response must be tailored to the intensity of arousal in the negative eWOM to limit the virality of potential online firestorms. The
impact of initiated firestorms can be mitigated by distinct firm responses over time, and the effectiveness of different disen-
gagement approaches also varies with their timing. For managers, these insights provide guidance on how to detect and reduce the
virality of online firestorms.
Keywords
message dynamics, online brand community, online firestorms, text mining, word of mouth
Online supplement: https://doi.org/10.1177/0022242918822300
More than 65 million firms leverage online brand communities
to connect with customers and achieve known performance
benefits, such as increased online reputation, brand patronage,
and customer spending (Baker, Donthu, and Kumar 2016; Hol-
lenbeck 2018, Kumar et al. 2016). However, online commu-
nities also engender significant risks of online firestorms—that
is, negative electronic word of mouth (eWOM) that receives
substantial support from other customers in a short period of
time (Pfeffer, Zorbach, and Carley 2014). Similar to prominent
online firestorm examples, such as #deleteUber and United
Airlines’ passenger removal incidents, less publicized negative
eWOM messages by dissatisfied customers also can go viral; a
single 466-word Facebook post by a disgruntled customer in
Odeon Cinemas’ Facebook brand community prompted more
than 94,000 likes, damaging the firm’s reputation and causing
it to lose thousands of customers (Dunphy 2012).
Detecting, preventing, and mitigating this virality of nega-
tive eWOM in online brand communities therefore constitutes
a critical managerial priority (Hewett et al. 2016), yet 72%of
firms rate their preparedness for online firestorms as “below
average” (Ethical Corporation 2012). Managers seem to have a
limited understanding of how to respond to negative eWOM
describing dissatisfactory consumption experiences (Wang and
Chaudhry 2018), nor do they know how to predict the evolution
of negative eWOM messages or address angered mass audi-
ences exposed to such negative eWOM. Lacking clear guide-
lines, firms continue to suffer damages from negative eWOM.
We aim to address this gap by identifying sources of firestorms
and detailing appropriate sequences for firm responses to neg-
ative viral content.
Extant marketing research has described the spreading of
word of mouth (WOM) as a contagious process, whereby recei-
vers “catch” others’ emotions through social transmission (Ber-
ger 2014). The relatively rare research that specifically
investigates negative WOM suggests that its contagiousness
Dennis Herhausen is Associate Professor of Marketing, KEDGE Business
School, France (email: dennis.herhausen@kedgebs.com). Stephan Ludwig is
Associate Professor of Marketing, University of Melbourne, Australia (email:
stephan.ludwig@unimelb.edu.au). Dhruv Grewal is Professor of Marketing,
Babson College (email: dgrewal@babson.edu). Jochen Wulf is Assistant
Professor of Information Management, University of St. Gallen, Switzerland
(email: jochen.wulf@unisg.ch). Marcus Schoegel is Associate Professor of
Marketing, University of St. Gallen, Switzerland (email: marcus.schoegel@
unisg.ch).
Journal of Marketing
1-21
ªAmerican Marketing Association 2019
Article reuse guidelines:
sagepub.com/journals-permissions
DOI: 10.1177/0022242918822300
journals.sagepub.com/home/jmx
primarily depends on the sender’s emotions (Berger and Milk-
man 2012; Heath, Bell, and Sternberg 2001) and the relation-
ship between senders and receivers (Brown and Reingen 1987;
Mittal, Huppertz, and Khare 2008). Yet few studies have
applied these valuable insights to an online brand community
context, so we identify sender and relational aspects pertinent
to firestorms of negative eWOM in online brand communities.
In addition to identifying sources of the spread of negative
eWOM, firms need to pinpoint how to respond (Chevalier,
Dover, and Mayzlin 2018). Services recovery research has
proposed several viable approaches to negative customer
experiences, including empathic and explanatory responses
(e.g., Bitner, Booms, and Tetreault 1990). In contrast with
traditional complaint channels, however, the online brand com-
munity makes customers’ complaints and the firm’s recovery
efforts visible. Therefore, beyond mitigating the complaining
customer’s unsatisfactory consumption experience, the firm
needs to craft a response that can minimize any negative effects
on the wider audience of online brand community members. By
investigating regulation strategies that can reduce receivers’
susceptibility to negative emotions (Gross and Thompson
2007), we investigate how firms should tailor their responses
to limit the virality of negative eWOM. In so doing, we do not
limit our assessment to a single response, because customers in
online brand communities often evaluate cross-message devel-
opments (Villarroel Ordenes et al. 2018). Therefore, in an
extension of Batra and Keller’s (2016) work, we consider how
sequences of firm responses might mitigate the virality of
online firestorms as they evolve.
With these empirical assessments of ways to detect, prevent,
and mitigate the virality of negative eWOM in online brand
communities, we offer three main contributions. First, we draw
on negative WOM research to investigate how sender and rela-
tional aspects aid in the detection of potential firestorms, then
specify how different levels of emotional arousal and the
strength of the senders’ structural ties and their similarity to
the online brand community relate to the virality of negative
eWOM. At an operational level, we establish a reliable, com-
puterized technique to determine the similarity of language use
within online brand communities. Second, our findings provide
insights into firms’ ability to prevent online firestorms by issu-
ing responses designed to engage with or disengage from cus-
tomers online. More explanatory responses are best for
negative eWOM messages containing above-average negative
high-arousal emotions; the effectiveness of disengaging
approaches varies. Third, we identify structured sequences of
different engaging responses across multiple firm messages as
a novel, actionable approach to mitigate the impact of evolved
online firestorms.
To achieve these aims, we first systematically delineate
sender and relational aspects that trigger greater virality of
negative eWOM. We then systematize extant knowledge on
common firm responses and contrast their effectiveness with
the arousal of negative eWOM and viable cross-message com-
position. We test our hypotheses on large-scale data, reflecting
negative customer posts from the online brand communities of
89 S&P 500 firms, which constitute potential online firestorms.
In the final section, we summarize the findings, discuss the
implications, and list some limitations.
Conceptual Foundations and Hypotheses
Extant marketing research primarily has focused on identifying
the presence and efficacy of positive eWOM (You, Vadakke-
patt, and Joshi 2015), but open customer communication also
bears the risk of unprecedented, rapidly discharged, large quan-
tities of negative eWOM (Pfeffer, Zorbach, and Carley 2014).
To cope with negative experiences and warn others, customers
share negative consumption experiences in online brand com-
munities. Often highly emotional, such posts may emerge, dif-
fuse, and dissolve quickly (Hauser et al. 2017). Similar to
positive eWOM, the extent to which other customers approve
of and share negative eWOM determines its virality and fire-
storm potential. Prestudy interviews with 16 social media man-
agers responsible for online brand communities suggest that
firms regard negative customer posts as evolved online fire-
storms if the firm’s initial response does not suffice to prevent
the negative eWOM from “catching fire” among other custom-
ers. Every like or comment that follows means that another
customer may be lost. At the outset, every negative post has
the potential to cause an online firestorm; not every post does
so. Compared with positive eWOM, negative WOM is trans-
mitted more often and is more influential (Hewett et al. 2016),
so firms must detect and adequately respond to negative posts
in online brand communities to avoid potential public debacles,
customer defections, and profit reductions (Pfeffer, Zorbach,
and Carley 2014).
Both product- and service-related WOM evaluations are
shared through a social transmission process, like emotional
contagion (Berger 2014). The Web Appendix contains an over-
view of studies that detail drivers of and firm responses to
eWOM. Various studies have indicated that eWOM conta-
giousness depends on the emotions conveyed, the structural
ties of the sender, or the perceived similarity between the sen-
der and receivers; however, the joint impact of these determi-
nants on the virality of negative eWOM is unknown. Moreover,
although some studies have examined the effectiveness of the
presence of firm responses, they do not differentiate the cir-
cumstances in which a certain type of response is more effec-
tive. Finally, firms might need to respond to the same negative
eWOM several times to resolve customers’ negative experi-
ences, and insight is lacking on how such responses should
be sequenced over time. To fill these critical research gaps,
we investigate ways to detect, prevent, and mitigate online
firestorms arising from negative eWOM messages.
Detecting Potential Online Firestorms
Conventional wisdom suggests that customers in online brand
communities first read about the cause of negative eWOM
messages and then decide whether to approve and share them.
However, faced with the information overload that tends to
2Journal of Marketing XX(X)
characterize communication exchanges on social media plat-
forms, customers might not elaborate in detail on the arguments
and instead could resort to heuristic processing (Hatfield et al.
2014). Accordingly, research has suggested that the relative
transmission of WOM is a result of the contagiousness of heur-
istics related to the sender’s message and the relationship
aspects between the sender and receivers (Brown and Reingen
1987; Heath, Bell, and Sternberg 2001; Mittal, Huppertz, and
Khare 2008).
For example, particularly expressive people seem to trans-
mit emotions effectively (Barsade 2002). Although emotions
are not verbal properties, the verbal use of emotional words
makes them relatively accessible and contagious. With
increased use of affective words in a post, it efficiently reveals
and makes accessible the intent or simplest raw feelings under-
lying the posting customer (Cohen et al. 2008). At a granular,
word-use level, increasing the number of negative emotion
words in eWOM translates directly into stronger behavioral
responses by message recipients (Ludwig et al. 2013). Even
if the content is unrealistic, more negative emotional messages
are shared more frequently (Blaine and Boyer 2018). However,
general negativity is a broad concept, and the influence of
negative emotional expressions might depend further on peo-
ple’s relative arousal levels (Russell and Barret 1999). For
example, in their study of urban legends, Heath, Bell, and
Sternberg (2001) investigate high-arousal disgust emotions
rather than just general negative emotions. Similarly, online
firestorms may be more likely to arise from high-arousal
(e.g., “This is so frustrating”) rather than low-arousal (e.g.,
“This is disappointing”) negative eWOM. Berger and Milkman
(2012) show that New York Times newspaper stories that
include more intensive high-arousal emotions (e.g., fear/anxi-
ety, anger) prompt emailed shares to others more frequently
than stories with more intensive low-arousal emotions (e.g.,
sadness). Thus, rather than simply being one-dimensional, the
contagiousness of emotionally charged negative eWOM in
online brand communities may depend on the level of arousal.
Therefore, we posit,
H
1
:The intensity of high-arousal emotion words in neg-
ative eWOM messages relates to greater virality in online
brand communities compared with the intensity of low-
arousal emotion words.
The decision to approve or share eWOM also depends on the
relationship and relational cues between the sender and recei-
ver (Berger and Schwartz 2011). Emotional contagion theorists
cite the importance of interpersonal relations that enable mes-
sage recipients to evaluate others and devise appropriate
responses (Barsade 2002). Marketing research on WOM sug-
gests that tie strength and perceptions of similarity are two
primary relational cues that cause receivers to regard senders
as more proximate (Brown and Reingen 1987).
Tie strength is relevant in various information-sharing con-
texts; it refers to both the frequency of communication and the
importance attached to the relations (Baker, Donthu, and
Kumar 2016; Brown and Reingen 1987; Risselada, Verhoef,
and Bijmolt 2014). Despite considerable debate about the rela-
tive advantages of weak and strong ties, researchers commonly
agree that strong ties increase the likelihood that social actors
will share sensitive information (Rapp et al. 2013) and engage
in collective action (Obstfeld 2005). Measures of tie strength
rely on a range of variables (Mittal, Huppertz, and Khare 2008;
Rapp et al. 2013), including frequency of contact (Risselada,
Verhoef, and Bijmolt 2014). Weak ties reflect members of a
community who interact less frequently with each other,
whereas strong ties describe relationships of members who
interact frequently (Burt 1987). Frequent, positive encounters
typically (if not always) lead to stronger structural ties (Mittal,
Huppertz, and Khare 2008) and increase opportunities to trans-
mit opinions (Frenzen and Nakamoto 1993). This assumption is
also in line with Burt’s (1987) suggestion that the more frequent
and empathic the communication is between two users, the more
likely that one user’s opinion will influence the other user’s
opinion. Strong structural ties, as characterized by more frequent
interactions, in turn increase imitative behavior within networks
(McFarland, Bloodgood, and Payan 2008). Members with an
exceptionally great number of ties within a community often act
as opinion leaders, who influence purchase decisions and prod-
uct adoptions (Katona, Zubcsek, and Sarvary 2011). Certainly,
the potential of well-connected customers to influence others in
a brand community is likely to be stronger than the potential of
less connected members in the same community (Goldenberg
et al. 2009). If the member who posts negative eWOM has
stronger structural ties in the online brand community, the fire-
storm potential of the post thus should be greater.
H
2
:Stronger structural ties between the sender of nega-
tive eWOM and the receiving online brand community
relate to greater virality.
In addition, perceived similarity (or homophily perceptions;
Brown and Reingen 1987) between the sender and customers in
online brand communities may relate to the virality of negative
eWOM messages. Although perceptions of similarity are not
required for contagion to occur, they can act as qualifiers of
information relevance (Hatfield et al. 2014). For example, Aral,
Muchnik, and Sundararajan (2009) find that perceptions of
similarity between customers explain more than half of the
effect of behavioral contagion on new product adoption.
Although interactions in online brand communities tend to be
relatively anonymous, studies drawing on psycholinguistic
research suggest that perceptions of similarity in computer-
mediated settings are an automatic outcome of a linguistic style
match (LSM). The similar use of function words—or LSM
between two or more conversation partners—represents a form
of psychological synchrony that elicits perceptions of similar-
ity, approval, and trust in receivers (Ireland and Pennebaker
2010). Just like conversation dyads, communities may develop
a distinctive collective communication style (Fayard and
DeSanctis 2010). An individual customer’s alignment with a
common, community-level communicative style may elicit
Herhausen et al. 3
similarity perceptions and in turn influence the approval like-
lihood by the collective (Gumperz and Levinson 1996). Lud-
wig et al. (2013) confirm that the congruence of a customer
review with the typical linguistic style demonstrated by a prod-
uct interest group on Amazon influences other customers’ pur-
chase behavior. Accordingly, negative eWOM that matches the
typical linguistic style of an online brand community (i.e.,
evokes perceptions of similarity) should induce greater online
firestorm potential.
H
3
:Closer LSM between the sender of negative eWOM
and the receiving online brand community relates to
greater virality.
Preventing Potential Online Firestorms
The growing influence of online evaluations on customer beha-
vior has increased managerial and research interest in firm
recovery strategies that can reduce the contagiousness of neg-
ative eWOM (Ma, Sun, and Kekre 2015). Recovery in the
context of online brand communities is unique though, in that
the customer’s complaint and a firm’s recovery efforts are
visible to thousands or even millions of other customers. Effec-
tive recovery thus must (1) adequately restore relationship
equity to the complaining customer and (2) prevent the nega-
tive eWOM message from spreading to other customers in the
online brand community. The viability of common recovery
approaches—offering an apology, compensation, responding
empathically, or providing explanations—has been investi-
gated mainly in bilateral firm–customer communication con-
texts (e.g., Hill, Roggeveen, and Grewal 2015). To gain further
insight into the suitability of these approaches for reducing the
contagiousness of emotions in negative eWOM, we turn to
theory about emotion regulation strategies (Gross and Thomp-
son 2007) and propose that firm responses to negative eWOM
might be classified as disengaging or engaging.
A disengaging approach to emotion regulation implies
reactinginwaystoavoidorblock elaboration, rather than
preparing an adaptive response (Sheppes et al. 2011). Observa-
tions and anecdotes suggest that avoidance and nonresponse is
the poorest approach to regulating the virality of negative
eWOM. As an alternative, firms might try to halt an ongoing
public online conversation by suggesting a communication
channel change (e.g., “Please contact our service center”). Such
a channel change suggestion might be effective for pushing
customers to the right channels (Ansari, Mela, and Neslin
2008), but it is unclear how other online brand community react
to being excluded from the continued conversation. The effec-
tiveness of offering compensation to the complaining customer
also is uncertain. Some service research has suggested that
halting further elaborations is an effective recovery strategy
(e.g., Bitner, Booms, and Tetreault 1990), yet Grewal, Roggev-
een, and Tsiros (2008) find that compensation is not always
effective and instead may depend on other response features.
Active engagement with negative eWOM messages instead
might be more appropriate (Wang and Chaudhry 2018). Ser-
vice recovery literature has outlined two primary response
approaches that represent active firm–customer conversational
elaboration (Hill, Roggeveen, and Grewal 2015): empathic or
explanatory. To express empathy, a spontaneous affective
response (Hoffman 1977), a firm might sympathize (e.g., “We
understand that you are unhappy”) or shift to a positive outlook
(e.g., “We hope you have a better experience next time”).
Highly empathic responses may enhance the complainant’s and
online brand community’s perceptions of interactional justice
and signal politeness and courtesy, which may reduce the vir-
ality of negative eWOM. An engaged firm response also might
include substantiated explanations, and the number of reasons
offered has more influence than the actual content of those
reasons on decision outcomes (Seibold, Lemus, and Kang
2010). When firms provide more substantiated arguments, it
may enhance perceptions of response quality and effort among
brand community audiences (e.g., “We could not assist you
quickly because the store was extremely busy”). By providing
more explanation, firms might enhance evaluations of their
recovery efforts (Bitner, Booms, and Tetreault 1990).
However, in line with cognitive appraisal theory and an
affect infusion model, Homburg, Grozdanovic, and Klarmann
(2007) posit that an affective approach, such as empathy, is
more effective in affect-intensive environments characterized
by social interactions and spontaneous decisions, such as online
brand communities. In general, then, more empathic responses
might be better suited to regulating the contagiousness of neg-
ative eWOM. According to the affect infusion model, the rela-
tive impact of cognitive responses, such as explanations,
increases with stronger affect and higher involvement (Forgas
1995). Research on emotion regulation strategies further indi-
cates that some stimuli may be too emotionally intense for an
empathic response to suffice, and instead, receivers may seek
explanations to reappraise the situation (Gross 2002). The more
contagious the emotions in a negative eWOM message, the
more attention customers will pay to the message and the stron-
ger their expectations about what needs to be done to remedy
the situation (Hess, Ganesan, and Klein 2003). In such situa-
tions, customers are more likely to engage in deliberate pro-
cessing of negative eWOM by cognitively reappraising the
situation; that is, they consider more information and perform
more intricate evaluations of the explanations (Lazarus 1991).
Empathic responses may help shift the attention of consumers
who experience low-arousal emotions, but firms might better
mitigate the virality of high-arousal emotions by offering more
explanations. Thus, the relative effectiveness of firm responses
for preventing online firestorms may be contingent on the
intensities of the high- and low-arousal levels in the negative
eWOM message:
H
4
:More explanation, rather than more empathy, in firm
responses is better suited to contain negative eWOM with
more intensive high-arousal emotions.
H
5
:More empathy, rather than more explanation, in firm
responses is better suited to contain negative eWOM with
more intensive low-arousal emotions.
4Journal of Marketing XX(X)
Mitigating Evolved Online Firestorms
Through observational learning processes, as an online fire-
storm evolves, and other members support the negative
eWOM, its perceived reliability should increase (Dholakia,
Basuroy, and Soltysinski 2002). Thus, customers pay even
more attention to the negative eWOM and form revised expec-
tations about what needs to be done to remedy the situation.
Therefore, beyond the compositional elements of individual
firm responses, when negative eWOM evolves into an online
firestorm, multiple firm responses become necessary to miti-
gate its detrimental impacts.
As Batra and Keller (2016) suggest, online messages build
on one another, and thus their sequence can determine their
success in terms of persuading customers, building brand
equity, or driving sales. Villarroel Ordenes et al. (2018) find
that posting the same (vs. mixed) consecutive brand message
decreases (vs. increases) customers’ engagement. Considera-
tions of cross-message dynamics in firm responses may
advance understanding of how to mitigate evolved online fire-
storms too. For example, by empathically sympathizing with
the customer in the first reply, then issuing a second, comple-
mentary response that provides explanations, the firm might
reduce the overall virality of the negative eWOM, compared
with a situation in which it repeats its offer of sympathy in the
second response, which might cause frustration (Kocielnik and
Hsieh 2017). We predict that such cross-message sequencing
should mitigate the virality of negative eWOM to a broad
customer audience (Batra and Keller 2016).
H
6
: Consecutive firm responses with varying rather than
repeated (a) empathic intensity and (b) explanatory inten-
sity are better suited to mitigating evolved online
firestorms.
Methods
Sampling Frame and Text Analysis Procedure
We used Facebook’s Application Programming Interface (API)
and processed detailed information on potential online fire-
storms from the official Facebook brand communities of all
U.S. firms listed on the S&P 500 between October 1, 2011 (the
introduction of the timeline feature), and January 31, 2016 (the
introduction of emojis). We chose this setting for three reasons.
First, it is common for customers to use Facebook to interact
with firms in their brand communities and to complain through
this channel. Second, unlike other rating and review sites that
encourage only customers to share their views, firms actively
participate in Facebook conversations and respond to customer
posts (Schweidel and Moe 2014). Third, in line with previous
research on online brand communities (e.g., Lee, Hosanagar,
and Nair 2018; Relling et al. 2016), the count of likes and
comments indicates the degree to which others approve and
share a message and provides an objective measure of virality.
We selected all firms that target private customers, have an
official Facebook brand community, and allow user posts in
their community (see the Web Appendix). Because of our focus
on negative eWOM, we analyzed text-based features to deter-
mine which posts were negative in two steps. We first applied
the R Quanteda package (Benoit et al. 2018) using Linguistic
Inquiry and Word Count (LIWC) text-mining dictionaries to
derive the intensity of positive- and negative-emotion words in
each post (for more details, see Humphreys and Wang [2018]).
We then applied the Stanford Sentence and Grammatical
Dependency Parser (Stanford Natural Language Processing
Group 2014) to subdivide each post into sentences and identify
dependencies between emotion words and negations (i.e.,
bigrams). The parser first identifies the presence of an emotion
word and then, in cases of negation, automatically assesses
whether there is a grammatical relationship (e.g., in the sen-
tence “The service was not nice,” the negation “not” is gram-
matically related to the adjective “nice” and therefore reverses
it). Negated positive emotion words were counted toward nega-
tivity and negated negative emotions words toward positivity.
If a customer post is more negative than positive overall, we
count it as a potential online firestorm.
We excluded 48,480 posts that contained a video, picture,
event, or external link, because we cannot control for this
external content. Furthermore, we excluded 140 posts with
fewer than three words, because every full thought requires
at least a subject, verb, and object to be understandable for
receivers. On Facebook, zero counts of likes and comments
from other customers might occur for two reasons: (1) The post
may have been viewed by other customers but prompted no
reaction, or (2) the post may not have been displayed to other
customers. Thus, we excluded 128,681 posts with no customer
reactions, because we are interested in the inflation of virality.
We also replicated our analysis with the sample including posts
with no customer reactions, and we report the results in the
Web Appendix.
With these restrictions, our final sample counted 472,995
negative customer posts in English across 89 online brand
communities. The posts averaged 99 words (SD ¼121.44,
ranging from 3 to 8,121 words) and received 2.95 likes and
comments on average (SD ¼59.22, ranging from 1 to 37,760
likes and comments). Included in this final sample are both
well-publicized incidents (e.g., customer post complaining that
a police officer was prohibited from using a coffeehouse chain
bathroom in September 2015) and less drastic cases in which
complaints were supported by a only small number of other
customers. Notably, for the well-publicized incidents, newspa-
pers and news portals cited the original Facebook post (e.g., “A
Facebook user reported that a police officer was prohibited
from using a bathroom” [www.snopes.com]). Figure 1 displays
illustrative examples of online firestorms and firm responses.
In terms of timing, 78%of the online firestorms emerged and
dissolved within a day, such that the last comment was posted
within 24 hours of the initial negative post, and 93%of firm
responses arrived within one hour. This short timeframe is a
unique feature of online firestorms that differentiates our study
from previous research on product-harm crises (e.g., Cleeren,
Dekimpe, and Van Heerde 2017).
Herhausen et al. 5
Example A: Illustrative example of an online firestorm that erupted
“We are aware of this situation,
and it is certainly not in line with
the experience we want any of our
customers to have in our stores.
We are taking all necessary steps
to ensure this doesn't happen in
the future.”
“So I walk in in full uniform and ask
the young blond liberal behind the
counter if I could use their public
bathroom for which I need a key code
and she states in a loud voice, so all the
other customers can hear that the
bathroom is for paying customers
only… It’s hip for this generation to
berate and totally disrespect cops in
front of the public… I hope my fellow
brothers and sisters in blue see this and
know that we have each other… and
not to patronize that coffeehouse”
Customer posted a high-arousal
negative complaint in the brand
community of a coffeehouse chain
6,604 likes,
2,347 comments,
20,879 shares
Other brand community members users support the complaint,
and an online firestorm takes place that spills over to other media
Firm responded late to the
customer post after the online
firestorm already took place
Example B: Illustrative example of a potential online firestorm that was successfully prevented
“Our apologies for any
inconvenience caused. We are
sorry for the mistake caused by
an update in our system. Hope
you already got reimbursed. If
you need anything furt her please
let us know.”
“I’m very disappointed. You tried
charging me two months late payments.
I knew I wasn’t late so I checked my
bank statements and was clearly not
late. Called the hotline and the gal says,
Oops! Our mistake. We'll adjust your
bill. I hung up only t o realize last
month was the same mistake and I paid
it! Now on the phone a second time but
again on hold for 15 minutes of course!
Shame on you!”
7 likes,
1 comment ,
0 shares
Some users start to
support the complaint
Firm responded with an
apology and explained the
reason for the service failure
“Thank you for your response
and explaining the reason for this
failure. I hope the system is fixed.
My reimbursement should arrive
soon. But please please also fix
your hotline…”
No further support of the complaint after the
firm response
The customer signaled an
understanding for the situation
and thanks the company
Example C: Illustrative example of an online firestorm that was successfully mitigated
“We are sorry for the
inconvenience. Please contact our
customer service team for any
requests regarding your luggage.
The Social Media Team”
“I write this completely outraged for
the treatment we received from part of
your airline. Since the beginning
everything went wrong. I traveled with
my family from Quito-Ecuador to New
York in this airline, but our bags got
sucked in Miami. They were sent to us
two days after, in which we had no
clothes, personal items or medicines. I
will not fly again with your airline!”
User posted a high-arousal
negative complaint in the
brand community of an airline
59 likes,
25 commen ts,
4 shares
Other users support
the complain
Firm responded with empathy for
the customer without explanation
or offering furthe r help
Despite the response, other users
continue to support the complain
After ongoing user reactions, the
second firm response explains the
situation and offers compensation
“Again, we apologize for your late
baggage caused by a damaged
plane door. Please mind that we
compensate you for any extra
costs caused and offer you a
compensation through our
customer service team.”
Customer posted a high-arousal
negative complaint in the brand
community of a service provider
6,604 likes,
2,347 comments,
20,879 shares
7 likes,
1 comment,
0 shares
21 likes,
9 comments,
2 shares
38 likes,
16 comments,
2 shares
No further
support
Figure 1. Illustrative examples of online firestorms and firm responses.
Notes: Posts are edited to exclude company names and customer names.
6Journal of Marketing XX(X)
Measurement
Table 1 contains the operationalizations and sources of all
variables, along with our rationale for including the control
variables, and Figure 2 displays hypotheses and the measure-
ment approach. The key variable of interest is the virality of
negative eWOM, measured as the total number of likes and
comments a post receives from other customers.
1
The total
number of likes and comments correlate closely (r ¼.81),
justifying their use as a composite variable. Because of
community-level differences in virality (some online brand
communities feature thousands of posts every day; others just
a few), we use the deviation from the community average.
Then, noting the data range and extreme values of virality (see
the Web Appendix), we add a constant to have only positive
values and apply a logarithmic transformation. To investigate
how firms prevent and mitigate online firestorms, we include
only likes and comments posted after the respective firm’s
response (Wang and Chaudhry 2018), which ensures that vir-
ality has been influenced by the firm response. Importantly, the
API does not allow us to capture time stamps for likes. How-
ever, comments are time-stamped and known to evolve simul-
taneously with likes over time (Rieder et al. 2015). Therefore,
we use the amount of comments following a firm response to
approximate the number of likes.
We measure the intensity of high and low arousal for each
negative post with computerized text analysis. In a top-down
manner (Humphreys and Wang 2018), we compared each word
in a message with predefined emotion word categories. We
then calculated an intensity score per emotion word category:
the proportion of total words that match each dictionary. In line
with the main four negative emotion types in the circumplex
model (Russell and Barret 1999), we classified the proportion
of word use related to fear/anxiety, anger, and disgust as the
intensity of high-arousal negative emotions and the proportion
of sadness as the intensity of low-arousal negative emotion. We
used existing LIWC dictionaries for fear/anxiety, anger, and
sadness (Pennebaker et al. 2015), but we needed to develop a
new dictionary to derive disgust. We provide this dictionary
and details on its development in the Web Appendix. To vali-
date the new dictionary, we compared statistical differences in
arousal levels, according to an extremity measure from the
Evaluative Lexicon 2.0 (EL 2.0; Rocklage, Rucker, and Nordg-
ren 2018), between our disgust words and words classified as
representing low negative arousal (i.e., sadness). We find a
significant difference (F ¼7.57, p<.01), with extremity mean
scores of 3.24 and 2.89 for disgust and low arousal, respec-
tively. This result confirms that our disgust dictionary matches
the EL 2.0 measure for expression extremity.
Strength of structural ties has been measured using different
variables, including subjective and objective measurements.
Because we cannot collect perceptions of tie strength across
the millions of brand community users, we followed Risselada,
Verhoef, and Bijmolt (2014) and operationalized strength of
structural ties (SST) as “the frequency of communication”
(Brown and Reignen 1987, p. 356). Formally, SST for cus-
tomer i posting negative eWOM at time t in community c is:
SST ic ¼X
t1
t¼0
Received Likest
ic þReceived Commentst
ic
þReceived Sharest
ic þX
t1
t¼0
Likes Givent
ic
þComments Given t
ic
ð1Þ
where t 1 is the entire period prior to the post at time t, and
SST
ic
is the sum of likes
c
, comments
c
, and shares
c
that cus-
tomer i received from others in the brand community c before
the post at t, as well as the sum of likes
i
and comments
i
the
customer gave to others in the brand community c prior to
the post at t (all calculated based on the comment timestamp).
The API does not allow us to identify customers who share a
certain post, due to privacy restrictions.
We derived the degree of LSM between customer i posting
negative eWOM at time t with the receiving brand community
c in three steps. First, we mined the use intensity of each of the
nine function word categories j separately in focal customer
i’s message and across all customer messages (negative and
positive) in the brand community c posted in the previous
three months in response to the focal negative eWOM post
(moving community average). Second, the degree of similar
use intensity LSM of each function word category (FW
j
)by
customer i posting the negative eWOM into community c
comes from the formula:
LSMjic ¼1jFWjiFWjic j
FWjiþFWjic þ:0001
ð2Þ
Third, by aggregating all nine LSM scores with equal
weights, we obtain an LSM score bound between 0 and 1, and
scores closer to 1 reflect greater degree of communication style
matching between customer i and the online brand community c.
We measured the intensity of empathy, or the degree of
spontaneous affective response (Hoffman 1977) a firm pro-
vided, as the proportion of affect words in the response text,
according to the LIWC dictionary for affect. Using the LIWC
dictionary for causal expressions, we also measured the inten-
sity of explanation in a firm’s responses. We then measured
variations in the response sequences as the standard deviation
in empathic and explanatory intensities across all firm
responses.
Following prior research, we account for multiple control
variables that might influence the virality of negative eWOM
(see Table 1). Firm-related aspects that influence eWOM
include industry membership, brand familiarity, and brand rep-
utation. At the online brand community level, we account for
1
Likes and comments are more common indicators of virality than shares on
Facebook. Of 472,995 negative eWOM incidences, 274,155 (52%) received at
least one like, and 235,545 (50%) received at least one comment, but only
9,434 (2%) were shared at least once by other customers (see the Web
Appendix). We also replicated the analyses with the number of likes,
comments, and shares as separate measures of virality (see the Web Appendix).
Herhausen et al. 7
Table 1. Operationalization and Sources of All Variables.
Variable and Time Operationalization Source Rational for Inclusion (Controls) and Related Studies
Dependent Variable
Virality Combined sum of likes and comments the post
received from other customers any time after it
was posted (Virality
t_1
; brand community-
centered and log-transformed). To investigate
how firms prevent and mitigate online firestorms,
we only considered virality after the first firm
responses (Virality
t_2
) or after the last firm
responses (Virality
t_3
).
Facebook API De Vries, Gensler, and Leeflang (2012); Lee,
Hosanagar, and Nair (2018); Relling et al. (2016);
Stephen, Sciandra, and Inman (2015)
Post Predictors
Intensity of high
arousal
LIWC dictionaries “anx” and “anger,” new dictionary
“disgust” for the focal post (number of matching
words in the post, expressed as percentage of
total words).
Text mining Berger and Milkman (2012); Hewett et al. (2016)
Intensity of low
arousal
LIWC dictionary “sad” for the focal post (number of
matching words in the post, expressed as
percentage of total words).
Text mining Berger and Milkman (2012); Hewett et al. (2016)
Strength of
structural ties
Frequency of communication with the online brand
community before the post (see Formula 1 in the
text).
Facebook API Risselada, Verhoef, and Bijmolt (2014)
LSM Degree of communication style matching with the
online brand community before the post (see
Formula 2 in the text).
Text mining Ireland and Pennebaker (2010); Ludwig et al. (2013)
Firm Response
Intensity of
empathy
LIWC dictionary “affect” for firm responses (number
of matching words in the response, expressed as
percentage of total words).
Text mining Fehr, Gelfand, and Nag (2010)
Intensity of
explanation
LIWC dictionary “cogproc” for firm responses
(number of matching words in the response,
expressed as percentage of total words).
Text mining Seibold and Meyers (2007)
Variation in firm
responses
Variance in the proportion of empathic and
explanatory words across all firm responses.
Text mining Villarroel Ordenes et al. (2018)
Firm Controls
GICS Global Industry Classification Standard: consumer
discretionary (49%) and consumer staples (21%)
versus other (30%).
Standard &
Poor’s
The amount of negative eWOM may depend on the
industry (Stephen, Sciandra, and Inman 2015).
Brand familiarity Familiarity of each firm (on a scale from 0% to 100%). Reputation
Institute
Low brand familiarity may lead to less engagement
with negative eWOM (Baker, Donthu, and Kumar
2016).
Brand reputation Reputation perceptions of each firm (on a scale from
0% to 100%).
Reputation
Institute
Support for negative eWOM may depend on the
firm’s reputation (Baker, Donthu, and Kumar
2016).
Brand Community
Controls
Brand community
size
Number of page likes for the online brand
community (in millions of page likes).
Facebook API Greater brand community size gives negative
eWOM a larger audience.
Brand community
attentiveness
Average number of likes and comments per
customer post in each online brand community
(for both positive and negative posts).
Facebook API More attentive members may be more susceptible
to negative eWOM (Hatfield, Cacioppo, and
Rapson 1994).
Brand community
expressiveness
Average of LIWC dictionary “affect” for all posts in
the online brand community (percentage of total
words).
Text mining More expressive members may be more susceptible
to negative eWOM (Hatfield, Cacioppo, and
Rapson 1994).
Average firm
engagement
Average number of firm responses per customer
post in each online brand community (for both
positive and negative posts).
Facebook API Firm engagement may stimulate negative eWOM
(Homburg, Ehm, and Artz 2015).
Average tie
strengths
Average tie strength of each customer who posted,
commented, or liked within the online brand
community.
Facebook API Average tie strengths may increase the effectiveness
of negative eWOM (Katona, Zubcsek, and Sarvary
2011).
(continued)
8Journal of Marketing XX(X)
community size, member attentiveness and expressiveness,
firm engagement frequency, average structural tie strength
among members, and variance in linguistic style. Post-related
aspects include the number of competing inputs at the time of
the post, the sentiment of the previous post, post length, post
complexity, and the frequency with which the customer had
complained on Facebook in the past. Furthermore, firm
response–related aspects include whether a firm responds or
not and the firm response time. We used dummy variables to
account for whether a firm offered an apology, compensation,
or suggested a communication channel change; a firm can use
more than one response approach in the same message for this
measure (e.g., combining apology and compensation). Finally,
as fixed effects, we account for the year and month to control
for seasonality in user activity and policy changes (see the Web
Appendix). We also include fixed effects for weekends and
time of day (Kanuri, Chen, and Sridhar 2018). The Web
Appendix reports correlations and descriptive statistics. Firms
responded at least once to 331,370 out of 472,995 negative
posts, yielding an average response rate of 70%. Across all firm
responses, suggesting a channel change (61%) and apologizing
(53%) were most commonly used while compensation was
used less often (3%). The degree to which explanations were
offered (8%of the time) was slightly more than the use of
empathy (6%). We found that 15,762 negative posts achieved
above-average virality and got multiple firm responses, sug-
gesting that 3%of potential online firestorms evolved during
the period of observation. The Web Appendix reveals the evo-
lution of the number of potential firestorms, average virality,
and average firm response rates over time. While the number of
potential online firestorms increased during the study period
(and we control for this increase with time-related fixed-
effect), both the average virality and the average percentage
of firm responses are rather stable over time.
Table 1. (continued)
Variable and Time Operationalization Source Rational for Inclusion (Controls) and Related Studies
Variance in
linguistic style
Variance in LSM of all customer posts in each online
brand community (for both positive and negative
posts).
Text mining Variance in linguistic style may decrease the
effectiveness of negative eWOM (Ludwig et al.
2014).
Post Controls
Competing inputs Number of other posts in the online brand
community on the day of the focal negative
customer post (both customer and firm posts).
Facebook API Exposure to competing stimuli may decrease the
virality of emotions (Coenen and Broekens 2012).
Sentiment
previous post
LIWC dictionaries “posemo” minus “negemo” for
the previous customer post in the online brand
community.
Text mining A preexisting mood state may increase the virality of
emotions (Coenen and Broekens 2012).
Post length Average words per sentence. Text mining Longer posts may convey more information and thus
increase virality (Berger and Milkman 2012).
Post complexity Average words with more than six letters per
sentence.
Text mining Posts that are more complex suggest eloquence in
writing which may increase virality (Va
´squez
2014).
Negation in post LIWC dictionary “negate” (in percentage of total
words).
Text mining Posts with negation may be more difficult to
understand and thus decrease virality.
Previous
complaints
Number of negative posts from the same customer
prior to the focal negative post.
Facebook API A high frequency of complaints from the same
customer may decrease virality (Ma, Sun, and
Kekre 2015).
No firm response No firm response on the negative customer post
(dummy coded).
Facebook API Virality may increase if members believe the firm is
ignoring them (Homburg, Ehm, and Artz 2015).
Firm response
time
Time stamp of negative customer post minus time
stamp of first firm response (converted to hours).
Facebook API A faster response should be beneficial for firms and
thus may decrease virality (Homburg, Ehm, and
Artz 2015).
Firm Response
Controls
Compensation Newly developed dictionary; see the Web Appendix
(dummy coded).
Text mining Offering compensation may satisfy the complaining
customer and decrease virality (Bitner, Booms,
and Tetreault 1990).
Apology Newly developed dictionary; see the Web Appendix
(dummy coded).
Text mining Apologizing may please the complaining customer
and decrease virality (Bitner, Booms, and
Tetreault 1990).
Channel change Newly developed dictionary; see the Web Appendix
(dummy coded).
Text mining Evoking a channel change may take the conversation
away from the brand community and decrease
virality (Ansari, Mela, and Neslin 2008).
Notes: We use time dummies to control for year, month, weekend, and time of day.
Herhausen et al. 9
Modeling Approach
The incidences of negative eWOM are nested within the online
brand communities, and thus the negative posts and firm
responses might be interdependent. To determine whether a
multilevel approach is warranted, we first conducted a one-
way analysis of variance with random effects to reveal any
systematic between-group variance in the virality of negative
eWOM. We find significant between-group variance (w
2
(88) ¼
818,729, p<.01). In addition, the design effect of 36.74 sug-
gests that a multilevel structure is possible (M´
uthen and Satorra
1995). The maximum variance inflation factor score across all
models is 3, indicating no potential threat of multicollinearity.
We specified a series of separate hierarchical models, with
parameters at the post and the firm/brand community level,
using full information maximum likelihood estimation and
grand mean-centering. Virality, our focal outcome measure,
is operationalized differently across these models along three
time periods. Virality
t_1
is the total number of likes and com-
ments from other customers any time after posting at t. Virali-
ty
t_2
is the number of likes and comments from other customers
any time after the first firm response, and Virality
t_3
is the
number of likes and comments from other customers any time
after the last firm response. We provide more detailed explana-
tions for each variable in Table 1. Thus, we assess the predic-
tors of virality for all 472,995 negative posts across the 89
brand communities as follows:
Viralityt 1
ic ¼g00 þg0104Firm Controlsc
þg0510Brand Community Controlsc
þg1117Post Controlsic
þg18Dum No Firm Responseic
þg1923Post Predictorsic
þg2443Dum Ti mingic þu0c þric;
ð3Þ
where t_1 is the time period after the time t of customer i’s post
in brand community c;
Viralityt 1
ic ¼the combined sum of likes and comments
post i receives from other customers in community c
any time after it was posted (brand community-
centered and log-transformed);
Firm Controlsc¼community-specific controls using the
Global Industry Classification Standard: GICS Con-
sumer Discretionary
c
, GICS Consumer Staples
c
,
Brand Familiarity
c
, and Brand Reputation
c
;
Brand Community Controlsc¼Brand Community Size
c
,
Brand Community Attentiveness
c
, Brand Community
Expressiveness
c,
Firm Engagement
c
,AverageTie
Strengths
c
, and Variance in Linguistic Style
c
;
Post Controlsic ¼Competing Inputs
ic
,Sentiment
Previous Postic , Post Length
ic
, Post Complexity
ic
,
Negation in Post
ic
, and Previous Complainsic;
First
Firm
Response
Time
Stage Detecting Potential Online Firestorms Preventing Potential Online Firestorms Mitigating Evolved Online Firestorms
Virality Viralityt_1 = total number of likes and comments from
other customers at any time after posting at t
Viralityt_2 = the number of likes and comments from other
customers at any time after the first firm response
Viralityt_3 = the number of likes and comments from other
customers at any time after the last firm response
Sample All posts with and without firm response
(472,995 negative posts)
All posts with at least one firm response
(331,370 negative posts)
All posts with more than one firm response and above-
average virality (15,762 negative posts)
Hypotheses H1: The intensity of high-arousal-emotion words in
negative eWOM messages relates to greater virality
in online brand communities, compared with the
intensity of low arousal emotion words.
H2: Stronger structural ties between the sender of
negative eWOM and the receiving online brand
community relate to greater virality.
H3: Closer LSM between the sender of negative eWOM
and the receiving online brand community relates to
greater virality.
H4: More explanation, rather than more empathy, in firm
responses is better suited to contain negative eWOM
with more intensive high-arousal emotions.
H5: More empathy, rather than more explanation, in firm
responses is better suited to contain negative eWOM
with more intensive low-arousal emotions.
H6: Consecutive firm responses with varying rather than
repeated (a) empathic intensity and (b) explanatory
intensity are better suited to mitigating evolved
online firestorms.
Results Table 2 Table 3 Table 4
Subsequent
Firm
Response(s)
Last
Firm
Response
Viralityt_1
Viralityt_2
Viralityt_3
Negative
Customer
Post
Figure 2. Hypotheses and measurement approach.
10 Journal of Marketing XX(X)
Dum No Firm Responseic ¼1 if there is no firm response
at any time, and 0 otherwise;
Post Predictorsic ¼Intensity of High Arousal
ic
, Intensity
of Low Arousal
ic
, SSTic , and LSM ic ;
Dum Timingic ¼dummy variables for years (baseline is
2015), month (baseline is December), weekend day
(baseline is week day), and time of the day (baseline
is night time, EST);
u0c ¼brand community–specific error term; and
ric ¼post-specific error term.
Next, to determine how firms can prevent viral online
firestorms, we examine 331,370 negative posts that received
at least one firm response according to the following
equation:
Viralityt 2
ic ¼g00 þg0104Firm Controlsc
þg0510Brand Community Controlsc
þg1117Post Controlsic
þg18Firm Response Timeic
þg1923Post Predictorsic
þg2443Dum Ti mingic
þg4448First Firm Responseic
þg4952Interactionsic þu0c þric ;
ð4Þ
where t_2 is the time period after the first firm response;
Viralityt 2
ic ¼the combined sum of likes and comments
post i received from other customers in community c
any time after the first firm response (brand
community-centered and log-transformed);
Firm Response Time ¼time until the first firm response;
First Firm Response ¼Compensation ic, Apology ic,
Channel Change ic , Intensity of Empathy ic,and
Intensity of Explanationic; and
Interactionsic ¼Intensity of High Arousal
ic
Intensity of Empathyic, Intensity of High Arousal
ic
Intensity of Explanation ic , Intensity of Low Arou-
sal
ic
Intensity of Empathy ic , and Intensity of Low
Arousal
ic
Intensity of Empathy ic .
Finally, we investigate how firms can mitigate the evolved
online firestorms represented by 15,762 negative posts that
achieved above-average virality and to which firms responded
multiple times:
Viralityt 3
ic ¼g00 þg0104Firm Controlsc
þg0510Brand Community Controlsc
þg1117Post Controlsic
þg18Firm Response Timeic
þg1923Post Predictorsic
þg2443Dum Ti mingic
þg4448First Firm Responseic
þg4951Subsequent Firm Responsesci
þg5253Variance in Firm Responsesci þu0c þric
ð5Þ
where t_3, ..., T is the time period after the last firm response;
Viralityt 3
ic ¼the combined sum of likes and comments
post i received from other customers in community c
any time after the last firm response (brand
community-centered and log-transformed);
Subsequent Firm Responsesic ¼include all firm
responses after the first firm respon se, Compensation ic ,
Apology ic, Channel Change ic , Intensity of Empathy ic ,
and Intensity of Explanationic ;and
Variance in Firm Responsesci ¼Variance in Empathyic
and Variance in Explanationic (across all Firm
Responses
ic
).
Results
Detecting Potential Online Firestorms
The results provide support for our hypotheses that the inten-
sity of high-arousal emotions (vs. low-arousal emotions),
SST, and LSM relate to the virality of negative eWOM
(Table 2, Model 3).
2
Both intensities of high arousal (g¼
.186, p<.01) and low arousal (g¼.026, p<.01) relate
positively to virality. However, a t-test reveals that the inten-
sity of high arousal is more strongly related to virality than
the intensity of low arousal (t ¼35.15, p<.01), in support
of H
1
. In addition, SST (g¼1.432, p<.01) and LSM (g¼
.025, p<.01) relate positively to virality, in support of H
2
and H
3
. Considering the relative influence of the drivers, we
find that SST exerts the strongest impact on virality (all t
77.97, p<.01).
Regarding brand community controls, we find that aver-
age tie strength relates negatively to virality (g¼.247,
p<.05), and greater variance in linguistic style relates
positively to virality (g¼5.550, p<.10). Online firestorms
thus appear to occur less in brand communities in which
members have stronger connections with one another and
are more similar. Regarding the post controls, we find that
competing inputs (g¼.001, p<.01) and previous com-
plaints (g¼.213, p<.01) relate negatively to virality.
Conversely, post length (g¼.005, p<.01) and post com-
plexity (g¼.017, p<.01) relate positively to virality.
Finally, a lack of firm response is significantly related to
increased virality (g¼.029, p<.01), clearly indicating the
importance of actively managing negative eWOM in online
brand communities.
Preventing Potential Online Firestorms
When considering the main effects of the intensities of empa-
thy and explanation in firm responses in Model 5 (Table 3), we
find that the increased use of empathy (g¼.069, p<.01)
leads to significantly lower virality than the increased use of
explanation (g¼.011, p<.01; t ¼14.42, p<.01).
Regarding other firm responses that reflect disengaging from
2
The fixed effects for year, month, weekend, and time of day for all models
appear in the Web Appendix.
Herhausen et al. 11
the conversation, we find that responses that contain an
apology (g¼.004, p<.01) or a suggestion for a channel
change (g¼.005, p<.01) relate negatively to virality.
However, immediately offering compensation fosters the vir-
ality of negative eWOM (g¼.003, p<.01).
When considering the interaction effects in Model 6 (Table
3), we find a positive interaction between the intensity of high
arousal and the increased use of empathy in the firm response
(g¼2.678, p<.01). The increased use of empathy in a
response to a post with a high intensity of high arousal thus
increases, rather than decreases, virality. Conversely, we find
a significant negative interaction between the increased inten-
sity of high arousal and the increased use of explanation in the
firm response (g¼1.437, p<.01). Therefore, the increased
use of explanation in a response to a post with a high intensity
of high arousal significantly reduces its virality, as depicted in
Figure 3. A t-test further reveals significant differences
between providing more empathy versus more explanation
in buffering the effect of high arousal (more empathy: g¼
.386, p<.01; more explanation: g¼.180, p<.01; t ¼23.43,
Table 2. Predictors of Potential Online Firestorms.
DV ¼Total Virality of Negative Customer Post
Sample ¼All Posts With and Without Firm Response
Model 1 Model 2 Model 3
gt/r gt/r gt/r
Level 2: Firm/Brand Community
Firm controls
GICS consumer discretionary .185 1.48 .190 1.56 .190 1.56
GICS consumer staples .100 .62 .100 .63 .100 .64
Brand familiarity .149 .57 .139 .55 .139 .55
Brand reputation .640 .69 .664 .74 .665 .74
Brand community controls
Brand community size .002 .18 .002 .20 .002 .20
Brand community attentiveness .009 .15 .011 .20 .011 .20
Brand community expressiveness 5.641 1.33 5.736 1.38 5.726 1.38
Average firm engagement .287 1.60 .290 1.66 .291 1.66
Average tie strengths .230
y
1.87 .247* 2.06 .247* 2.06
Variance in linguistic style 5.363
y
1.80 5.498
y
1.89 5.500
y
1.89
Level 1: Customer Post
Post controls
Competing inputs .001** .03 .001** .03 .001** .03
Sentiment previous post .001 .00 .001 .00 .002 .00
Post length .005** .07 .005** .07 .005** .06
Post complexity .018** .02 .012** .01 .017** .01
Negation in post .021** .01 .022 .01 .004 .00
Previous complaints .017** .00 .212** .05 .213** .05
No firm response .030** .14 .030** .14 .029** .14
Post predictors
Negative emotions .133** .06
Intensity of high arousal (H
1
).186** .07
Intensity of low arousal (H
1
).026** .01
Strength of structural ties (H
2
) 1.435** .13 1.432** .13
LSM (H
3
) .032** .05 .025** .04
Log-likelihood 1,019,495 1,030,248 1,031,025
Change in log-likelihood 10,752** 11,530**
N
Level 2
89 brand communities
N
Level 1
472,995 negative posts
y
p<.10.
*p<.05.
**p<.01.
Notes: Significance is based on two-tailed tests. We report t-values for Level 2 and effect size r for Level 1. Change in fit in comparison with the model with control
variables only. We test GICS consumer discretionary and GICS consumer staples against GICS other. Fixed effects for year, month, weekend, and time of day are
included in the model and reported in the Web Appendix. Slopes of high-arousal and low-arousal emotions are different at p<.01 and t ¼35.15. An additional
model that compared the effects of different high-arousal emotions (fear/anxiety, anger, disgust) with the low-arousal emotion (sadness) revealed that fear/anxiety
(t ¼10.75, p<.01), anger (t ¼37.52, p<.01), and disgust (t ¼1.74, p<.10) are all more strongly related to virality than sadness. The small difference between
disgust and sadness is in line with Berger and Milkman’s (2012) finding and can be attributed to the relative scarcity of disgust in the negative eWOM messages. We
also find that anger relates more strongly to virality then other high-arousal emotions (all ts 21.34, p<.01).
12 Journal of Marketing XX(X)
p<.01). Overall, these results support H
4
and indicate that
intensive high arousal is better contained with more
explanation.
Contrary to our expectations, we find no negative interac-
tion between the intensity of low arousal and the increased use
of empathy in the firm response at conventional significance
levels (g¼.042, p¼.30) and a negative interaction between
the intensity of low-arousal emotions and the increased use of
explanation in the firm response (g¼.206, p<.01). A t-test
reveals no significant differences between providing more
empathy versus more explanation in buffering intensive low-
arousal emotions (more empathy: g¼.025, p<.01; more
explanation: g¼.016, p<.01; t ¼1.63, p¼.10). Thus, H
5
is not supported.
Table 3. Preventing Potential Online Firestorms.
DV ¼Virality After the First Firm Response
Sample ¼All Posts With at Least One Firm Response
Model 4 Model 5 Model 6
gt/r gt/r gt/r
Level 2: Firm/Brand Community
Firm controls
GICS consumer discretionary .206 1.45 .206 1.44 .206 1.44
GICS consumer staples .104 .57 .105 .57 .104 .57
Brand familiarity .119 .40 .118 .40 .117 .40
Brand reputation .842 .80 .833 .79 .834 .79
Brand community controls
Brand community size .003 .24 .003 .24 .003 .24
Brand community attentiveness .019 .28 .018 .28 .019 .28
Brand community expressiveness 5.971 1.23 5.968 1.23 5.960 1.23
Average firm engagement .323 1.57 .323 1.58 .323 1.58
Average tie strengths .310* 2.20 .309* 2.20 .310* 2.20
Variance in linguistic style 5.996
y
1.76 5.995
y
1.76 5.993
y
1.76
Level 1: Customer Post
Post controls
Competing inputs .003** .03 .003** .03 .003** .03
Sentiment previous post .005** .01 .005** .01 .005** .01
Post length .005** .08 .006** .08 .006** .08
Post complexity .017** .01 .018** .02 .018** .02
Negation in post .005 .00 .002 .00 .002 .00
Previous complaints .171** .03 .179** .03 .178** .03
Firm response time .001** .02 .001** .02 .001** .02
Post predictors
Intensity of high arousal .220** .08 .217** .08 .252** .09
Intensity of low arousal .025** .01 .025** .01 .027** .02
Strength of structural ties 1.047** .08 1.044** .08 1.038** .08
LSM .022** .04 .024** .04 .023** .04
First firm response
Compensation .003** .01 .003** .01
Apology .004** .02 .004** .02
Channel change .005** .03 .006** .03
Intensity of empathy (EMP) .069** .04 .065** .04
Intensity of explanation (EXP) .011** .01 .006* .00
Intensity of high arousal EMP (H
4
)2.678** .05
Intensity of high arousal EXP (H
4
)1.437** .04
Intensity of low arousal EMP (H
5
).042 .00
Intensity of low arousal EXP (H
5
).206** .01
Log-likelihood 770,911 771,760 773,267
Change in log-likelihood 849** 1,508**
N
Level 2
89 brand communities
N
Level 1
331,370 negative posts
y
p<.10.
*p<.05.
**p<.01.
Notes: Significance is based on two-tailed tests. We report t-values for Level 2 and effect size r for Level 1. Fixed effects for year, month, weekend, and time of day
are included in the model and reported in the Web Appendix.
Herhausen et al. 13
Mitigating Evolved Online Firestorms
Variations in the intensity of empathy (g¼.185, p<.05)
and intensity of explanation (g¼.205, p<.01) relate nega-
tively to virality in Model 9 (Table 4), in support of H
6a
and
H
6b
. These findings suggest that firms should vary their
response formulation to decrease the virality of evolved
online firestorms. We further find that an increased use of
explanation in subsequent responses increases virality (g¼
.083, p<.01). To mitigate the virality of negative eWOM,
firms should vary their response, focusing on more empathy
in later responses. Notably, at an evolved stage of an online
firestorm, firm responses that contain an apology (g¼.031, p
<.01) or a channel change (g¼.038, p<.01) positively
relate to virality. These findings reveal that such response
strategies not only are ineffective in mitigating evolved online
firestorms but even fuel the fire. In contrast, offering com-
pensation in a subsequent response negatively relates to vir-
ality (g¼.045, p<.01).
Additional Analyses
Addressing potential endogeneity. In our research design, endo-
geneity might arise out of reverse causality, omitted variables,
or a learning effect. First, the timestamps provided allow us to
avoid reverse causality by considering only customer com-
ments that occur after the respective firm response. Second,
we address omitted variables stemming from the 89 firms with
firm-fixed-effects regressions to account for the unobservable
heterogeneity of each brand community (Allison 2009). All
results are fully in line with the main analyses, as displayed in
the Web Appendix. Third, we account for a learning effect in
an analysis in which we include a continuous time variable
that captures the 52 months of our study period (1 ¼October
2011 to 52 ¼January 2016). If endogeneity from a learning
effect biases our results, this time variable would reduce the
virality of negative posts and/or increase the effectiveness of
the response strategies. Instead, when we include the time
variable in our prevention model, we find a nonsignificant
main effect on virality (p¼.98). The positive interaction
effects of the time variable with empathy (g¼.010, p<
.01) and explanation (g¼.002, p<.08) indicate that both
response strategies become less effective over time. Taken
together, these additional analyses suggest it is unlikely that
endogeneity biases our results.
Additional interaction effects. Recent research has suggested that
different drivers of virality may reinforce one another (Ludwig
et al. 2013). Thus, we tested for potential interaction effects
among the four drivers of online firestorms in Model 3. We find
that both SST (g¼28.840, p<.01) and LSM (g¼.743, p<
.01) increase the effect of high-arousal emotions. In addition,
LSM increases the effect of low-arousal emotions (g¼.107, p
<.01), and SST and LSM reinforce each other (g¼2.079, p<
.01). When we tested whether SST influences the effectiveness
of empathy and explanation, we find that both empathy (g¼
4.00, p<.01) and explanation (g¼5.83, p<.01) are more
effective in reducing the virality of negative eWOM from cus-
tomers with strong structural ties in the brand community.
Alternative measures of virality. In line with previous research
(e.g., Relling et al. 2016), we regard the combined number of
likes and comments as the most appropriate measure for vir-
ality. For robustness, we tested our results with three separate,
alternative measures of virality: likes, comments, and shares.
Previous research has indicated mixed results, such that Lee,
Hosanagar, and Nair (2018) find no major difference in using
likes or comments as measures, but De Vries, Gensler, and
Leeflang (2012) find different results using likes versus com-
ments to measure virality. We replicate all the models with
these different outcome measures and report the results in the
Web Appendix; nearly all estimates for the hypothesized
effects are directionally similar and significant at conventional
levels.
A: Interaction Effect of Intensity of High Arousal and Intensity of Empathy
B: Interaction Effect of Intensity of High Arousal and Intensity of Explanation
3.80
3.81
3.82
3.83
3.84
3.85
High Arousal (-1 SD) High Arousal (+1 SD)
Viralityt2 (log-transformed)
Empathy (-1 SD) Empathy (+1 SD)
3.80
3.81
3.82
3.83
3.84
3.85
High Arousal (-1 SD) High Arousal (+1 SD)
Viralityt2 (log-transformed)
Explanation (-1 SD) Explanation (+1 SD)
Figure 3. Firm response strategies moderate the effect of high-
arousal emotions on virality.
Notes: Virality
t_2
is measured after the first firm response.
14 Journal of Marketing XX(X)
Table 4. Mitigating Evolved Online Firestorms.
DV ¼Virality After the Last Firm Response
Sample ¼All Posts With More Than One Firm Response and Above-Average Virality
Model 7 Model 8 Model 9
gt/r gt/r gt/r
Level 2: Firm/Brand Community
Firm controls
GICS consumer discretionary .021 1.34 .022 1.40 .022 1.40
GICS consumer staples .012 .54 .015 .67 .015 .67
Brand familiarity .006 .16 .007 .16 .007 .17
Brand reputation .009 .08 .005 .04 .004 .04
Brand community controls
Brand community size .001 1.08 .002 1.15 .002 1.15
Brand community attentiveness .008 1.24 .007 1.19 .007 1.19
Brand community expressiveness .304 .50 .295 .48 .296 .49
Average firm engagement .014 .53 .012 .46 .012 .46
Average tie strengths .020 1.45 .020 1.45 .020 1.45
Variance in linguistic style .616 1.08 .615 1.07 .616 1.07
Level 1: Customer Post
Post controls
Competing inputs .001 .00 .001 .00 .001 .00
Sentiment previous post .030 .01 .029 .01 .029 .01
Post length .013** .07 .013** .07 .013** .07
Post complexity .017 .01 .016 .00 .016 .00
Negation in post .087 .01 .082 .01 .082 .01
Previous complaints .707** .03 .710** .03 .710** .03
Firm response time .003** .05 .003** .05 .003** .05
Post predictors
Intensity of high arousal .363** .06 .364** .06 .364** .06
Intensity of low arousal .122** .03 .124** .03 .124** .03
Strength of structural ties 5.796** .11 5.813** .11 5.812** .11
LSM .045** .03 .045** .03 .045** .03
First firm response
Compensation .018
y
.01 .017 .01 .017 .01
Apology .011* .02 .012** .02 .012** .02
Channel change .008
y
.02 .010* .02 .010* .02
Intensity of empathy .084 .01 .072 .01 .072 .01
Intensity of explanation .129** .03 .127** .03 .127** .03
Subsequent firm responses
Compensation .044** .04 .045** .04 .045** .04
Apology .032** .06 .031** .06 .031** .06
Channel change .040** .08 .038** .08 .038** .08
Intensity of empathy .110** .03 .016 .00 .021 .00
Intensity of explanation .042 .01 .082** .02 .083** .02
Variance in firm response .198** .04
Variance in empathy (H
6a
).185** .02
Variance in explanation (H
6b
).205** .03
Log-likelihood 1,463.20 1,480.45 1,475.86
Change in log-likelihood 17.25** 12.66**
N
Level 2
72 brand communities
N
Level 1
15,762 negative posts
y
p<.10.
*p<.05.
**p<.01.
Notes: Significance is based on two-tailed tests. We report t-values for Level 2 and effect size r for Level 1. Fixed effects for year, month, weekend, and time of day
are included in the model and reported in the Web Appendix.
Herhausen et al. 15
Discussion
Theoretical Implications and Extensions
Extensive literature has addressed the benefits of eWOM (e.g.,
Baker, Donthu, and Kumar 2016), but theoretical and empirical
work devoted to negative eWOM in brand communities is
scarce. Drawing on research on negative WOM (e.g., Brown
and Reingen 1987; Heath, Bell, and Sternberg 2001), we com-
bine multiple sender and relational aspects that likely increase
the virality of negative eWOM in online brand communities.
We then empirically assess their role in driving virality across
472,995 potential online firestorms in 89 online brand commu-
nities of S&P 500 firms. Integrating common recovery
approaches from service literature (Hill, Roggeveen, and Gre-
wal 2015) with emotion regulation strategies (Gross and
Thompson 2007), we also highlight the relative effectiveness
of different firm response approaches and cross-response var-
iations to prevent and mitigate online firestorms. Thereby, our
study makes three primary contributions to extant marketing
research (for a summary of the results, see Table 5).
First, we advance research on how to detect the online fire-
storm potential of negative eWOM with an empirical investi-
gation of prototypical conceptions of different drivers of
virality and their interrelations. In line with prior marketing
research on negative WOM (Brown and Reingen 1987; Heath,
Bell, and Sternberg 2001, Mittal, Huppertz, and Khare 2008),
we show that the virality of negative eWOM in online brand
communities varies depending on sender and relational aspects.
As an extension of Berger and Milkman’s (2012) findings
about sharing newspaper articles, we find that in online brand
communities, the use of more high-arousal-emotion words in
negative eWOM increases its virality and makes it relatively
more contagious than the use of low-arousal-emotion words.
We also find that stronger structural ties between the complain-
ing customer and the receiving online brand community relate
to greater virality of negative eWOM. Thus, Brown and Rein-
gen’s (1987) conclusions, gathered from a small offline com-
munity, hold in large digital communities. Notably, owing to
data limitations, we were only able to use the frequency of
communication as tie strength indicator, not the importance
attached to the relationship. Moreover, in this otherwise anon-
ymous context, we use the degree of LSM as an indicator of
interpersonal similarity between senders and receivers. In line
with Ludwig et al. (2013), we find that closer LSM between the
complaining customer and the receiving online brand commu-
nity relates to greater virality of negative eWOM. In contrast
with previous studies, we consider a broader set of drivers of
virality and test their relative importance and interrelation-
ships. Structural ties are the strongest driver of virality. Further-
more, strong structural ties and a close LSM amplify the
virality effect of high-arousal emotions in negative eWOM.
To put our estimated effects into perspective with related
research, we calculate effect sizes r using the formula from
Rosenthal and Rosnow (2007). The identified drivers show
effect sizes ranging from .04 to .13. Taken together, our study
thus advances negative eWOM research with a theoretically
grounded framework of sender and relational aspects, useful
for detecting potential online firestorms in brand communities.
Second, as suggested in prior research (e.g., Ma, Sun, and
Kekre 2015), we consider the effectiveness of firm response
approaches to reduce the contagiousness of negative eWOM in
online brand communities. Contributing to this emergent
stream of research, we reconcile research on service recovery
(Hill, Roggeveen, and Grewal 2015) and emotion regulation
(Gross and Thompson 2007) to delineate common, theoreti-
cally grounded firm response alternatives. We empirically con-
firm the common knowledge that not responding to a negative
customer post is a firm’s worst choice and should be avoided.
In line with Homburg, Ehm, and Artz (2015), we further find
that responding fast is important. When actively engaging in
elaboration with the complaining customer within the online
brand community, the increased use of empathy is more effec-
tive overall. However, if a negative eWOM messages contains
exceptionally intense high-arousal emotions, increasing the
amount of explanation is more effective for preventing and
mitigating its virality. Contrary to our expectations, we find
that explanation rather than empathy best contains negative
eWOM containing severe low-arousal emotions (e.g., sadness).
Indeed, in line with Gross and Thompson (2007), customers
who are experiencing severe emotions (either anger or sadness)
are not able to shift attention but are always looking for expla-
nations beyond empathy.
Initially aiming to block and disengage, rather than engage
in elaborate online discussions, firms are best advised to offer
an apology or suggest a channel change. At a later stage, once
the negative eWOM has gathered support within the online
brand community, these disengagement approaches are not
only ineffective but may further increase a post’s virality.
Table 5. Overview of Results.
Hypotheses Effect on Virality of Negative eWOM Support
H
1
High- versus low-arousal emotions Intensity of high arousal >Intensity of low arousal Supported (Table 3)
H
2
Strength of structural ties Strong structural ties >Weak structural ties Supported (Table 3)
H
3
LSM Closer LSM >More distant LSM Supported (Table 2)
H
4
Firm response to high-arousal emotions For high arousal: More explanation >More empathy Supported (Table 3)
H
5
Firm response to low-arousal emotions For low arousal: More empathy >More explanation Not supported (Table 3)
H
6a
Variation in empathy in firm responses Variation in empathy >Similar intensity of empathy Supported (Table 4)
H
6b
Variation in explanation in firm responses Variation in explanation >Similar intensity of explanation Supported (Table 4)
16 Journal of Marketing XX(X)
Offering to take a customer’s complaint offline seems to
remove at least some negative conversations before they go
viral. However, at a later stage, when the support of others for
the negative eWOM has increased its perceived reliability
(Dholakia, Basuroy, and Soltysinski 2002), both the complain-
ing customer and the community likely feel further encouraged
by a firm’s admittance of guilt (i.e., apology) and disgruntled if
forcefully removed from the conversation (i.e., channel
change). Moreover, the switching effect of channel change is
in line with recent research from Gr´egoire et al. (2018), who
find that for customers using indirect revenge behaviors with
public exposure (e.g., complaining online in a brand commu-
nity), desire for revenge increases over time. Offering compen-
sation only mitigates the virality of the negative eWOM
message when used as a later response by the firm. Some
controversy exists regarding the effectiveness of compensation
as a means to recover from a service failure. That is, compen-
sation may dissipate customers’ frustrations (Bitner, Booms,
and Tetreault 1990), but offering it without explanation
increases attributions of control and indicates an admission of
guilt, evoking more negative evaluations (Bitner 1990). In line
with Grewal, Roggeveen, and Tsiros (2008), we therefore sug-
gest that compensation is most effective if it follows an offer of
an explanation or empathy. With these findings on how to
prevent online firestorms, we extend debates about service
recovery strategies to negative eWOM in online brand
communities.
Third, online firestorms that have evolved often require
multiple responses. We contribute to research on social media
sharing by analyzing the implications of variations in firm
response sequences (Batra and Keller 2016). Several firm
responses are likely to be interpreted jointly rather than in
isolation (Villarroel Ordenes et al. 2018). We find that varying
response approaches, rather than consistently responding in the
same way, can reduce the virality of evolved online firestorms
in brand communities. Firms that use the same intensity of
empathy or explanation to respond to negative eWOM
increase, rather than mitigate, its contagiousness to other com-
munity members. Together with our findings that the effective-
ness of compensation, apology, and channel change depends on
when they are used, we add new insights into how best to
sequence multiple firm responses in social media. Finally, as
a useful resource for research, we have developed text-mining
dictionaries to derive firms’ response approaches automatically
using a top-down text-mining approach (see the Web Appen-
dix). We carefully followed traditional dictionary development
standards (Humphreys and Wang 2018), so researchers who
want to examine written or transcribed firm responses in
firm–customer exchanges may use these dictionaries as a start-
ing point for their own investigations of firm response
strategies.
Managerial Implications
By investigating how to detect, prevent, and mitigate the vir-
ality of negative eWOM in online brand communities, we offer
several actionable implications for managers. We discuss these
in the following subsections.
Detecting potential online firestorms. Brand community manag-
ers, who struggle to identify potentially threatening negative
eWOM messages, should consider complainers’ message for-
mulations, beyond what is literally said, as well as their rela-
tionship with other members of the community. First, by using
our dictionary-based, straightforward, automatic text-mining
approach, managers can assess the high- and low-arousal levels
of negative messages to predict their potential virality. The
higher-arousal-emotion words a message contains, the more
likely it is to go viral. Second, managers should assess the tie
strength of the customer posting the negative message. Nega-
tive messages by customers who frequently interact with other
community members are more likely to go viral than messages
by customers who are relative strangers in the community.
Third, text-mining tools can track the brand community’s
dominant communication style continuously and contrast it
with the style of each negative customer post. Posts that closely
match the dominant communication style are more likely to go
viral. Taken together, the identified drivers explain 25%of
virality across all examined brand communities. Finally, the
different drivers amplify one another, so managers should be
particularly cautious of complaining customers with strong
structural ties who closely match the community’s dominant
communication style.
Preventing potential online firestorms. Unlike in a traditional ser-
vice recovery setting, the success of managers’ responses in
preventing online firestorms critically depends on their ability
to satisfy both the complainant and the brand community. Not
responding is the worst choice, but the firm response also needs
to be fast and tailored to the customer’s message. In an initial
response, empathy is generally most effective for containing
negative eWOM. However, very negative messages that use an
exceptional amount of high-arousal emotion words (e.g.,
“angry,” “hate”) demand more explanation. To disengage from
the conversation and reduce the virality of negative eWOM
upfront, managers should apologize or ask the unsatisfied cus-
tomer to use another channel to raise the issue. Offering com-
pensation immediately is not advised; it is effective only as a
later response. We find that an appropriate response strategy
can reduce virality of an intensive high-arousal post by up to
10%, which may equal hundreds of angry customers supporting
and sharing negative eWOM.
3
Mitigating evolved online firestorms. Some negative eWOM can-
not be prevented from going viral or “catching fire” among
3
For the calculation, we compared the expected virality of an intensive
high-arousal post where firms respond with above-average empathy,
below-average explanation, and with a compensation with the expected
virality of an intensive high-arousal post where firms respond with
below-average empathy, above-average explanation, and with apology and
channel change (while keeping all other predictors constant).
Herhausen et al. 17
other customers, and managers will need to respond multiple
times. These responses are likely to be viewed collectively,
rather than in isolation, so managers should consider each
response as part of an overall response sequence. Rather than
consistently posting the same message, managers should vary
the use of empathy and explanation to mitigate the further
virality of negative eWOM messages. An explanatory
approach is viable as a first response to exceptionally intense
high-arousal negative eWOM, and later firm responses should
use increased empathy. If used at a later response stage, apol-
ogizing or suggesting a different communication channel will
“feed the fire” and increase the virality of the negative eWOM.
Instead, offering compensation should be the last resort for
managers to prevent further elaborations and reduce the virality
of negative eWOM. Using an appropriate response strategy
over time can reduce subsequent virality by up to 11%.
4
Limitations and Directions for Further Research
Our results are consistent with the proposition that firms need
to manage negative eWOM in their online brand communities
actively to prevent or mitigate their detrimental effects (e.g.,
Hewett et al. 2016; Pfeffer, Zorbach, and Carley 2014).
Although we believe our findings have broad applicability,
managing online firestorms is a vast and largely neglected field
of research that is of critical importance to managers. Thus, it is
important to recognize some limitations of our study and sug-
gest further research. Although our large-scale study offers
theoretical and empirical insights into textual aspects related
to the virality of negative eWOM, using what Humphreys and
Wang (2018) call a top-down approach, we also acknowledge
that alternative bottom-up approaches might usefully derive
specific service or product failures and their severity to test the
suitability of the response approaches we outline. Similarly, in
other communication contexts where heuristic processing is
less prevalent, the implications of systematic content in firm
responses should be assessed (e.g., size of the compensation,
legitimacy of the argument).
Moreover, the scope of our study is limited to all posts
visible in the communities. On Facebook, firms have the option
to remove comments. Deletion criteria may include especially
offensive (e.g., racist, derogatory) messages. This option may
bias our results for high-arousal emotions, because we do not
observe deleted posts. Managerial reports strongly discourage
deleting negative customer posts on social media (e.g., Duncan
2016), but extreme posts missing from our data set could fur-
ther increase the virality effect, or the effect may taper off or
even reverse with an extreme use of high-arousal emotions.
The consequences of deleting customer posts remain to be
investigated. In addition, we could not assess perceptions of
source credibility or how source credibility may interact with
the use of empathy or other firm responses to make them more
or less effective. Therefore, further research might seek novel
ways to determine the importance of source credibility for
message acceptance. Similarly, lacking an ability to account
for the attitudinal importance that customers attribute to their
ties in brand communities, continued research could extend our
study by assessing the implications of weak and strong tie
perceptions for sharing negative eWOM in brand communities.
Potential measures that could be adapted for this purpose could
be obtained from Umashankar, Ward, and Dahl (2017).
Furthermore, we were not able to obtain data on the number
of friends due to privacy restrictions in Facebook’s API terms
and conditions. However, in line with Peng et al. (2018), we
believe that strength of structural ties (i.e., the number of
encounters with other users in the brand community) influences
virality regardless of the number of friends.
Interestingly, we found that customers expressing intense
negative emotions, irrespective of whether they are high
(e.g., anger) or low (e.g., sadness) on arousal, are looking for
explanations rather than empathy. Future research should con-
sider how the relative lack of emotionality might relate to the
suitability of firm’s response options. Finally, it was surprising
that immediate compensation leads to more virality. Poten-
tially, if firms offering an initial compensation that is perceived
as not high enough could lead to an offense. If firms would then
offer a higher compensation at a later stage this might lead to
less virality. Thus, future research should investigate whether
offering greater levels of compensation at later stages lead to
the observed effects.
Acknowledgments
The authors thank Dinesh Gauri, Yu Ma, and Hannes Datta for meth-
odological advice, and Andrew Stephen, Jonah Berger, Grant Packard,
Luigi De Luca, and Oliver Emrich as well as participants of research
seminars at Cardiff University, Nova School of Business and Econom-
ics, Copenhagen Business School, the University of St. Gallen, and the
Interactive Marketing Research Conference 2018 for their helpful
comments. The authors are grateful to Stephen Hahngriffiths from the
Reputation Institute for sharing their data on brand familiarity and
brand reputation. The first author handled the data analysis and the
fourth author processed the social media data.
Associate Editor
Christian Homburg served as associate editor for this article.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to
the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for
the research, authorship, and/or publication of this article: The first
author received a grant from the Basic Research Fund (GFF) of the
University of St. Gallen for this project.
4
For the calculation, we compared the expected virality of a post where firms
first respond with compensation, later with apology and channel change, and
no variation in empathy and explanation to the expected virality of a post where
firms first respond with apology and channel change, later with compensation,
and with variation in empathy and explanation (all other predictors constant).
18 Journal of Marketing XX(X)
References
Allison, Paul D. (2009), Fixed Effects Regression Models. Thousand
Oaks, CA: SAGE Publications.
Ansari, Asim, Carl F. Mela, and Scott A. Neslin (2008), “Customer
Channel Migration,” Journal of Marketing Research,45(1),
60–76.
Aral, Sinan, Lev Muchnik, and Arun Sundararajan (2009),
“Distinguishing Influence-Based Contagion from Homophily-
Driven Diffusion in Dynamic Networks,” Proceedings of the
National Academy of Sciences, 106 (51), 21544–49.
Baker, Andrew M., Naveen Donthu, and V. Kumar (2016),
“Investigating How Word-of-Mouth Conversations About Brands
Influence Purchase and Retransmission Intentions,” Journal of
Marketing Research, 53 (2), 225–39.
Barsade, Sigal G. (2002), “The Ripple Effect: Emotional Contagion
and Its Influence on Group Behavior,” Administrative Science
Quarterly, 47 (4), 644–75.
Batra, Rajeev, and Kevin Lane Keller (2016), “Integrating Marketing
Communications: New Findings, New Lessons, and New Ideas,”
Journal of Marketing, 80 (6), 122–45.
Benoit, Kenneth, Paul Nulty, Pablo Barber, Kohei Watanabe, and
Benjamin Lauderdale (2018), Quantitative Analysis of Textual
Data, https://cran.r-project.org/web/packages/quanteda/.
Berger, Jonah (2014), “Word-of-Mouth and Interpersonal Communi-
cation: A Review and Directions for Future Research,” Journal of
Consumer Psychology, 24 (4), 586–607.
Berger, Jonah, and Eric M. Schwartz (2011), “What Drives Immediate
and Ongoing Word of Mouth?” Journal of Marketing Research,48
(5), 869–80.
Berger, Jonah, and Katherine L. Milkman (2012), “What Makes
Online Content Viral?” Journal of Marketing Research, 49 (2),
192–205.
Bitner, Mary J. (1990), “Evaluating Service Encounters: The Effects
of Physical Surroundings and Employee Responses,” Journal of
Marketing, 54 (2), 69–82.
Bitner, Mary J., Bernard H. Booms, and Mary Stanfield Tetreault
(1990), “The Service Encounter: Diagnosing Favorable and Unfa-
vorable Incidences,” Journal of Marketing, 54 (1), 71–84.
Blaine,