Almost everything a consumer learns about a product arrives through someone else. A friend’s offhand recommendation, a one-star review, a forwarded video, a retweeted complaint—these consumer-to-consumer transmissions move demand more reliably than most paid messages, and they do so at a cost the firm does not pay. This chapter is about how such transmission works: what gets passed along, why people pass it, how far and how fast it travels, and how much of the resulting spread a manager can actually engineer rather than merely observe.
Two questions organize the field, and they are easy to conflate. The first is psychological: given that a person is talking, what makes them choose one piece of content over another to share? This is the literature on word of mouth (WOM) and its drivers—social currency, emotion, usefulness, narrative. The second is structural: given that content is shared, what shape does the resulting diffusion take through the social network, and is that shape predictable? This is the literature on cascades and structural virality. A serious treatment needs both. Content drivers explain the probability that any individual relays a message; network structure determines whether those individual relays compound into an epidemic or fizzle. We develop the content account first, formalize the diffusion account second, and close on the measurement and identification problems that make virality genuinely hard to study—chief among them the fact that what looks “viral” is often neither viral nor caused by the content at all.
By the end the reader should be able to define virality precisely, distinguish it from popularity, write down a model of cascade size and state what limits its predictability, and reason about seeding and measurement with the right identification caveats in hand.
31.1 Defining the Construct
Word of mouth predates the internet by the entire history of language, and its canonical definition is correspondingly old. We adopt Westbrook (1987)’s formulation as the field’s landmark statement.
Word of mouth is “informal communications directed at other consumers about the ownership, usage, or characteristics of particular goods and services or their sellers” (Westbrook 1987).
Three features of this definition do real work. The communication is informal (it is not the firm’s paid voice), it is consumer-to-consumer (the firm is at best an indirect party), and it is about market objects (products, services, sellers). The migration of this activity online produces electronic word of mouth (e-WOM, sometimes “word of mouse”): the same construct, now mediated by platforms that record, timestamp, and graph every utterance. Both classical WOM and e-WOM exploit the network effect to reach audiences the firm could not address directly (Vilpponen, Winter, and Sundqvist 2006), but e-WOM does so with lower marginal cost, faster diffusion, and sharper targeting (Bampo et al. 2008; Feroz Khan and Vong 2014).
Virality is the construct that e-WOM makes salient, and the literature uses the word in two incompatible ways that must be kept apart. One usage is processual: virality is the act or trajectory of successive sharing—“a process whereby an ad is successively shared by viewers.” The other is dispositional: virality is a property of content, namely its latent potential to be spread. We adopt the dispositional reading throughout, because it is the one that supports prediction and design: we want to know what makes a given piece of content likely to diffuse, holding the network fixed.
WOM, virality, and sharing as near-synonyms
We treat WOM, virality, and sharing interchangeably, following the strong construct overlap documented by Camarero and San José (2011) and Phelps et al. (2004). The distinctions that matter for this chapter are not lexical but mechanical: the content drivers of a single relay versus the network dynamics of aggregate diffusion. We are explicit about which we mean at each point.
Formally, fix a population of potential relayers indexed by \(i\) and a piece of content \(c\). Let \(S_{ic} \in \{0,1\}\) indicate that \(i\) relays \(c\). The dispositional virality of \(c\) is the share probability marginalized over the population, \[
v(c) \;=\; \mathbb{E}_i\!\left[\,\Pr(S_{ic} = 1 \mid \mathbf{z}_c, \mathbf{w}_i)\,\right],
\tag{31.1}\] where \(\mathbf{z}_c\) are content/context features (the levers a marketer can pull) and \(\mathbf{w}_i\) are person- and tie-level moderators. The content literature in this chapter is, in effect, a catalog of the partial derivatives \(\partial v / \partial \mathbf{z}_c\): which features of content raise the per-person relay probability. The structural literature (Section 31.5) asks what happens when these individual relays are composed through the edges of a social graph.
Why this matters commercially is blunt: sharing is necessary but not sufficient for value. Intentions overstate behavior, and behavior overstates profit. In one telecommunications referral program, 81% of customers said they would refer others, only 33% actually did, and of those referred only 8% became profitable customers (V. Kumar, Petersen, and P.Leone 2007). The funnel from stated willingness to share, to actual sharing, to monetized outcome leaks at every stage—a recurring caution when reading self-reported sharing studies.
31.2 Theoretical Foundations
The two literatures this chapter joins, content drivers and network structure, each rest on established theory, and naming the governing accounts shows why the empirical regularities are not coincidental. The structural side descends from social contagion and diffusion of innovations: new products and ideas spread through a population by person-to-person communication over time, the process that the formal model of Bass (1969) partitions into innovators (external influence) and imitators (internal, word-of-mouth influence). Diffusion theory supplies the contagion hypothesis that the cascade literature (Section 31.5) then tests and qualifies, and it is the reason an S-curve is so often, and so often wrongly, read as evidence of transmission. Network structure enters through the strength of weak ties(Granovetter 1973): novel information flows disproportionately across bridging ties that connect otherwise separate clusters, which is why tie structure governs whether individual relays compound into reach and why the weak-tie thesis organizes the seeding and tie-strength results below.
On the content side, three social-psychological theories account for why a person relays. Social influence and social proof hold that descriptive and injunctive norms shift behavior, often more than people anticipate (Cialdini, Influence, a monograph cited here by name; for the empirical norm result see Goldstein, Cialdini, and Griskevicius (2008)): people share, adopt, and review in part because others do, which is why consensus cues and visible popularity move transmission. The emotion/arousal account of sharing holds that the operative driver is not the valence of content but its physiological arousal: high-arousal states, in both positive and negative valence, potentiate sharing, while low-arousal states suppress it (Berger and Milkman 2012), a claim formalized in Equation 31.2. And social identity holds that consumption and communication are identity-expressive: people relay content that signals who they are and that affirms valued group memberships, the mechanism underneath the impression-management and social-bonding functions developed next (Tajfel and Turner’s social identity theory, an edited-volume statement cited by name). Together these theories predict both the probability that any individual relays content and the shape of the diffusion that results.
The empirical literature that operationalized these ideas developed in three waves, and it is worth knowing which question each was answering. The first established that online word of mouth predicts commercial outcomes at all: Yong Liu (2006) showed that pre-release and opening-week message volume forecasts box office, with volume rather than valence carrying most of the signal — a finding that reframed WOM from a soft construct into a measurable leading indicator. The second asked how transmission actually proceeds, decomposing viral marketing into the distinct stages of click-through, opening, and forwarding, each with its own drivers, so that a campaign optimized for one stage can fail at another (De Bruyn and Lilien 2008); the parallel question of what determines transmission within an online community rather than across a dyad is taken up in Brown, Broderick, and Lee (2007). The third turned to what firms can do about it. Godes et al. (2005) models the firm’s management of social interactions as a decision variable rather than an environmental parameter; Chen, Fay, and Wang (2011) distinguishes online from offline conversation and finds the two are less substitutable than practice assumes; and A. Kumar et al. (2016) traces firm-generated social-media content through to purchase, showing the path runs via brand consideration rather than directly to sale. Rapp et al. (2013) generalizes across seller, retailer, and consumer sources, which is the comparison most single-source studies cannot make.
31.3 Why People Share: A Functional Account
The single most useful organizing idea in the WOM literature is that sharing is goal-directed. People do not relay content at random; they relay it because doing so serves an end, often below the level of conscious awareness. Berger (2014) consolidates this into five functions that WOM serves, and—crucially—each function makes predictions about what kind of content will be shared. The framework is generative rather than descriptive: name the goal, and you can forecast the content that satisfies it. Figure 31.1 maps each function to the content properties it predicts will be shared.
flowchart LR
WOM["Word of mouth<br/>(goal-directed)"]
WOM --> IM["Impression<br/>management"]
WOM --> ER["Emotion<br/>regulation"]
WOM --> IA["Information<br/>acquisition"]
WOM --> SB["Social<br/>bonding"]
WOM --> P["Persuasion"]
IM --> IMc["entertaining, useful,<br/>status-, identity-relevant"]
ER --> ERc["emotional, high-arousal,<br/>valence-managed"]
IA --> IAc["risky / complex /<br/>uncertain decisions"]
SB --> SBc["common ground,<br/>shared emotion"]
P --> Pc["polarized valence,<br/>arousing"]
Figure 31.1: Five functions of word of mouth and the content properties each one predicts will be shared. Multiple functions can operate simultaneously on the same act of sharing.
That these are functions and not mere correlates is supported by the self-referential bias in everyday communication: over 70% of ordinary speech concerns the self, and the same proportion holds for social-media posts. People talk to accomplish something for themselves. We take the five functions in turn.
31.3.1 Impression management
Consumers share to curate how others see them. The mechanism rests on the premise that consumption is identity-expressive—people acquire and discuss goods partly to signal who they are—and it operates through self-enhancement (looking good), identity-signaling (advertising specific traits or expertise), and the mundane need to fill conversational space. The self-enhancement motive is strong enough that people will distort to satisfy it: roughly 60% of retold personal stories are altered in the retelling (Marsh and Tversky 2004), and people share content specifically because it burnishes their image (Wojnicki and Godes 2017).
The motive predicts a recognizable content profile. People preferentially share material that is entertaining (novel, surprising, funny, or extreme), because relaying it makes the sharer look interesting and in-the-know. Novelty must be paired with comprehensibility to function—novel-and-understandable is interesting, novel-and-incomprehensible is merely confusing(Berlyne 1960). The starkest evidence on novelty comes from Vosoughi, Roy, and Aral (2018), who tracked roughly 126,000 rumor cascades on Twitter, classified by six independent fact-checking organizations: false stories spread significantly farther, faster, deeper, and more broadly than true ones, and the gap is explained by the greater novelty of false content and the surprise and disgust it evokes. The effect survives controls for bot activity, implicating human sharers rather than automation.
Spread is only half of the misinformation problem; the other half is what the first few exposures do to the person who sees them. Puthineedi and Jha (2026) ran three controlled experiments in which participants scrolled Instagram-style posts about topics unfamiliar to them. Impressions stabilized after a short, bounded run of consistent posts—roughly five, by the authors’ account of the results (INFORMS 2026)—a “sufficiency” point they call the Point of Critical Information, after which further posts added little. The stance formed there then filtered what came next: congruent posts were more readily believed and shared, and contradictory ones were more easily dismissed. Early judgments tracked factual accuracy only weakly, since participants shown inaccurate information reacted much like those shown accurate information, so a correction that arrives after the stance has formed works against a frame already in place. Set beside the novelty account, this is an argument about sequence: what matters is not only how far a false story travels but whether it reaches a user first and consistently, which is why the authors favor upstream interventions (getting authoritative content into the earliest exposures) over later fact-checks. Only the published abstract and the INFORMS press release were accessible when this section was written; the five-post figure and the experimental design come from the release.
People also share what is useful (it makes the sharer helpful and smart) and what is status-relevant (premium brands and knowledge are discussed more, because both signal standing). De Angelis et al. (2012) separate the two ways self-enhancement operates: a sharer can inflate their own experience (bragging) or deflate someone else’s (gossip), and which route is taken depends on whether the sharer’s self-view has just been threatened or affirmed. A subtle reversal appears for uniqueness: unique products draw more WOM in general, yet consumers with a high need for uniqueness withhold positive WOM to keep adoption—and therefore distinctiveness—scarce (Cheema and Kaikati 2010). Finally, sharing is shaped by accessibility: products that are more frequently cued in the environment are discussed more, so heavily advertised products generate more WOM (Onishi and Manchanda 2012). Accessibility is also a route through which incidental states matter—unrelated physiological arousal raises sharing across the board (Berger and Schwartz 2011), a point we formalize under emotion below.
31.3.2 Emotion regulation
People share to manage how they feel—seeking social support, venting, making sense of an experience, reducing post-decision dissonance, taking revenge on a firm, or reliving a positive episode. Sharing a negative experience can raise well-being through perceived social support (Buechel and Berger 2017). This function predicts that emotional content is shared more, but with two refinements that the next subsection makes precise: not all emotions promote sharing (shame and guilt suppress it, because relaying them is self-incriminating), and the impression- management function can oppose emotion regulation when venting a negative episode would reflect badly on the sharer.
31.3.3 Information acquisition, social bonding, and persuasion
The remaining functions are quicker to state. WOM acquires information: rather than learn by trial and error, consumers solicit advice, especially for risky, complex, or uncertain decisions where talking reduces perceived risk. WOM builds social bonds: people share common ground and emotional content because both increase felt connection and group cohesion. And WOM persuades: when the goal is to move another’s choice, people share polarized (extremely positive or negative) and arousing content, because moderation does not convince. These functions are not mutually exclusive—a single share can manage an impression, regulate emotion, and bond simultaneously—which is why the five-function lens predicts content profiles rather than partitioning shares into types.
31.4 Drivers of Virality
The functional account tells us why people share; the empirical literature isolates which content features raise share probability. We group the robust drivers into social currency, emotion, usefulness, and narrative, and then treat the moderators—audience and channel—that determine when each driver dominates.
31.4.1 Social currency
The dominant driver of posting behavior is image-related utility rather than intrinsic enjoyment of the content: Toubia and Stephen (2013), in a field experiment that randomized whether contributions were attributed to their authors, find that image-related utility drives posting more than intrinsic utility. People share to accrue social currency—to look interesting, smart, or in-the-know. A consequence worth flagging is that “useful” and “interesting” are not interchangeable currencies: useful information signals helpfulness, but the signal is valued less than the signal of being entertaining (Berger 2014).
Social currency interacts with the channel, and the interaction is identifying. Interesting content is shared more in written (asynchronous) settings but not in oral (synchronous) ones. The mechanism is time to self-edit: asynchronous communication lets people polish what they say, which amplifies self-enhancement motives, whereas synchronous conversation must be filled in real time, pushing people toward mundane, accessible topics merely to avoid silence. This is a clean example of a moderator (channel) switching a driver (interestingness) on and off.
31.4.2 Emotion: valence and arousal
Emotion is the most-studied driver, and the literature’s central correction is that valence is not the operative variable—arousal is. It is tempting to predict sharing from whether content is positive or negative; the better predictor is how physiologically activating it is.
Berger and Milkman (2012) make the case with both archival and experimental evidence. Examining roughly 7,000 New York Times articles, they find that positive content is more viral on average, but emotionality matters over and above valence, and the result holds after controlling for usefulness, interestingness, surprise, prominence of placement, author fame, length, and timing. The experiments then dissociate valence from arousal: high-arousal negative emotions (anger, anxiety) raise sharing, while a low-arousal negative emotion (sadness) lowers it; high-arousal positive emotions (awe, excitement, amusement) raise sharing relative to a low-arousal positive state (contentment). The clinching test is that incidental arousal— arousal induced by a task unrelated to the content, such as physical exertion—also raises sharing (Berger and Schwartz 2011), which it could not do if valence were the true cause.
The mechanism is physiological: arousal is an excitatory state that potentiates action of all kinds, including sharing and other prosocial behavior (Gaertner and Dovidio 1977). Different emotions sit at different points on the activation dimension (C. A. Smith and Ellsworth 1985), and the relevant contrast is activation (high arousal, behavioral readiness) versus deactivation (low arousal, relaxation) (Feldman Barrett and Russell 1998; Heilman 1997). Table 31.1 organizes the predictions; note that the regulation goal differs by quadrant—high-arousal negative states drive venting, high-arousal positive states drive rehearsal(Rimé et al. 1991).
Table 31.1: Sharing predicted by the arousal × valence quadrant. Arousal, not valence, is the operative dimension: high-arousal states in both valences raise sharing, while low-arousal states suppress it (Berger and Schwartz 2011; Berger and Milkman 2012).
Low arousal (deactivation)
High arousal (activation)
Positive valence
Contentment — less shared
Awe, excitement, amusement — more shared (rehearsal)
Negative valence
Sadness — less shared
Anger, anxiety — more shared (venting)
A formal statement makes the claim testable. Let \(a(c) \in [0,1]\) be the arousal a content piece evokes and \(\text{val}(c) \in [-1,1]\) its valence. The arousal model asserts \[
\frac{\partial \, v(c)}{\partial \, a(c)} > 0
\quad\text{and}\quad
\frac{\partial \, v(c)}{\partial \, |\text{val}(c)|} \;\text{small once } a(c)\text{ is held fixed},
\tag{31.2}\] i.e., once arousal is conditioned on, residual valence carries little additional explanatory power for sharing. Identification of Equation 31.2 is exactly what the incidental-arousal manipulation buys: by moving \(a\) without moving any property of \(c\), the experiment severs the confound between arousing content and interesting or useful content, which would otherwise make \(a\) endogenous.
The managerial corollary is valuable virality(Akpinar and Berger 2017). Emotional ads are shared more but often fail to lift brand outcomes because the brand is incidental to the story; informative ads lift brand evaluation (the brand is integral to the message) but are shared less. The resolution is emotional, brand-integral content—embed the brand into an arousing narrative so that sharing and brand lift move together. The same prescription recurs in the advertising literature (Chapter 14): Tellis et al. (2019) find across field studies that information-focused content is shared less (except in risky contexts), that positive high-arousal emotions and “drama” elements (surprise, plot, characters, babies, animals, celebrities) raise sharing through arousal, that brand prominence matters, and that platform moderates fit—emotional ads travel on general platforms, informational ads on professional ones—with an optimal ad length of roughly 1.2 to 1.7 minutes. Emotion can be operationalized directly from language (Yin, Bond, and Zhang 2017), which makes these predictions deployable at scale.
31.4.3 Usefulness and narrative
Beyond emotion, practical utility is a robust driver: useful newspaper articles are more viral (Berger and Milkman 2012), and the broader pattern—useful, interesting, and emotionally arousing content is shared more—replicates in Milkman and Berger (2014), who add the striking finding that articles by women are shared more because they are written to be more comprehensible, useful, and interesting. Three theoretical lenses explain why usefulness travels: sharing useful content signals knowledge and so serves impression management (Wojnicki and Godes 2017); it generates reciprocity, the sharer expecting a return favor (Fehr, Kirchsteiger, and Riedl 1998); and it functions as social exchange with exchange value (Homans 1958). The channel again moderates: usefulness and novelty predict email forwarding most strongly, whereas emotional evocativeness and content familiarity predict social-media retransmission (Kim 2015), because the two channels serve different goals (private helpfulness versus public self-presentation).
Narrative is the vehicle that carries emotion and brand together. In social media, brands and consumers co-create brand performances, and the ongoing process of improvisation matters more than any single output; managing a brand becomes the work of keeping the performance alive (Singh and Sonnenburg 2012). Narrative is also where the valuable-virality prescription is executed—the brand rides inside the story rather than interrupting it.
31.4.4 Storytelling design: narrativity, emotional sequence, and reading level
“Narrative” is usually treated as a yes-or-no property of content. Ghose et al. (2026) unbundle it into design choices a publisher controls, and test them across four studies in Information Systems Research. The three features are narrativity (a linear, chronological story versus a summary-first structure that states the outcome up front), emotional sequence (good-to-bad versus bad-to-good), and reading level (simple versus complex language). The paper asks what each does to engagement, why (through perceived interestingness and the gratifications readers get), and when and for whom (news format, reading level, and consumer motivation as boundary conditions).
The design is the instructive part. Study 1 is observational: it programmatically codes more than 5,000 satirical and 35,000 traditional news stories that one publisher posted on Twitter, which holds the source constant while format varies. Studies 2 to 4 are controlled experiments in which large language models rewrite the same story to vary one feature at a time, so the manipulation is clean and cheap to scale (the SSRN version (Ghose et al. 2025) carries the full research abstract). The press coverage reports more than 500 participants and an Indian publisher as the source (Temple University 2026; Poynter Institute 2026).
The results are conditional rather than universal. For satirical news, a summary-first structure with a good-to-bad sequence works best when the language is simple, while a linear structure with a bad-to-good sequence works best when the language is complex. For traditional news, simple language helps and complex language hurts, but complex language paired with linear narrativity and a bad-to-good sequence recovers much of the loss. Motivation moderates the reading-level effect: high self-brand connection offsets the cognitive cost of hard text. Satire must also clear a higher interestingness threshold than traditional news before readers share it. The authors’ summary is that engagement depends on how the features combine, and for whom, not on any single tactic (Ghose et al. 2026; Poynter Institute 2026).
Three lessons carry beyond news. First, they refine the “emotion drives sharing” finding earlier in this chapter: the order of valence matters, and its sign depends on format and language, so a model with a single valence score discards signal. Second, they sit beside the linguistic-style literature in Section 31.10 as another manipulable, text-extractable lever distinct from content. Third, they show a method: using an LLM to generate matched variants gives a within-story manipulation that observational text cannot, and the authors report the same patterns for video, podcasts, and interactive media. (Only the abstracts, working-paper record, and press coverage were accessible when this section was written; the effect sizes are in the paper.)
31.4.5 The source’s stance as a signal
What a story says is only part of what readers take from it; they also draw inferences from who published it. Huang, Meng, and Weng (2026) study reposting by Chinese microblog users, combining field data with an online experiment that varies whether a story’s source and stance are disclosed. As expected, users repost stories that match their own attitudes more often. They are also more likely to repost stories that run against the reporting outlet’s usual stance. The authors’ explanation is strategic inference. An outlet trades popularity off against its stance, so a story that cuts against the outlet’s line was presumably worth running for its popularity, and readers infer greater popularity from the inconsistency. Two results support that reading. The effect is larger among users with stronger Bayesian-updating and level-k reasoning, the abilities the inference requires. And simply reminding users that media are slanted makes their responses more strategic. For sharing research, the implication is that the same content is a different signal depending on who carries it, so a model of share probability built only on content features leaves out the inference readers draw from the source. (Only the published abstract was accessible when this section was written; the effect sizes are in the paper.)
31.4.6 Perceptual features: visual motion and children’s viewing
Every driver so far works through what content means to a viewer who might pass it on: the status it confers, the emotion it evokes, the use it offers, the story it tells, and the source that carries it. The YouTube evidence fits that mold. Tellis et al. (2019) trace the sharing of online video ads to information, emotion, and brand prominence, and Feroz Khan and Vong (2014) find that view counts of the most-viewed YouTube videos depend on network links and the uploader’s offline fame as well as on the platform itself. Tucker (2015) adds a warning that views and persuasion can come apart: across 400 video ads, relative persuasiveness was on average 10% lower for every million views. The exceptions were ads that drew many comments or comments naming the product, which seem to have earned their views through humor rather than outrage. Even the regulatory work on YouTube engagement, such as the effect of FTC affiliate-disclosure rules on product review videos (Zhang, Lee, and Tan 2026), is about what viewers infer from content.
S. Kumar, Viswanathan, and Bapna (2026) isolate a channel that runs beneath meaning. Their audience is children, who, the authors argue, rely less on the social and participatory cues that carry adult video. For this audience the relay mechanism behind Equation 31.1 is largely absent, and views accumulate through how well a video holds attention. The feature they study is visual motion, measured with optical flow, the computer-vision estimate of how fast brightness patterns move between frames (Section 51.3.3). Figure 31.2 lays out the three-study design. In more than 32,000 videos from leading children’s YouTube channels, higher motion predicts viewership after controlling for visual, audio, and affective characteristics. When the authors build educational videos that differ only in background motion and run them in randomized ad campaigns through YouTube’s A/B testing infrastructure, the high-motion versions earn significantly higher view-rates and views. In a lab eye-tracking study with children, motion raises screen engagement mainly by cutting short long off-screen lapses, rather than by preventing attention from breaking at all. It draws more gaze to the background without reducing gaze on the instructional content. The authors read this as re-engagement at the level of the screen, not displacement of the lesson.
flowchart LR
S1["Study 1: observational<br/>32,000+ children's YouTube videos"] --> F1["Higher optical flow,<br/>more views (with controls)"]
F1 --> S2["Study 2: field experiment<br/>background motion manipulated,<br/>randomized YouTube ad campaigns"]
S2 --> F2["Higher view-rates<br/>and views"]
F2 --> S3["Study 3: lab eye-tracking<br/>with children"]
S3 --> F3["Shorter long off-screen lapses;<br/>more gaze on background,<br/>no loss on instructional content"]
Figure 31.2: The three-study design of Kumar, Viswanathan, and Bapna (2026). An observational association across children’s YouTube videos is replicated causally in a randomized field experiment and then traced to a mechanism with lab eye-tracking.
For the framework of this chapter, the paper adds a separate class of content features \(\mathbf{z}_c\). Perceptual properties such as motion, pace, and visual intensity act on attention directly, apart from the informational and emotional content that the Tellis et al. (2019) decomposition captures, so a model of video engagement that codes only meaning leaves out a lever creators actually pull. It also marks a boundary on the outcome. The dispositional virality of Equation 31.1 is a relay probability, while children’s viewership is closer to retention, and the features that drive the two need not coincide. The design is a model of triangulation: an observational association, a randomized field replication that rules out the recommender as the source of the correlation, and a lab measure of mechanism. Finally, the governance stakes are unusually direct. A feature that raises views by shortening children’s attentional lapses is commercially valuable to creators and platforms, and at the same time it is the kind of design choice that policy toward child-directed media and attention-capturing design is starting to scrutinize. (Only the published abstract and an earlier working-paper abstract were accessible when this section was written; the effect sizes are in the paper.)
31.5 Structural Virality
Everything so far concerns the per-person relay probability \(v(c)\). We now turn to what happens when those relays compose through a network, and here the field’s hard-won lesson is that viral is not the same as popular.
31.5.1 Cascades versus broadcasts
A piece of content can reach a million people two very different ways. It can be broadcast—a single hub (a celebrity, a front page) transmits directly to a mass audience, producing a wide but shallow diffusion tree. Or it can cascade—person relays to person relays to person, producing a deep, multi-generation tree. Both yield large reach; only the second is “viral” in the mechanistic sense. The total number of retweets is therefore a poor proxy for virality, because it cannot distinguish a hub blasting to its followers from a genuine person-to-person epidemic (Goel et al. 2015).
The empirical reality is that genuine cascades are vanishingly rare. In a corpus of roughly a billion diffusion events, about 90% of content was adopted by no one beyond the seed, and fewer than 1% of events achieved even a handful of second-generation adoptions (Goel, Watts, and Goldstein 2012). Virality, in the strict sense, is an extreme-tail phenomenon; most content dies at the source.
Goel et al. (2015) formalize the broadcast–cascade distinction with a single statistic, structural virality, defined as the average distance between all pairs of nodes in a diffusion tree (the Wiener index, normalized by tree size). For a tree \(T\) on \(n\) nodes with shortest-path distances \(d_{ij}\), \[
\nu(T) \;=\; \frac{1}{n(n-1)} \sum_{i=1}^{n} \sum_{j=1}^{n} d_{ij}.
\tag{31.3}\] A pure broadcast (a star: one seed, \(n-1\) direct children) has small \(\nu\)—every pair is at most two hops apart—no matter how large \(n\) grows. A deep chain of person-to-person relays has large \(\nu\). The statistic thus separates size from shape: two events with identical reach can have wildly different structural virality. The following code makes the contrast concrete by computing Equation 31.3 for a broadcast star and for a deep cascade of equal size.
Code
set.seed(606)# Mean pairwise distance (Wiener index / [n(n-1)]) on a tree given as a# parent vector: parent[i] is the index of node i's parent (NA for the root).structural_virality<-function(parent){n<-length(parent)# Build n x n shortest-path distances via repeated parent-walks (tree => unique path).depth<-integer(n)path<-vector("list", n)for(iinseq_len(n)){p<-i; chain<-iwhile(!is.na(parent[p])){p<-parent[p]; chain<-c(chain, p)}path[[i]]<-chaindepth[i]<-length(chain)-1L}D<-matrix(0, n, n)for(iinseq_len(n))for(jinseq_len(n)){common<-intersect(path[[i]], path[[j]])# nearest common ancestor setlca_depth<-max(depth[common])D[i, j]<-(depth[i]-lca_depth)+(depth[j]-lca_depth)}sum(D)/(n*(n-1))}n<-40# Broadcast: one root, everyone else attaches directly to it (a star).broadcast<-c(NA, rep(1L, n-1))# Cascade: a single deep chain, 1 -> 2 -> ... -> n.cascade<-c(NA, 1:(n-1))c(broadcast =structural_virality(broadcast), cascade =structural_virality(cascade))#> broadcast cascade #> 1.95000 13.66667
The broadcast tree returns a structural-virality near 2 (almost all paths route through the hub), while the chain returns a value that grows with depth—mechanically identical reach, opposite diffusion shape. Real cascades fall between these poles, and Goel et al. (2015) document that large structural virality does occur in the wild but is rare and only weakly predictable from content alone.
31.5.2 The limits of prediction: skill, luck, and an \(R^2\) ceiling
Can we predict which content goes viral? Goel et al. (2015) give the question a sharp, discouraging answer by decomposing realized cascade success \(S\) into a part explained by observable “skill” \(Q\) (content quality, seed characteristics) and an irreducible “luck” component \(E\), \[
S \;=\; f(Q) + E.
\tag{31.4}\] The natural measure of predictability is how much variance in \(S\) survives after conditioning on everything knowable, \(Q\): \[
F \;=\; \frac{\mathbb{E}\!\left[\,\operatorname{Var}(S \mid Q)\,\right]}{\operatorname{Var}(S)} \;=\; 1 - R^2 .
\tag{31.5}\] If \(R^2 \to 1\), success is pure skill and perfectly forecastable; if \(R^2 \to 0\), it is pure luck and unforecastable. The substantive result is that there is a theoretical ceiling on \(R^2\) that is well below one, arising from the intrinsic stochasticity of the branching process: the more sensitive the cascade is to the initial seed, the larger the variance in outcomes for fixed \(Q\), and the lower the attainable \(R^2\). Virality is partly designable and partly irreducibly chancy, and no amount of feature engineering closes the gap. This is the structural counterpart to the funnel leakage of V. Kumar, Petersen, and P.Leone (2007): even perfect content does not guarantee a cascade.
31.5.3 What the S-curve does and does not tell you
Diffusion is conventionally pictured as an S-shaped cumulative-adoption curve, and it is tempting to read an S-curve as evidence of contagion—of person-to-person transmission. This inference is invalid, for two distinct reasons that any identification strategy must confront.
First, an S-curve does not imply contagion. The same sigmoid can be generated by mechanisms that involve no social transmission at all. Population heterogeneity in adoption thresholds produces an S-curve mechanically as the marginal adopter’s threshold is crossed over time (Chatterjee and Eliashberg 1990), and time-varying marketing effort can trace out the identical shape (Bulte and Lilien 2001). Attributing an S-curve to word of mouth therefore requires ruling out heterogeneity and marketing as alternative explanations—a problem of separating contagion from correlated unobservables, addressed structurally in Chapter 26 and in the diffusion-modeling literature.
Second, selection (survivorship) bias contaminates retrospective virality studies: we observe the cascades that happened to grow and rarely the near-identical content that died, so any pattern estimated on “viral” cases is conditioned on success. Combined with the rarity result of Goel, Watts, and Goldstein (2012)—90% of events go nowhere—this means naive comparisons of viral and non-viral content overstate whatever distinguishes the survivors. Figure 31.3 collects the competing data-generating processes that yield the same aggregate curve.
flowchart TD
S["Observed S-shaped<br/>adoption curve"]
S --> C["Social contagion<br/>(person-to-person)"]
S --> H["Population heterogeneity<br/>in thresholds"]
S --> M["Time-varying<br/>marketing effort"]
C -.->|"observationally<br/>equivalent"| H
H -.-> M
SB["Survivorship bias:<br/>dead cascades unobserved"] --> S
Figure 31.3: Why an observed S-shaped adoption curve is not evidence of contagion. At least three distinct data-generating processes produce the same aggregate shape, and survivorship in the sample compounds the inference problem.
31.6 Network Position, Ties, and Seeding
If structure governs whether relays compound, then where in the network a message starts and which ties it crosses should matter. This section covers the position/seeding literature and the surprisingly subtle role of tie strength.
31.6.1 Seeding strategy
The seeding question is whether a firm should inject content at well-connected hubs or at ordinary nodes. Analytical models and simulations had suggested that targeting hubs does not pay—hubs are no more influential over any given peer than ordinary nodes. Hinz et al. (2011) overturn this in small-scale field experiments and real campaigns: seeding to well-connected people yields up to eight times more adoptions, not because hubs are more persuasive per contact but because they are more active and exploit their greater reach. The lesson is that seeding returns come from reach and activity, not from per-tie influence—a distinction that matters because it cannot be recovered from peer-level persuasion data alone. Seeding can also have competitive effects: a well-placed campaign raises focal-brand WOM while suppressing conversation about rival products in the same category (Chae et al. 2017).
31.6.2 Network management and identification
Studying how a firm’s own networking activity drives online success raises an acute endogeneity problem: firms that succeed may simply network more, so position and success are jointly determined. Ansari et al. (2018) model the co-evolution of an actor’s ego-network and success with stochastic actor-based models, characterizing each actor by ego-network density and degree, betweenness, and closeness centrality. To identify the effect of networking on success they combine time fixed effects with a control-function approach instrumented by the firm’s activity on an independent US social platform—activity correlated with the firm’s European networking effort but plausibly excludable from European-market success. The design is a template for network studies generally: position is endogenous, and credible identification needs an instrument that shifts position without directly shifting the outcome.
31.6.3 Tie strength and the strength of weak ties
Granovetter’s classic thesis is that weak ties—acquaintances rather than close friends—are disproportionately valuable for information diffusion, because they bridge otherwise disconnected clusters and so carry novel information. The modern e-WOM literature both confirms and qualifies this. (The positional version of the argument, in which the advantage belongs to the node that bridges rather than to the tie, is the structural-holes tradition; Section 52.3 covers its measures, including the pair-based opportunity and exclusivity scores of Borgatti and Everett (2026), which separate how many disconnected pairs a node spans from how many rivals span the same pairs.)
On the confirmation side, Peng et al. (2018) show that sharing propensity rises with network overlap between sender and receiver—common followees, common followers, and common mutual followers—using a dyadic hazard model on Twitter and Digg data. High overlap signals shared interests, so the receiver infers the content suits her own audience and relays it; popular content, conversely, is shared less, because many others already have. Tie strength also conditions persuasion: Lee and Kronrod (2020) find that consensus language—expressions implying widespread agreement—is more persuasive coming from weak ties than strong ones, because a weak tie evokes a broader, more diverse agreeing group and thus signals higher validity, providing social proof.
The sharpest recent test qualifies the weak-tie thesis itself. Rajkumar et al. (2022) exploit large-scale randomized experiments on LinkedIn’s “People You May Know” algorithm—rare causal variation in tie formation—to estimate how tie strength affects job mobility. They find that weak ties do increase job transmission but with diminishing returns: the relationship is an inverted U, not monotone. Moderately weak ties, not the weakest possible, maximize mobility, and the effect is industry-dependent—weak ties dominate in digital industries, strong ties in less digital ones. The headline revision is that “the strength of weak ties” is nonlinear and context-contingent, not a universal law.
Table 31.2 contrasts the three studies, which agree that tie structure matters but disagree on the direction once the outcome and mechanism change.
Table 31.2: How tie strength enters WOM and diffusion across three studies. The apparent disagreements reflect different outcomes (sharing, persuasion, mobility) and identification strategies (observational dyadic models versus randomized tie formation).
Inverted-U; moderately weak ties maximize job mobility
31.7 Accessibility and the Mere-Exposure Channel
A driver that operates orthogonally to emotion is accessibility: content and brands that are easier to retrieve from memory are talked about more, however mundane they are. The theoretical basis is spreading activation in semantic memory: priming activates related constructs, and a primed brand becomes more likely to enter the consideration set (Berger and Fitzsimons 2008). Berger and Fitzsimons (2008) show that more frequent exposure to perceptually or conceptually related cues raises product accessibility and, through it, evaluation and choice—conceptual priming with real consumer consequences. Accessibility is why publicly visible and heavily advertised products sustain ongoing WOM (Onishi and Manchanda 2012), and why familiar content is selected more often for retransmission (Kim 2015). It also supplies an alternative, non-emotional reason content travels, which any well-specified sharing model must include as a control lest the emotion coefficients absorb a pure-accessibility effect.
31.8 Moderators: Audience and Channel
Which WOM function dominates—and therefore which content travels—depends on two moderators introduced above and worth consolidating. The audience is characterized by tie strength, audience size, and the status of the tie; the channel by whether communication is written or oral, whether the communicator is identifiable, and whether the audience is salient during composition. The recurring result is that asynchronous, identifiable, audience-salient channels amplify self-presentational motives and so favor interesting, status-relevant content, whereas synchronous channels favor accessible, mundane content that fills conversational time. Demographics moderate too: the 18–24 cohort responds less to viral content because saturation raises their activation threshold, making their emotions harder to move (Libert 2014). And conversation has its own dynamics—a wide gap between two opposing opinions stimulates future discussion, whereas fluctuation around a single opinion does not, with informational content strengthening the opposition-to-discussion link (Nagle and Riedl 2014).
31.9 Cultural and Dispositional Antecedents
Sharing is not culturally invariant. Two constructs from the cross-cultural WOM literature illustrate how disposition shapes transmission.
Horizontal/vertical individualism–collectivism shapes what communicative role people adopt. Singelis et al. (1995) cross individualism–collectivism with a horizontal (equality-emphasizing) versus vertical (hierarchy-emphasizing) dimension, yielding four cultural patterns. Applied to WOM, Choi and Kim (2019) find that vertical orientations—where standing is signaled by display—channel into both opinion leadership and opinion seeking, while horizontal orientations feed opinion leadership through a norm of being socially appropriate. The broader regularity is that collectivist cultures do more opinion seeking and individualist cultures more information giving(Fong and Burton 2008).
31.10 Linguistic Style
Because e-WOM is text, how something is said is measurable and consequential, and a sizable literature maps linguistic features onto WOM outcomes. The features that have been shown to matter include the abstraction level of language (Schellekens, Verlegh, and Smidts 2010), the explanatory framing a storyteller uses—which feeds back to shape the storyteller’s own attitudes (Moore 2012)—figurative language (Kronrod and Danziger 2010), markers of modesty versus boastfulness (Packard, Gershoff, and Wooten 2016), “dispreferred” markers that soften disagreement (Hamilton, Vohs, and McGill 2014), the personal pronouns a communicator selects (Packard and Wooten 2013), and linguistic mimicry between parties (Moore and McFerran 2017). The practical upshot is that the style of a message is a manipulable lever distinct from its content, and—because style is extractable from text at scale—an empirically tractable one.
31.11 Negative Virality and Firestorms
Virality is not always the firm’s friend. Negative e-WOM can compound into a firestorm—a sudden, large-scale wave of hostile messages. Herhausen et al. (2019) supply a framework for the drivers of negative e-WOM and the firm responses that mitigate it, identifying the level of arousal, structural tie strength, and linguistic-style match as the factors that fuel a firestorm. Their central prescription follows directly from the arousal account of Equation 31.2: the firm’s response should match the intensity of arousal in the negative e-WOM—neither under- nor over-reacting—because mismatched responses either dismiss aggrieved high-arousal customers or inflame low-arousal ones. Negative virality thus obeys the same arousal mechanics as positive virality, with the valence of the firm’s exposure reversed.
31.12 Contractuality: Engineering Authentic WOM
Firms that try to manufacture WOM through incentives confront a credibility trade-off captured by contractuality, defined by Lisjak, Bonezzi, and Rucker (2021) as “the extent to which a perk is perceived to be conditional on specific behaviors and contingencies dictated by a company.” Consumers prefer perks with less salient contractuality, and a prior negative relationship with the firm leads them to read a high- contractuality perk as a manipulative quid pro quo with ulterior motives. Low-contractuality perks foster authentic WOM but trade off against effectiveness and efficiency—the firm gives up the control that an explicit share-for-reward contract would provide. The construct formalizes the intuition that the most valuable WOM is the kind the firm can least directly buy.
Contractuality asks whether the sender was paid. A second credibility test asks whether the sender is doing what they say, and it becomes observable on platforms that publish both. Ye Liu et al. (2026) exploit a social trading platform where a user’s posted opinions and actual trades are visible side by side, and study early cryptocurrency adoption—a setting chosen because it is exactly the high-uncertainty, high-return regime where social learning should matter most.
Peers’ adoption raises an individual’s probability of adopting, which is the baseline social-learning result. The informative part is the interaction: the effect is markedly stronger when a peer’s trading behavior aligns with their expressed opinion. Talk that is corroborated by observable action carries the signal; talk that is not, does not. This is the same economics as contractuality approached from the other side—there, credibility is threatened by an incentive the audience infers; here, it is established by a costly action the audience can verify.
Two boundary conditions sharpen it. When uncertainty or returns run high, investors lean on positive opinions regardless of whether the poster’s trades back them up—the alignment check is a luxury that gets abandoned under exactly the conditions that make it most valuable. And the responsiveness is heterogeneous in a direction worth noting: older, more experienced, higher-income, more risk-tolerant users respond most to credible aligned signals, while younger users are more skeptical of positive sentiment, plausibly because they have more experience with noisy social media. For a firm seeking authentic WOM about a new technology, the design implication is architectural rather than contractual—make advocates’ behavior visible alongside their advocacy, and the credibility follows without a contract to disclose.
31.13 Empirical Regularities and Open Problems
A handful of well-replicated facts and one persistent confound deserve listing as the chapter’s empirical bottom line. WOM dispersion predicts higher future ratings, but its impact declines over the product life cycle, so it should be measured early (Godes and Mayzlin 2009)—WOM is both a precursor and a consequence of behavior, which complicates causal reading. Contagion is detectable even in consumer packaged goods (Du and Kamakura 2011), a category once thought too low-involvement for social transmission. Authorship matters: articles by business academics, psychologists, and economists are shared more than those by physicists and biochemists, and—net of content—men rate scientific summaries as more comprehensible, interesting, and useful and share them more (Milkman and Berger 2014). And identifying influential users in a network is itself a modeling problem with no off-the-shelf solution (Trusov, Bodapati, and Bucklin 2010).
The persistent confound is structural virality itself. Many sharing experiments—including the influential arousal studies—ask respondents how likely they would be to share, an intention that (i) carries little social risk in the lab, inflating the estimate, and (ii) abstracts away the network through which real diffusion runs. Two pieces of content with identical per-person share probability \(v(c)\) can diffuse to radically different sizes depending on where they seed and how the branching process resolves (Equation 31.4, Equation 31.5). The frontier of the field is precisely the integration we have kept separate for expository clarity: joining the content model of \(v(c)\) to the structural model of how \(v(c)\) propagates, so that designable content features and irreducible network stochasticity are estimated in one frame rather than two.
31.14 Key Takeaways
Word of mouth is goal-directed. The five functions—impression management, emotion regulation, information acquisition, social bonding, persuasion—predict what content travels, not merely that sharing happens (Berger 2014).
Arousal, not valence, drives sharing. High-activation states in both valences raise transmission; incidental-arousal manipulations identify the effect by moving arousal without moving content (Berger and Schwartz 2011; Berger and Milkman 2012).
Viral \(\neq\) popular. Broadcasts and cascades both reach many people; only cascades are structurally viral, and genuine cascades are extreme-tail rare (Goel, Watts, and Goldstein 2012; Goel et al. 2015).
Predictability is capped. Cascade success decomposes into skill and luck, and the irreducible luck component imposes a theoretical \(R^2\) ceiling
(Equation 31.5) (Goel et al. 2015).
Seeding pays through reach, not per-tie influence, and tie-strength effects are nonlinear and context-dependent (Hinz et al. 2011; Rajkumar et al. 2022).
Akpinar, Ezgi, and Jonah Berger. 2017. “Valuable Virality.”Journal of Marketing Research 54 (2): 318–30. https://doi.org/10.1509/jmr.13.0350.
Ansari, Asim, Florian Stahl, Mark Heitmann, and Lucas Bremer. 2018. “Building a Social Network for Success.”Journal of Marketing Research 55 (3): 321–38. https://doi.org/10.1509/jmr.12.0417.
Bampo, Mauro, Michael T. Ewing, Dineli R. Mather, David Stewart, and Mark Wallace. 2008. “The Effects of the Social Structure of Digital Networks on Viral Marketing Performance.”Information Systems Research 19 (3): 273–90. https://doi.org/10.1287/isre.1070.0152.
Berger, Jonah. 2014. “Word of Mouth and Interpersonal Communication: A Review and Directions for Future Research.”Journal of Consumer Psychology 24 (4): 586–607. https://doi.org/10.1016/j.jcps.2014.05.002.
Berger, Jonah, and Gráinne Fitzsimons. 2008. “Dogs on the Street, Pumas on Your Feet: How Cues in the Environment Influence Product Evaluation and Choice.”Journal of Marketing Research 45 (1): 1–14. https://doi.org/10.1509/jmkr.45.1.1.
Berger, Jonah, and Katherine L. Milkman. 2012. “What Makes Online Content Viral?”Journal of Marketing Research 49 (2): 192–205. https://doi.org/10.1509/jmr.10.0353.
Berger, Jonah, and Eric M. Schwartz. 2011. “What Drives Immediate and Ongoing Word of Mouth?”Journal of Marketing Research 48 (5): 869–80. https://doi.org/10.1509/jmkr.48.5.869.
Berlyne, D. E. 1960. “Toward a Theory of Exploratory Behavior: I. Arousal and Drive.” In Conflict, Arousal, and Curiosity., 163–92. McGraw-Hill Book Company. https://doi.org/10.1037/11164-007.
Borgatti, Stephen P., and Martin G. Everett. 2026. “Beyond Constraint: Measuring Brokerage Opportunity and Exclusivity.”Strategy Science. https://doi.org/10.1287/stsc.2026.0799.
Brown, Jo, Amanda J. Broderick, and Nick Lee. 2007. “Word of Mouth Communication Within Online Communities: Conceptualizing the Online Social Network.”Journal of Interactive Marketing 21 (3): 2–20. https://doi.org/10.1002/dir.20082.
Buechel, Eva C., and Jonah Berger. 2017. “Microblogging and the Value of Undirected Communication.”Journal of Consumer Psychology 28 (1): 40–55. https://doi.org/10.1002/jcpy.1013.
Bulte, Christophe Van den, and Gary L. Lilien. 2001. “Medical Innovation Revisited: Social Contagion Versus Marketing Effort.”American Journal of Sociology 106 (5): 1409–35. https://doi.org/10.1086/320819.
Camarero, Carmen, and Rebeca San José. 2011. “Social and Attitudinal Determinants of Viral Marketing Dynamics.”Computers in Human Behavior 27 (6): 2292–2300. https://doi.org/10.1016/j.chb.2011.07.008.
Chae, Inyoung, Andrew T. Stephen, Yakov Bart, and Dai Yao. 2017. “Spillover Effects in Seeded Word-of-Mouth Marketing Campaigns.”Marketing Science 36 (1): 89–104. https://doi.org/10.1287/mksc.2016.1001.
Chatterjee, Rabik Ar, and Jehoshua Eliashberg. 1990. “The Innovation Diffusion Process in a Heterogeneous Population: A Micromodeling Approach.”Management Science 36 (9): 1057–79. https://doi.org/10.1287/mnsc.36.9.1057.
Cheema, Amar, and Andrew M. Kaikati. 2010. “The Effect of Need for Uniqueness on Word of Mouth.”Journal of Marketing Research 47 (3): 553–63. https://doi.org/10.1509/jmkr.47.3.553.
Chen, Yubo, Scott Fay, and Qi Wang. 2011. “The Role of Marketing in Social Media: How Online Consumer Reviews Evolve.”Journal of Interactive Marketing 25 (2): 85–94. https://doi.org/10.1016/j.intmar.2011.01.003.
Choi, Youngtae, and Junga Kim. 2019. “Influence of Cultural Orientations on Electronic Word-of-Mouth (eWOM) in Social Media.”Journal of Intercultural Communication Research 48 (3): 292–313. https://doi.org/10.1080/17475759.2019.1627388.
De Angelis, Matteo, Andrea Bonezzi, Alessandro M. Peluso, Derek D. Rucker, and Michele Costabile. 2012. “On Braggarts and Gossips: A Self-Enhancement Account of Word-of-Mouth Generation and Transmission.”Journal of Marketing Research 49 (4): 551–63. https://doi.org/10.1509/jmr.11.0136.
De Bruyn, Arnaud, and Gary L. Lilien. 2008. “A Multi-Stage Model of Word-of-Mouth Influence Through Viral Marketing.”International Journal of Research in Marketing 25 (3): 151–63. https://doi.org/10.1016/j.ijresmar.2008.03.004.
Du, Rex Yuxing, and Wagner A. Kamakura. 2011. “Measuring Contagion in the Diffusion of Consumer Packaged Goods.”Journal of Marketing Research 48 (1): 28–47. https://doi.org/10.1509/jmkr.48.1.28.
Fehr, Ernst, Georg Kirchsteiger, and Arno Riedl. 1998. “Gift Exchange and Reciprocity in Competitive Experimental Markets.”European Economic Review 42 (1): 1–34. https://doi.org/10.1016/s0014-2921(96)00051-7.
Feldman Barrett, Lisa, and James A. Russell. 1998. “Independence and Bipolarity in the Structure of Current Affect.”Journal of Personality and Social Psychology 74 (4): 967–84. https://doi.org/10.1037/0022-3514.74.4.967.
Feroz Khan, Gohar, and Sokha Vong. 2014. “Virality over YouTube: An Empirical Analysis.”Internet Research 24 (5): 629–47. https://doi.org/10.1108/intr-05-2013-0085.
Fong, John, and Suzan Burton. 2008. “A Cross-Cultural Comparison of Electronic Word-of-Mouth and Country-of-Origin Effects.”Journal of Business Research 61 (3): 233–42. https://doi.org/10.1016/j.jbusres.2007.06.015.
Gaertner, Samuel L., and John F. Dovidio. 1977. “The Subtlety of White Racism, Arousal, and Helping Behavior.”Journal of Personality and Social Psychology 35 (10): 691–707. https://doi.org/10.1037/0022-3514.35.10.691.
Ghose, Debashish, Susan M. Mudambi, Subodha Kumar, and Joydeep Srivastava. 2025. “How to Tell a (News) Story? Quantifying the Impact of News Format and Storytelling on Engagement.” Working Paper. SSRN. https://doi.org/10.2139/ssrn.5649030.
———. 2026. “How to Tell a (News) Story? Quantifying the Impact of News Format and Storytelling on Engagement.”Information Systems Research 37 (3): 1363–89. https://doi.org/10.1287/isre.2022.0650.
Godes, David, and Dina Mayzlin. 2009. “Firm-Created Word-of-Mouth Communication: Evidence from a Field Test.”Marketing Science 28 (4): 721–39. https://doi.org/10.1287/mksc.1080.0444.
Godes, David, Dina Mayzlin, Yubo Chen, Sanjiv Das, Chrysanthos Dellarocas, Bruce Pfeiffer, Barak Libai, Subrata Sen, Mengze Shi, and Peeter Verlegh. 2005. “The Firm’s Management of Social Interactions.”Marketing Letters 16 (3-4): 415–28. https://doi.org/10.1007/s11002-005-5902-4.
Goel, Sharad, Ashton Anderson, Jake Hofman, and Duncan J. Watts. 2015. “The Structural Virality of Online Diffusion.”Management Science, July, 150722112809007. https://doi.org/10.1287/mnsc.2015.2158.
Goel, Sharad, Duncan J. Watts, and Daniel G. Goldstein. 2012. “The Structure of Online Diffusion Networks.” In Proceedings of the 13th ACM Conference on Electronic Commerce - EC 12. ACM Press. https://doi.org/10.1145/2229012.2229058.
Goldstein, Noah J., Robert B. Cialdini, and Vladas Griskevicius. 2008. “A Room with a Viewpoint: Using Social Norms to Motivate Environmental Conservation in Hotels.”Journal of Consumer Research 35 (3): 472–82. https://doi.org/10.1086/586910.
Granovetter, Mark S. 1973. “The Strength of Weak Ties.”American Journal of Sociology 78 (6): 1360–80. https://doi.org/10.1086/225469.
Hamilton, Ryan, Kathleen D. Vohs, and Ann L. McGill. 2014. “We’ll Be Honest, This Won’t Be the Best Article You’ll Ever Read: The Use of Dispreferred Markers in Word-of-Mouth Communication.”Journal of Consumer Research 41 (1): 197–212. https://doi.org/10.1086/675926.
Heilman, Kenneth. 1997. “The Neurobiology of Emotional Experience.”The Journal of Neuropsychiatry and Clinical Neurosciences 9 (3): 439–48. https://doi.org/10.1176/jnp.9.3.439.
Herhausen, Dennis, Stephan Ludwig, Dhruv Grewal, Jochen Wulf, and Marcus Schoegel. 2019. “Detecting, Preventing, and Mitigating Online Firestorms in Brand Communities.”Journal of Marketing 83 (3): 1–21. https://doi.org/10.1177/0022242918822300.
Hinz, Oliver, Bernd Skiera, Christian Barrot, and Jan U. Becker. 2011. “Seeding Strategies for Viral Marketing: An Empirical Comparison.”Journal of Marketing 75 (6): 55–71. https://doi.org/10.1509/jm.10.0088.
Hoffman, Donna L., Thomas P. Novak, and Ann E. Schlosser. 2003. “Locus of Control, Web Use, and Consumer Attitudes Toward Internet Regulation.”Journal of Public Policy and Marketing 22 (1): 41–57. https://doi.org/10.1509/jppm.22.1.41.17628.
Homans, George C. 1958. “Social Behavior as Exchange.”American Journal of Sociology 63 (6): 597–606. https://doi.org/10.1086/222355.
Huang, Yihong, Juanjuan Meng, and Xi Weng. 2026. “How Awareness of Media Slant Affects Reaction to News: Field and Experimental Evidence from Microblog Users.”Management Science. https://doi.org/10.1287/mnsc.2024.07339.
INFORMS. 2026. “Just Five Posts May Be Enough to Shape What People Believe Online, New Study Finds.” News release, distributed via EurekAlert! https://www.eurekalert.org/news-releases/1129596.
Kim, Hyun Suk. 2015. “Attracting Views and Going Viral: How Message Features and News-Sharing Channels Affect Health News Diffusion.”Journal of Communication 65 (3): 512–34. https://doi.org/10.1111/jcom.12160.
Kronrod, Ann, and Shai Danziger. 2010. “The Effect of Figurative Language in Consumer Online Reviews.” American Psychological Association (APA). https://doi.org/10.1037/e621072012-028.
Kumar, Ashish, Ram Bezawada, Rishika Rishika, Ramkumar Janakiraman, and P. K. Kannan. 2016. “From Social to Sale: The Effects of Firm-Generated Content in Social Media on Customer Behavior.”Journal of Marketing 80 (1): 7–25. https://doi.org/10.1509/jm.14.0249.
Kumar, Sumeet, Madhu Viswanathan, and Ravi Bapna. 2026. “Attention Trap? How Visual Motion Shapes Children’s Screen Time and Engagement on Video Platforms.”Information Systems Research. https://doi.org/10.1287/isre.2023.0718.
Kumar, V., Andrew Petersen, and Robert P.Leone. 2007. “How Valuable Is Word of Mouth?”Harvard Business Review.
Lachman, Margie E., and Rona Leff. 1989. “Perceived Control and Intellectual Functioning in the Elderly: A 5-Year Longitudinal Study.”Developmental Psychology 25 (5): 722–28. https://doi.org/10.1037/0012-1649.25.5.722.
Lam, Desmond, and Dick Mizerski. 2005. “The Effects of Locus of Control on Word-of-Mouth Communication.”Journal of Marketing Communications 11 (3): 215–28. https://doi.org/10.1080/1352726042000333180.
Lee, Jeffrey K., and Ann Kronrod. 2020. “The Strength of Weak-Tie Consensus Language.”Journal of Marketing Research 57 (2): 353–74. https://doi.org/10.1177/0022243720904957.
Libert, Kelsey. 2014. “Age and Gender Matter in Viral Marketing.”Harvard Business Review.
Lisjak, Monika, Andrea Bonezzi, and Derek Rucker. 2021. “EXPRESS: How Marketing Perks Influence Word-of-Mouth.”Journal of Marketing, January, 002224292199179. https://doi.org/10.1177/0022242921991798.
Liu, Ye, Mingwen Yang, Matthias Pelster, and Yong Tan. 2026. “The Interplay of Opinions and Behavior in Social Trading: The Dynamics of Early Cryptocurrency Adoption.”Information Systems Research. https://doi.org/10.1287/isre.2024.1300.
Liu, Yong. 2006. “Word of Mouth for Movies: Its Dynamics and Impact on Box Office Revenue.”Journal of Marketing 70 (3): 74–89. https://doi.org/10.1509/jmkg.70.3.074.
Marsh, Elizabeth J., and Barbara Tversky. 2004. “Spinning the Stories of Our Lives.”Applied Cognitive Psychology 18 (5): 491–503. https://doi.org/10.1002/acp.1001.
Milkman, K. L., and J. Berger. 2014. “The Science of Sharing and the Sharing of Science.”Proceedings of the National Academy of Sciences 111 (Supplement_4): 13642–49. https://doi.org/10.1073/pnas.1317511111.
Moore, Sarah G. 2012. “Some Things Are Better Left Unsaid: How Word of Mouth Influences the Storyteller.”Journal of Consumer Research 38 (6): 1140–54. https://doi.org/10.1086/661891.
Moore, Sarah G., and Brent McFerran. 2017. “She Said, She Said: Differential Interpersonal Similarities Predict Unique Linguistic Mimicry in Online Word of Mouth.”Journal of the Association for Consumer Research 2 (2): 229–45. https://doi.org/10.1086/690942.
Nagle, Frank, and Christoph Riedl. 2014. “Online Word of Mouth and Product Quality Disagreement.”Academy of Management Proceedings 2014 (1): 15681. https://doi.org/10.5465/ambpp.2014.214.
Onishi, Hiroshi, and Puneet Manchanda. 2012. “Marketing Activity, Blogging and Sales.”International Journal of Research in Marketing 29 (3): 221–34. https://doi.org/10.1016/j.ijresmar.2011.11.003.
Packard, Grant, Andrew D. Gershoff, and David B. Wooten. 2016. “When Boastful Word of Mouth Helps Versus Hurts Social Perceptions and Persuasion.”Journal of Consumer Research 43 (1): 26–43. https://doi.org/10.1093/jcr/ucw009.
Packard, Grant, and David B. Wooten. 2013. “Compensatory Knowledge Signaling in Consumer Word-of-Mouth.”Journal of Consumer Psychology 23 (4): 434–50. https://doi.org/10.1016/j.jcps.2013.05.002.
Peng, Jing, Ashish Agarwal, Kartik Hosanagar, and Raghuram Iyengar. 2018. “Network Overlap and Content Sharing on Social Media Platforms.”Journal of Marketing Research 55 (4): 571–85. https://doi.org/10.1509/jmr.14.0643.
Phelps, Joseph E., Regina Lewis, Lynne Mobilio, David Perry, and Niranjan Raman. 2004. “Viral Marketing or Electronic Word-of-Mouth Advertising: Examining Consumer Responses and Motivations to Pass Along Email.”Journal of Advertising Research 44 (4): 333–48. https://doi.org/10.1017/s0021849904040371.
Puthineedi, Venu, and Ashish Kumar Jha. 2026. “Where the Ball Starts Rolling? An Empirical Investigation into Initial Opinion Formation on Social Media Platforms.”Information Systems Research 37 (3): 2114–32. https://doi.org/10.1287/isre.2024.1589.
Rajkumar, Karthik, Guillaume Saint-Jacques, Iavor Bojinov, Erik Brynjolfsson, and Sinan Aral. 2022. “A Causal Test of the Strength of Weak Ties.”Science 377 (6612): 1304–10.
Rapp, Adam, Lauren Skinner Beitelspacher, Dhruv Grewal, and Douglas E. Hughes. 2013. “Understanding Social Media Effects Across Seller, Retailer, and Consumer Interactions.”Journal of the Academy of Marketing Science 41 (5): 547–66. https://doi.org/10.1007/s11747-013-0326-9.
Rimé, Bernard, Batja Mesquita, Stefano Boca, and Pierre Philippot. 1991. “Beyond the Emotional Event: Six Studies on the Social Sharing of Emotion.”Cognition and Emotion 5 (5-6): 435–65. https://doi.org/10.1080/02699939108411052.
Schellekens, Gaby A. C., Peeter W. J. Verlegh, and Ale Smidts. 2010. “Language Abstraction in Word of Mouth.”Journal of Consumer Research 37 (2): 207–23. https://doi.org/10.1086/651240.
Singelis, Theodore M., Harry C. Triandis, Dharm P. S. Bhawuk, and Michele J. Gelfand. 1995. “Horizontal and Vertical Dimensions of Individualism and Collectivism: A Theoretical and Measurement Refinement.”Cross-Cultural Research 29 (3): 240–75. https://doi.org/10.1177/106939719502900302.
Singh, Sangeeta, and Stephan Sonnenburg. 2012. “Brand Performances in Social Media.”Journal of Interactive Marketing 26 (4): 189–97. https://doi.org/10.1016/j.intmar.2012.04.001.
Smith, Craig A., and Phoebe C. Ellsworth. 1985. “Patterns of Cognitive Appraisal in Emotion.”Journal of Personality and Social Psychology 48 (4): 813–38. https://doi.org/10.1037/0022-3514.48.4.813.
Smith, Peter B., Shaun Dugan, and Fons Trompenaars. 1997. “Locus of Control and Affectivity by Gender and Occupational Status: A 14 Nation Study.”Sex Roles 36 (1-2): 51–77. https://doi.org/10.1007/bf02766238.
Tellis, Gerard J., Deborah J. MacInnis, Seshadri Tirunillai, and Yanwei Zhang. 2019. “What Drives Virality (Sharing) of Online Digital Content? The Critical Role of Information, Emotion, and Brand Prominence.”Journal of Marketing 83 (4): 1–20. https://doi.org/10.1177/0022242919841034.
Toubia, Olivier, and Andrew T. Stephen. 2013. “Intrinsic Vs. Image-Related Utility in Social Media: Why Do People Contribute Content to Twitter?”Marketing Science 32 (3): 368–92. https://doi.org/10.1287/mksc.2013.0773.
Trusov, Michael, Anand V. Bodapati, and Randolph E. Bucklin. 2010. “Determining Influential Users in Internet Social Networks.”Journal of Marketing Research 47 (4): 643–58. https://doi.org/10.1509/jmkr.47.4.643.
Vilpponen, Antti, Susanna Winter, and Sanna Sundqvist. 2006. “Electronic Word-of-Mouth in Online Environments.”Journal of Interactive Advertising 6 (2): 8–77. https://doi.org/10.1080/15252019.2006.10722120.
Vosoughi, Soroush, Deb Roy, and Sinan Aral. 2018. “The Spread of True and False News Online.”Science 359 (6380): 1146–51. https://doi.org/10.1126/science.aap9559.
Westbrook, Robert A. 1987. “Product/Consumption-Based Affective Responses and Postpurchase Processes.”Journal of Marketing Research 24 (3): 258. https://doi.org/10.2307/3151636.
Wojnicki, Andrea C., and David Godes. 2017. “Signaling Success: Word of Mouth as Self-Enhancement.”Customer Needs and Solutions 4 (4): 68–82. https://doi.org/10.1007/s40547-017-0077-8.
Yin, Dezhi, Samuel D. Bond, and Han Zhang. 2017. “Keep Your Cool or Let It Out: Nonlinear Effects of Expressed Arousal on Perceptions of Consumer Reviews.”Journal of Marketing Research 54 (3): 447–63. https://doi.org/10.1509/jmr.13.0379.
Zhang, Jingwen, Stephanie Lee, and Yong Tan. 2026. “The Effects of the FTC Policy and Affiliation Disclosures on Product Review Video Engagement: Evidence from YouTube.”Information Systems Research 37 (3): 1508–25. https://doi.org/10.1287/isre.2023.0158.
@inbook{Nguyen2026Marketing-06-virality,
author = {Nguyen, Mike},
title = {Marketing Research: A Technical and Academic Treatment},
chapter = {Virality and Word of Mouth},
year = {2026},
url = {https://mikenguyen13.github.io/marketing_research/06-virality.html}
}
Use this book with your AI assistant
Ask questions about this book in Claude, Cursor, VS Code, or ChatGPT, with answers that cite the section.
# Virality and Word of Mouth {#sec-virality}Almost everything a consumer learns about a product arrives through someone else.A friend's offhand recommendation, a one-star review, a forwarded video, aretweeted complaint—these *consumer-to-consumer* transmissions move demand morereliably than most paid messages, and they do so at a cost the firm does not pay.This chapter is about how such transmission works: what gets passed along, whypeople pass it, how far and how fast it travels, and how much of the resultingspread a manager can actually engineer rather than merely observe.Two questions organize the field, and they are easy to conflate. The first is*psychological*: given that a person is talking, what makes them choose one pieceof content over another to share? This is the literature on **word of mouth**(WOM) and its drivers—social currency, emotion, usefulness, narrative. The secondis *structural*: given that content is shared, what shape does the resultingdiffusion take through the social network, and is that shape predictable? This isthe literature on **cascades** and **structural virality**. A serious treatmentneeds both. Content drivers explain the *probability* that any individual relays amessage; network structure determines whether those individual relays compoundinto an epidemic or fizzle. We develop the content account first, formalize thediffusion account second, and close on the measurement and identification problemsthat make virality genuinely hard to study—chief among them the fact that whatlooks "viral" is often neither viral nor caused by the content at all.By the end the reader should be able to define virality precisely, distinguish itfrom popularity, write down a model of cascade size and state what limits itspredictability, and reason about seeding and measurement with the rightidentification caveats in hand.## Defining the ConstructWord of mouth predates the internet by the entire history of language, and itscanonical definition is correspondingly old. We adopt @westbrook1987'sformulation as the field's landmark statement.> Word of mouth is "informal communications directed at other consumers about the> ownership, usage, or characteristics of particular goods and services or their> sellers" [@westbrook1987].Three features of this definition do real work. The communication is *informal*(it is not the firm's paid voice), it is *consumer-to-consumer* (the firm is atbest an indirect party), and it is *about market objects* (products, services,sellers). The migration of this activity online produces **electronic word ofmouth** (e-WOM, sometimes "word of mouse"): the same construct, now mediated byplatforms that record, timestamp, and graph every utterance. Both classical WOMand e-WOM exploit the network effect to reach audiences the firm could not addressdirectly [@vilpponen2006], but e-WOM does so with lower marginal cost, fasterdiffusion, and sharper targeting [@bampo2008; @ferozkhan2014].*Virality* is the construct that e-WOM makes salient, and the literature uses theword in two incompatible ways that must be kept apart. One usage is **processual**:virality is the act or trajectory of successive sharing—"a process whereby an ad issuccessively shared by viewers." The other is **dispositional**: virality is a*property* of content, namely its latent potential to be spread. We adopt thedispositional reading throughout, because it is the one that supports predictionand design: we want to know what makes a given piece of content *likely* todiffuse, holding the network fixed.::: {.callout-note}## WOM, virality, and sharing as near-synonymsWe treat WOM, virality, and sharing interchangeably, following the strongconstruct overlap documented by @camarero2011 and @phelps2004. The distinctionsthat matter for this chapter are not lexical but mechanical: the *content drivers*of a single relay versus the *network dynamics* of aggregate diffusion. We areexplicit about which we mean at each point.:::Formally, fix a population of potential relayers indexed by $i$ and a piece ofcontent $c$. Let $S_{ic} \in \{0,1\}$ indicate that $i$ relays $c$. Thedispositional virality of $c$ is the share probability marginalized over thepopulation,$$v(c) \;=\; \mathbb{E}_i\!\left[\,\Pr(S_{ic} = 1 \mid \mathbf{z}_c, \mathbf{w}_i)\,\right],$$ {#eq-virality-def}where $\mathbf{z}_c$ are content/context features (the levers a marketer can pull)and $\mathbf{w}_i$ are person- and tie-level moderators. The content literature inthis chapter is, in effect, a catalog of the partial derivatives$\partial v / \partial \mathbf{z}_c$: which features of content raise theper-person relay probability. The structural literature (@sec-structural-virality)asks what happens when these individual relays are composed through the edges of asocial graph.Why this matters commercially is blunt: sharing is necessary but not sufficientfor value. Intentions overstate behavior, and behavior overstates profit. In onetelecommunications referral program, 81% of customers said they would referothers, only 33% actually did, and of those referred only 8% became profitablecustomers [@kumar_2007]. The funnel from stated willingness to share, to actualsharing, to monetized outcome leaks at every stage—a recurring caution when readingself-reported sharing studies.## Theoretical Foundations {#sec-thv-foundations}The two literatures this chapter joins, content drivers and network structure, eachrest on established theory, and naming the governing accounts shows why the empiricalregularities are not coincidental. The structural side descends from **social contagionand diffusion of innovations**: new products and ideas spread through a population byperson-to-person communication over time, the process that the formal model of@Bass_1969 partitions into innovators (external influence) and imitators (internal,word-of-mouth influence). Diffusion theory supplies the contagion hypothesis that thecascade literature (@sec-structural-virality) then tests and qualifies, and it is thereason an S-curve is so often, and so often wrongly, read as evidence of transmission.Network structure enters through the **strength of weak ties** [@granovetter1973]:novel information flows disproportionately across bridging ties that connect otherwiseseparate clusters, which is why tie structure governs whether individual relays compoundinto reach and why the weak-tie thesis organizes the seeding and tie-strength resultsbelow.On the content side, three social-psychological theories account for *why* a personrelays. **Social influence and social proof** hold that descriptive and injunctive normsshift behavior, often more than people anticipate (Cialdini, *Influence*, a monographcited here by name; for the empirical norm result see @goldstein2008norms): people share,adopt, and review in part because others do, which is why consensus cues and visiblepopularity move transmission. The **emotion/arousal account of sharing** holds that theoperative driver is not the valence of content but its physiological *arousal*:high-arousal states, in both positive and negative valence, potentiate sharing, whilelow-arousal states suppress it [@berger2012a], a claim formalized in @eq-arousal-model.And **social identity** holds that consumption and communication are identity-expressive:people relay content that signals who they are and that affirms valued group memberships,the mechanism underneath the impression-management and social-bonding functions developednext (Tajfel and Turner's social identity theory, an edited-volume statement cited byname). Together these theories predict both the *probability* that any individual relayscontent and the *shape* of the diffusion that results.The empirical literature that operationalized these ideas developed in three waves,and it is worth knowing which question each was answering. The first established thatonline word of mouth *predicts* commercial outcomes at all: @liu2006wordmouth showedthat pre-release and opening-week message volume forecasts box office, with volumerather than valence carrying most of the signal — a finding that reframed WOM from asoft construct into a measurable leading indicator. The second asked how transmissionactually proceeds, decomposing viral marketing into the distinct stages ofclick-through, opening, and forwarding, each with its own drivers, so that a campaignoptimized for one stage can fail at another [@debruyn2008multistage]; the parallelquestion of what determines transmission *within* an online community rather thanacross a dyad is taken up in @lee2007wordmouth. The third turned to what firms can doabout it. @shi2005firmmanagemen models the firm's management of social interactions asa decision variable rather than an environmental parameter; @fay2011socialmediadistinguishes online from offline conversation and finds the two are lesssubstitutable than practice assumes; and @janakiraman2016socialsale tracesfirm-generated social-media content through to purchase, showing the path runs viabrand consideration rather than directly to sale. @hughes2013understansocialgeneralizes across seller, retailer, and consumer sources, which is the comparisonmost single-source studies cannot make.## Why People Share: A Functional AccountThe single most useful organizing idea in the WOM literature is that sharing is*goal-directed*. People do not relay content at random; they relay it because doingso serves an end, often below the level of conscious awareness. @Berger_2014consolidates this into five functions that WOM serves, and—crucially—each functionmakes predictions about *what kind* of content will be shared. The framework isgenerative rather than descriptive: name the goal, and you can forecast the contentthat satisfies it. @fig-wom-functions maps each function to the content propertiesit predicts will be shared.```{mermaid}%%| label: fig-wom-functions%%| fig-cap: "Five functions of word of mouth and the content properties each one predicts will be shared. Multiple functions can operate simultaneously on the same act of sharing."flowchart LR WOM["Word of mouth<br/>(goal-directed)"] WOM --> IM["Impression<br/>management"] WOM --> ER["Emotion<br/>regulation"] WOM --> IA["Information<br/>acquisition"] WOM --> SB["Social<br/>bonding"] WOM --> P["Persuasion"] IM --> IMc["entertaining, useful,<br/>status-, identity-relevant"] ER --> ERc["emotional, high-arousal,<br/>valence-managed"] IA --> IAc["risky / complex /<br/>uncertain decisions"] SB --> SBc["common ground,<br/>shared emotion"] P --> Pc["polarized valence,<br/>arousing"]```That these are *functions* and not mere correlates is supported by theself-referential bias in everyday communication: over 70% of ordinary speechconcerns the self, and the same proportion holds for social-media posts. Peopletalk to accomplish something *for themselves*. We take the five functions in turn.### Impression managementConsumers share to curate how others see them. The mechanism rests on the premisethat consumption is identity-expressive—people acquire and discuss goods partly tosignal who they are—and it operates through self-enhancement (looking good),identity-signaling (advertising specific traits or expertise), and the mundane needto fill conversational space. The self-enhancement motive is strong enough thatpeople will distort to satisfy it: roughly 60% of retold personal stories arealtered in the retelling [@marsh2004], and people share content specificallybecause it burnishes their image [@wojnicki2017].The motive predicts a recognizable content profile. People preferentially sharematerial that is *entertaining* (novel, surprising, funny, or extreme), becauserelaying it makes the sharer look interesting and in-the-know. Novelty must bepaired with comprehensibility to function—novel-and-understandable is *interesting*,novel-and-incomprehensible is merely *confusing* [@Berlyne_1960]. The starkestevidence on novelty comes from @vosoughi2018, who tracked roughly 126,000 rumorcascades on Twitter, classified by six independent fact-checking organizations:false stories spread significantly farther, faster, deeper, and more broadly thantrue ones, and the gap is explained by the greater *novelty* of false content andthe surprise and disgust it evokes. The effect survives controls for bot activity,implicating human sharers rather than automation.Spread is only half of the misinformation problem; the other half is what the firstfew exposures do to the person who sees them. @puthineedi2026ballrolling ran threecontrolled experiments in which participants scrolled Instagram-style posts abouttopics unfamiliar to them. Impressions stabilized after a short, bounded run ofconsistent posts—roughly five, by the authors' account of the results[@informs2026fiveposts]—a "sufficiency" point they call the *Point of CriticalInformation*, after which further posts added little. The stance formed there thenfiltered what came next: congruent posts were more readily believed and shared, andcontradictory ones were more easily dismissed. Early judgments tracked factualaccuracy only weakly, since participants shown inaccurate information reacted muchlike those shown accurate information, so a correction that arrives after the stancehas formed works against a frame already in place. Set beside the novelty account,this is an argument about *sequence*: what matters is not only how far a false storytravels but whether it reaches a user first and consistently, which is why theauthors favor upstream interventions (getting authoritative content into the earliestexposures) over later fact-checks. Only the published abstract and the INFORMS pressrelease were accessible when this section was written; the five-post figure and theexperimental design come from the release.People also share what is *useful* (it makes the sharer helpful and smart) and whatis *status-relevant* (premium brands and knowledge are discussed more, because bothsignal standing). @deangelis2012braggartsgossips separate the two waysself-enhancement operates: a sharer can inflate their own experience (bragging) ordeflate someone else's (gossip), and which route is taken depends on whether thesharer's self-view has just been threatened or affirmed. A subtle reversal appearsfor *uniqueness*: unique products draw more WOM in general, yet consumers with a highneed for uniqueness *withhold* positive WOM to keep adoption—and thereforedistinctiveness—scarce [@cheema2010needuniquenes]. Finally,sharing is shaped by *accessibility*: products that are more frequently cued in theenvironment are discussed more, so heavily advertised products generate more WOM[@onishi2012]. Accessibility is also a route through which *incidental* statesmatter—unrelated physiological arousal raises sharing across the board[@berger2011a], a point we formalize under emotion below.### Emotion regulationPeople share to manage how they feel—seeking social support, venting,making sense of an experience, reducing post-decision dissonance, taking revenge ona firm, or reliving a positive episode. Sharing a negative experience can raisewell-being through perceived social support [@buechel2017]. This function predictsthat *emotional* content is shared more, but with two refinements that the nextsubsection makes precise: not all emotions promote sharing (shame and guiltsuppress it, because relaying them is self-incriminating), and the impression-management function can *oppose* emotion regulation when venting a negative episodewould reflect badly on the sharer.### Information acquisition, social bonding, and persuasionThe remaining functions are quicker to state. WOM **acquires information**: ratherthan learn by trial and error, consumers solicit advice, especially for *risky,complex, or uncertain* decisions where talking reduces perceived risk. WOM builds**social bonds**: people share *common ground* and *emotional* content because bothincrease felt connection and group cohesion. And WOM **persuades**: when the goalis to move another's choice, people share *polarized* (extremely positive ornegative) and *arousing* content, because moderation does not convince. Thesefunctions are not mutually exclusive—a single share can manage an impression,regulate emotion, and bond simultaneously—which is why the five-function lenspredicts content profiles rather than partitioning shares into types.## Drivers of ViralityThe functional account tells us *why* people share; the empirical literatureisolates *which content features* raise share probability. We group the robustdrivers into social currency, emotion, usefulness, and narrative, and then treatthe moderators—audience and channel—that determine when each driver dominates.### Social currencyThe dominant driver of posting behavior is *image-related* utility rather thanintrinsic enjoyment of the content: @toubia2013, in a field experiment thatrandomized whether contributions were attributed to their authors, find thatimage-related utility drives posting more than intrinsic utility. People share toaccrue social currency—to look interesting, smart, or in-the-know. A consequenceworth flagging is that "useful" and "interesting" are *not* interchangeablecurrencies: useful information signals helpfulness, but the signal is valued lessthan the signal of being entertaining [@Berger_2014].Social currency interacts with the *channel*, and the interaction is identifying.Interesting content is shared more in *written* (asynchronous) settings but not in*oral* (synchronous) ones. The mechanism is time to self-edit: asynchronouscommunication lets people polish what they say, which amplifies self-enhancementmotives, whereas synchronous conversation must be filled in real time, pushingpeople toward mundane, accessible topics merely to avoid silence. This is a cleanexample of a moderator (channel) switching a driver (interestingness) on and off.### Emotion: valence and arousalEmotion is the most-studied driver, and the literature's central correction is that**valence is not the operative variable—arousal is.** It is tempting to predictsharing from whether content is positive or negative; the better predictor is howphysiologically *activating* it is.@berger2012a make the case with both archival and experimental evidence. Examiningroughly 7,000 *New York Times* articles, they find that positive content is moreviral on average, but emotionality matters over and above valence, and the resultholds after controlling for usefulness, interestingness, surprise, prominence ofplacement, author fame, length, and timing. The experiments then dissociate valencefrom arousal: high-arousal *negative* emotions (anger, anxiety) raise sharing,while a low-arousal negative emotion (sadness) lowers it; high-arousal *positive*emotions (awe, excitement, amusement) raise sharing relative to a low-arousalpositive state (contentment). The clinching test is that *incidental* arousal—arousal induced by a task unrelated to the content, such as physical exertion—alsoraises sharing [@berger2011a], which it could not do if valence were the truecause.The mechanism is physiological: arousal is an excitatory state that potentiatesaction of all kinds, including sharing and other prosocial behavior[@gaertner1977]. Different emotions sit at different points on the activationdimension [@smith1985a], and the relevant contrast is **activation** (high arousal,behavioral readiness) versus **deactivation** (low arousal, relaxation)[@feldmanbarrett1998; @theneur1997]. @tbl-arousal-valence organizes thepredictions; note that the regulation goal differs by quadrant—high-arousalnegative states drive *venting*, high-arousal positive states drive *rehearsal*[@rimé1991].|| Low arousal (deactivation) | High arousal (activation) ||---|---|---|| **Positive valence** | Contentment — *less* shared | Awe, excitement, amusement — *more* shared (rehearsal) || **Negative valence** | Sadness — *less* shared | Anger, anxiety — *more* shared (venting) |: Sharing predicted by the arousal × valence quadrant. Arousal, not valence, is the operative dimension: high-arousal states in *both* valences raise sharing, while low-arousal states suppress it [@berger2011a; @berger2012a]. {#tbl-arousal-valence}A formal statement makes the claim testable. Let $a(c) \in [0,1]$ be the arousal acontent piece evokes and $\text{val}(c) \in [-1,1]$ its valence. The arousal modelasserts$$\frac{\partial \, v(c)}{\partial \, a(c)} > 0\quad\text{and}\quad\frac{\partial \, v(c)}{\partial \, |\text{val}(c)|} \;\text{small once } a(c)\text{ is held fixed},$$ {#eq-arousal-model}i.e., once arousal is conditioned on, residual valence carries little additionalexplanatory power for sharing. Identification of @eq-arousal-model is exactly whatthe incidental-arousal manipulation buys: by moving $a$ without moving any propertyof $c$, the experiment severs the confound between arousing content and*interesting* or *useful* content, which would otherwise make $a$ endogenous.The managerial corollary is **valuable virality** [@Akpinar_2017]. Emotional adsare shared more but often fail to lift brand outcomes because the brand isincidental to the story; informative ads lift brand evaluation (the brand isintegral to the message) but are shared less. The resolution is *emotional,brand-integral* content—embed the brand into an arousing narrative so that sharingand brand lift move together. The same prescription recurs in the advertisingliterature (@sec-advertising): @tellis2019 find across field studies thatinformation-focused content is shared less (except in risky contexts), thatpositive high-arousal emotions and "drama" elements (surprise, plot, characters,babies, animals, celebrities) raise sharing through arousal, that brand prominencematters, and that platform moderates fit—emotional ads travel on general platforms,informational ads on professional ones—with an optimal ad length of roughly 1.2 to1.7 minutes. Emotion can be operationalized directly from language[@yin2017], which makes these predictions deployable at scale.### Usefulness and narrativeBeyond emotion, *practical utility* is a robust driver: useful newspaper articlesare more viral [@berger2012a], and the broader pattern—useful, interesting, andemotionally arousing content is shared more—replicates in @milkman2014, who add thestriking finding that articles by women are shared more because they are written tobe more comprehensible, useful, and interesting. Three theoretical lenses explain*why* usefulness travels: sharing useful content signals knowledge and so servesimpression management [@wojnicki2017]; it generates reciprocity, the sharerexpecting a return favor [@fehr1998]; and it functions as social exchange withexchange value [@homans1958]. The channel again moderates: usefulness and noveltypredict *email* forwarding most strongly, whereas emotional evocativeness andcontent familiarity predict *social-media* retransmission [@kim2015], because thetwo channels serve different goals (private helpfulness versus publicself-presentation).*Narrative* is the vehicle that carries emotion and brand together. In socialmedia, brands and consumers co-create brand performances, and the ongoing *process*of improvisation matters more than any single output; managing a brand becomes thework of keeping the performance alive [@singh2012]. Narrative is also where thevaluable-virality prescription is executed—the brand rides inside the story ratherthan interrupting it.### Storytelling design: narrativity, emotional sequence, and reading level {#sec-virality-storytelling}"Narrative" is usually treated as a yes-or-no property of content. @ghose2026newsstoryunbundle it into design choices a publisher controls, and test them acrossfour studies in *Information Systems Research*. The three features are*narrativity* (a linear, chronological story versus a summary-first structure thatstates the outcome up front), *emotional sequence* (good-to-bad versusbad-to-good), and *reading level* (simple versus complex language). The paper askswhat each does to engagement, why (through perceived interestingness and thegratifications readers get), and when and for whom (news format, reading level,and consumer motivation as boundary conditions).The design is the instructive part. Study 1 is observational: it programmaticallycodes more than 5,000 satirical and 35,000 traditional news stories that onepublisher posted on Twitter, which holds the source constant while format varies.Studies 2 to 4 are controlled experiments in which large language models rewritethe same story to vary one feature at a time, so the manipulation is clean andcheap to scale (the SSRN version [@ghose2025newsstoryssrn] carries the fullresearch abstract). The press coverage reports more than 500 participants and an Indianpublisher as the source [@temple2026newsstory; @poynter2026newsstory].The results are conditional rather than universal. For *satirical* news, asummary-first structure with a good-to-bad sequence works best when the languageis simple, while a linear structure with a bad-to-good sequence works best whenthe language is complex. For *traditional* news, simple language helps and complexlanguage hurts, but complex language paired with linear narrativity and abad-to-good sequence recovers much of the loss. Motivation moderates the reading-leveleffect: high self-brand connection offsets the cognitive cost of hard text.Satire must also clear a higher interestingness threshold than traditional newsbefore readers share it. The authors' summary is that engagement depends on how thefeatures combine, and for whom, not on any single tactic [@ghose2026newsstory;@poynter2026newsstory].Three lessons carry beyond news. First, they refine the "emotion drives sharing"finding earlier in this chapter: the *order* of valence matters, and its signdepends on format and language, so a model with a single valence score discardssignal. Second, they sit beside the linguistic-style literature in@sec-virality-linguistic as another manipulable, text-extractable lever distinctfrom content. Third, they show a method: using an LLM to generate matchedvariants gives a within-story manipulation that observational text cannot, and theauthors report the same patterns for video, podcasts, and interactive media.(Only the abstracts, working-paper record, and press coverage were accessiblewhen this section was written; the effect sizes are in the paper.)### The source's stance as a signalWhat a story says is only part of what readers take from it; they also drawinferences from who published it. @huang2026mediaslant study reposting by Chinesemicroblog users, combining field data with an online experiment that varies whethera story's source and stance are disclosed. As expected, users repost stories thatmatch their own attitudes more often. They are also more likely to repost storiesthat run against the reporting outlet's usual stance. The authors' explanation isstrategic inference. An outlet trades popularity off against its stance, so a storythat cuts against the outlet's line was presumably worth running for its popularity,and readers infer greater popularity from the inconsistency. Two results supportthat reading. The effect is larger among users with stronger Bayesian-updating andlevel-k reasoning, the abilities the inference requires. And simply reminding usersthat media are slanted makes their responses more strategic. For sharing research,the implication is that the same content is a different signal depending on whocarries it, so a model of share probability built only on content features leavesout the inference readers draw from the source. (Only the published abstract wasaccessible when this section was written; the effect sizes are in the paper.)### Perceptual features: visual motion and children's viewing {#sec-virality-visual-motion}Every driver so far works through what content *means* to a viewer who might pass it on:the status it confers, the emotion it evokes, the use it offers, the story it tells, andthe source that carries it. The YouTube evidence fits that mold. @tellis2019 trace thesharing of online video ads to information, emotion, and brand prominence, and@ferozkhan2014 find that view counts of the most-viewed YouTube videos depend on networklinks and the uploader's offline fame as well as on the platform itself. @tucker2015 addsa warning that views and persuasion can come apart: across 400 video ads, relativepersuasiveness was on average 10% lower for every million views. The exceptions were adsthat drew many comments or comments naming the product, which seem to have earned theirviews through humor rather than outrage. Even the regulatorywork on YouTube engagement, such as the effect of FTC affiliate-disclosure rules onproduct review videos [@zhang2026ftc], is about what viewers infer from content.@kumar2026attention isolate a channel that runs beneath meaning. Their audience ischildren, who, the authors argue, rely less on the social and participatory cues thatcarry adult video. For this audience the relay mechanism behind @eq-virality-def islargely absent, and views accumulate through how well a video holds attention. Thefeature they study is *visual motion*, measured with optical flow, the computer-visionestimate of how fast brightness patterns move between frames (@sec-video-motion).@fig-virality-motion-studies lays out the three-study design. In more than 32,000 videosfrom leading children's YouTube channels, higher motion predicts viewership aftercontrolling for visual, audio, and affective characteristics. When the authors buildeducational videos that differ only in background motion and run them in randomized adcampaigns through YouTube's A/B testing infrastructure, the high-motion versions earnsignificantly higher view-rates and views. In a lab eye-tracking study with children,motion raises screen engagement mainly by cutting short long off-screen lapses, rather thanby preventing attention from breaking at all. It draws more gaze to the background withoutreducing gaze on the instructional content. The authors read this as re-engagement at thelevel of the screen, not displacement of the lesson.```{mermaid}%%| label: fig-virality-motion-studies%%| fig-cap: "The three-study design of Kumar, Viswanathan, and Bapna (2026). An observational association across children's YouTube videos is replicated causally in a randomized field experiment and then traced to a mechanism with lab eye-tracking."flowchart LR S1["Study 1: observational<br/>32,000+ children's YouTube videos"] --> F1["Higher optical flow,<br/>more views (with controls)"] F1 --> S2["Study 2: field experiment<br/>background motion manipulated,<br/>randomized YouTube ad campaigns"] S2 --> F2["Higher view-rates<br/>and views"] F2 --> S3["Study 3: lab eye-tracking<br/>with children"] S3 --> F3["Shorter long off-screen lapses;<br/>more gaze on background,<br/>no loss on instructional content"]```For the framework of this chapter, the paper adds a separate class of content features$\mathbf{z}_c$. Perceptual properties such as motion, pace, and visual intensity act onattention directly, apart from the informational and emotional content that the@tellis2019 decomposition captures, so a model of video engagement that codes only meaningleaves out a lever creators actually pull. It also marks a boundary on the outcome. Thedispositional virality of @eq-virality-def is a relay probability, while children'sviewership is closer to *retention*, and the features that drive the two need notcoincide. The design is a model of triangulation: an observational association, arandomized field replication that rules out the recommender as the source of thecorrelation, and a lab measure of mechanism. Finally, the governance stakes are unusuallydirect. A feature that raises views by shortening children's attentional lapses iscommercially valuable to creators and platforms, and at the same time it is the kind ofdesign choice that policy toward child-directed media and attention-capturing design isstarting to scrutinize. (Only the published abstract and an earlier working-paper abstractwere accessible when this section was written; the effect sizes are in the paper.)## Structural Virality {#sec-structural-virality}Everything so far concerns the per-person relay probability $v(c)$. We now turn towhat happens when those relays compose through a network, and here the field'shard-won lesson is that **viral is not the same as popular.**### Cascades versus broadcastsA piece of content can reach a million people two very different ways. It can be*broadcast*—a single hub (a celebrity, a front page) transmits directly to a massaudience, producing a wide but shallow diffusion tree. Or it can *cascade*—personrelays to person relays to person, producing a deep, multi-generation tree. Bothyield large reach; only the second is "viral" in the mechanistic sense. The totalnumber of retweets is therefore a poor proxy for virality, because it cannotdistinguish a hub blasting to its followers from a genuine person-to-personepidemic [@Goel_2015].The empirical reality is that genuine cascades are vanishingly rare. In a corpus ofroughly a billion diffusion events, about 90% of content was adopted by *no one*beyond the seed, and fewer than 1% of events achieved even a handful ofsecond-generation adoptions [@Goel_2012]. Virality, in the strict sense, is anextreme-tail phenomenon; most content dies at the source.@Goel_2015 formalize the broadcast–cascade distinction with a single statistic,**structural virality**, defined as the average distance between all pairs of nodesin a diffusion tree (the Wiener index, normalized by tree size). For a tree$T$ on $n$ nodes with shortest-path distances $d_{ij}$,$$\nu(T) \;=\; \frac{1}{n(n-1)} \sum_{i=1}^{n} \sum_{j=1}^{n} d_{ij}.$$ {#eq-structural-virality}A pure broadcast (a star: one seed, $n-1$ direct children) has small $\nu$—everypair is at most two hops apart—no matter how large $n$ grows. A deep chain ofperson-to-person relays has large $\nu$. The statistic thus separates *size* from*shape*: two events with identical reach can have wildly different structuralvirality. The following code makes the contrast concrete by computing@eq-structural-virality for a broadcast star and for a deep cascade of equal size.```{r structural-virality, message=FALSE, warning=FALSE}set.seed(606)# Mean pairwise distance (Wiener index / [n(n-1)]) on a tree given as a# parent vector: parent[i] is the index of node i's parent (NA for the root).structural_virality <-function(parent) { n <-length(parent)# Build n x n shortest-path distances via repeated parent-walks (tree => unique path). depth <-integer(n) path <-vector("list", n)for (i inseq_len(n)) { p <- i; chain <- iwhile (!is.na(parent[p])) { p <- parent[p]; chain <-c(chain, p) } path[[i]] <- chain depth[i] <-length(chain) -1L } D <-matrix(0, n, n)for (i inseq_len(n)) for (j inseq_len(n)) { common <-intersect(path[[i]], path[[j]]) # nearest common ancestor set lca_depth <-max(depth[common]) D[i, j] <- (depth[i] - lca_depth) + (depth[j] - lca_depth) }sum(D) / (n * (n -1))}n <-40# Broadcast: one root, everyone else attaches directly to it (a star).broadcast <-c(NA, rep(1L, n -1))# Cascade: a single deep chain, 1 -> 2 -> ... -> n.cascade <-c(NA, 1:(n -1))c(broadcast =structural_virality(broadcast),cascade =structural_virality(cascade))```The broadcast tree returns a structural-virality near 2 (almost all paths routethrough the hub), while the chain returns a value that grows with depth—mechanicallyidentical reach, opposite diffusion shape. Real cascades fall between these poles,and @Goel_2015 document that large structural virality does occur in the wild but israre and only weakly predictable from content alone.### The limits of prediction: skill, luck, and an $R^2$ ceilingCan we predict which content goes viral? @Goel_2015 give the question a sharp,discouraging answer by decomposing realized cascade success $S$ into a partexplained by observable "skill" $Q$ (content quality, seed characteristics) and anirreducible "luck" component $E$,$$S \;=\; f(Q) + E.$$ {#eq-skill-luck}The natural measure of predictability is how much variance in $S$ survives afterconditioning on everything knowable, $Q$:$$F \;=\; \frac{\mathbb{E}\!\left[\,\operatorname{Var}(S \mid Q)\,\right]}{\operatorname{Var}(S)} \;=\; 1 - R^2 .$$ {#eq-predictability}If $R^2 \to 1$, success is pure skill and perfectly forecastable; if $R^2 \to 0$,it is pure luck and unforecastable. The substantive result is that there is a*theoretical ceiling* on $R^2$ that is well below one, arising from the intrinsicstochasticity of the branching process: the more sensitive the cascade is to theinitial seed, the larger the variance in outcomes for fixed $Q$, and the lower theattainable $R^2$. Virality is partly designable and partly irreducibly chancy, andno amount of feature engineering closes the gap. This is the structural counterpartto the funnel leakage of @kumar_2007: even perfect content does not guarantee acascade.### What the S-curve does and does not tell youDiffusion is conventionally pictured as an S-shaped cumulative-adoption curve, andit is tempting to read an S-curve as *evidence of contagion*—of person-to-persontransmission. This inference is invalid, for two distinct reasons that anyidentification strategy must confront.First, **an S-curve does not imply contagion.** The same sigmoid can be generatedby mechanisms that involve no social transmission at all. *Population heterogeneity*in adoption thresholds produces an S-curve mechanically as the marginal adopter'sthreshold is crossed over time [@Chatterjee_1990], and time-varying *marketingeffort* can trace out the identical shape [@Van_den_Bulte_2001]. Attributing anS-curve to word of mouth therefore requires ruling out heterogeneity and marketingas alternative explanations—a problem of separating *contagion* from *correlatedunobservables*, addressed structurally in @sec-marketing-finance and in thediffusion-modeling literature.Second, **selection (survivorship) bias** contaminates retrospective viralitystudies: we observe the cascades that happened to grow and rarely thenear-identical content that died, so any pattern estimated on "viral" cases isconditioned on success. Combined with the rarity result of @Goel_2012—90% of eventsgo nowhere—this means naive comparisons of viral and non-viral content overstatewhatever distinguishes the survivors. @fig-scurve-confounds collects the competingdata-generating processes that yield the same aggregate curve.```{mermaid}%%| label: fig-scurve-confounds%%| fig-cap: "Why an observed S-shaped adoption curve is not evidence of contagion. At least three distinct data-generating processes produce the same aggregate shape, and survivorship in the sample compounds the inference problem."flowchart TD S["Observed S-shaped<br/>adoption curve"] S --> C["Social contagion<br/>(person-to-person)"] S --> H["Population heterogeneity<br/>in thresholds"] S --> M["Time-varying<br/>marketing effort"] C -.->|"observationally<br/>equivalent"| H H -.-> M SB["Survivorship bias:<br/>dead cascades unobserved"] --> S```## Network Position, Ties, and SeedingIf structure governs whether relays compound, then *where* in the network a messagestarts and *which* ties it crosses should matter. This section covers theposition/seeding literature and the surprisingly subtle role of tie strength.### Seeding strategyThe seeding question is whether a firm should inject content at well-connected hubsor at ordinary nodes. Analytical models and simulations had suggested that targetinghubs does not pay—hubs are no more *influential* over any given peer than ordinarynodes. @hinz2011 overturn this in small-scale field experiments and real campaigns:seeding to well-connected people yields up to *eight times* more adoptions, notbecause hubs are more persuasive per contact but because they are more *active* andexploit their greater reach. The lesson is that seeding returns come from reach andactivity, not from per-tie influence—a distinction that matters because it cannot berecovered from peer-level persuasion data alone. Seeding can also have *competitive*effects: a well-placed campaign raises focal-brand WOM while suppressingconversation about rival products in the same category [@chae2017].### Network management and identificationStudying how a firm's own networking activity drives online success raises an acuteendogeneity problem: firms that succeed may simply network more, so position andsuccess are jointly determined. @ansari2018 model the co-evolution of an actor'sego-network and success with stochastic actor-based models, characterizing eachactor by ego-network density and degree, betweenness, and closeness centrality. Toidentify the effect of networking on success they combine time fixed effects with acontrol-function approach instrumented by the firm's activity on an *independentUS* social platform—activity correlated with the firm's European networking effortbut plausibly excludable from European-market success. The design is a template fornetwork studies generally: position is endogenous, and credible identificationneeds an instrument that shifts position without directly shifting the outcome.### Tie strength and the strength of weak tiesGranovetter's classic thesis is that *weak* ties—acquaintances rather than closefriends—are disproportionately valuable for information diffusion, because theybridge otherwise disconnected clusters and so carry *novel* information. The moderne-WOM literature both confirms and qualifies this. (The positional version of theargument, in which the advantage belongs to the node that bridges rather than to thetie, is the structural-holes tradition; @sec-graph-brokerage covers its measures,including the pair-based opportunity and exclusivity scores of @borgatti2026beyond,which separate how many disconnected pairs a node spans from how many rivals span thesame pairs.)On the confirmation side, @peng2018 show that sharing propensity rises with*network overlap* between sender and receiver—common followees, common followers,and common mutual followers—using a dyadic hazard model on Twitter and Digg data.High overlap signals shared interests, so the receiver infers the content suits herown audience and relays it; popular content, conversely, is shared *less*, becausemany others already have. Tie strength also conditions *persuasion*: @lee2020 findthat *consensus language*—expressions implying widespread agreement—is morepersuasive coming from *weak* ties than strong ones, because a weak tie evokes abroader, more diverse agreeing group and thus signals higher validity, providingsocial proof.The sharpest recent test qualifies the weak-tie thesis itself. @rajkumar2022causalexploit large-scale randomized experiments on LinkedIn's "People You May Know"algorithm—rare *causal* variation in tie formation—to estimate how tie strengthaffects job mobility. They find that weak ties do increase job transmission but with*diminishing returns*: the relationship is an inverted U, not monotone. Moderatelyweak ties, not the weakest possible, maximize mobility, and the effect isindustry-dependent—weak ties dominate in digital industries, strong ties in lessdigital ones. The headline revision is that "the strength of weak ties" isnonlinear and context-contingent, not a universal law.@tbl-tie-strength contrasts the three studies, which agree that tie structurematters but disagree on the direction once the outcome and mechanism change.| Study | Setting / identification | Tie variable | Effect on outcome ||---|---|---|---|| @peng2018 | Twitter, Digg; dyadic hazard | Network overlap | Overlap raises sharing; popularity lowers it || @lee2020 | Experiments | Tie strength × consensus language | Weak ties *more* persuasive with consensus cues || @rajkumar2022causal | LinkedIn field experiments (causal) | Tie strength | Inverted-U; moderately weak ties maximize job mobility |: How tie strength enters WOM and diffusion across three studies. The apparent disagreements reflect different outcomes (sharing, persuasion, mobility) and identification strategies (observational dyadic models versus randomized tie formation). {#tbl-tie-strength}## Accessibility and the Mere-Exposure ChannelA driver that operates orthogonally to emotion is *accessibility*: content andbrands that are easier to retrieve from memory are talked about more, howevermundane they are. The theoretical basis is spreading activation in semantic memory:priming activates related constructs, and a primed brand becomes more likely toenter the consideration set [@Berger_2008]. @Berger_2008 show that more frequentexposure to perceptually or conceptually related cues raises product accessibilityand, through it, evaluation and choice—conceptual priming with real consumerconsequences. Accessibility is why publicly visible and heavily advertised productssustain ongoing WOM [@onishi2012], and why *familiar* content is selected more oftenfor retransmission [@kim2015]. It also supplies an alternative, non-emotional reasoncontent travels, which any well-specified sharing model must include as a controllest the emotion coefficients absorb a pure-accessibility effect.## Moderators: Audience and ChannelWhich WOM function dominates—and therefore which content travels—depends on twomoderators introduced above and worth consolidating. The **audience** ischaracterized by tie strength, audience size, and the status of the tie; the**channel** by whether communication is written or oral, whether the communicatoris *identifiable*, and whether the audience is *salient* during composition. Therecurring result is that asynchronous, identifiable, audience-salient channelsamplify self-presentational motives and so favor interesting, status-relevantcontent, whereas synchronous channels favor accessible, mundane content that fillsconversational time. Demographics moderate too: the 18–24 cohort responds *less* toviral content because saturation raises their activation threshold, making theiremotions harder to move [@Libert_2014]. And conversation has its own dynamics—a widegap between two opposing opinions stimulates future discussion, whereas fluctuationaround a single opinion does not, with informational content strengthening theopposition-to-discussion link [@Nagle_2014].## Cultural and Dispositional AntecedentsSharing is not culturally invariant. Two constructs from the cross-cultural WOMliterature illustrate how disposition shapes transmission.*Locus of control*—the generalized expectancy about whether one's outcomes areinternally or externally caused [@lam2005]—predicts *whom* people talk to.Internals (who believe they control their lives) tend to be more educated[@lachman1989], higher-income [@hoffman2003], and more senior and male[@smith1997], and they engage in WOM with *out-groups*; externals, oriented towardaffiliation and avoidance, engage more with *in-groups* [@lam2005].*Horizontal/vertical individualism–collectivism* shapes *what* communicative rolepeople adopt. @singelis1995 cross individualism–collectivism with ahorizontal (equality-emphasizing) versus vertical (hierarchy-emphasizing)dimension, yielding four cultural patterns. Applied to WOM, @choi2019 find thatvertical orientations—where standing is signaled by display—channel into bothopinion *leadership* and opinion *seeking*, while horizontal orientations feedopinion leadership through a norm of being socially appropriate. The broaderregularity is that collectivist cultures do more opinion *seeking* andindividualist cultures more information *giving* [@fong2008].## Linguistic Style {#sec-virality-linguistic}Because e-WOM is text, *how* something is said is measurable and consequential, anda sizable literature maps linguistic features onto WOM outcomes. The features thathave been shown to matter include the abstraction level of language[@schellekens2010], the explanatory framing a storyteller uses—which feeds back toshape the storyteller's own attitudes [@moore2012]—figurative language[@kronrod2010], markers of modesty versus boastfulness [@packard2016],"dispreferred" markers that soften disagreement [@hamilton2014], the personalpronouns a communicator selects [@packard2013], and linguistic mimicry betweenparties [@moore2017]. The practical upshot is that the *style* of a message is amanipulable lever distinct from its content, and—because style is extractable fromtext at scale—an empirically tractable one.## Negative Virality and FirestormsVirality is not always the firm's friend. Negative e-WOM can compound into a*firestorm*—a sudden, large-scale wave of hostile messages. @herhausen2019 supplya framework for the drivers of negative e-WOM and the firm responses that mitigateit, identifying the level of arousal, structural tie strength, and linguistic-stylematch as the factors that fuel a firestorm. Their central prescription followsdirectly from the arousal account of @eq-arousal-model: the firm's response should*match the intensity of arousal* in the negative e-WOM—neither under- norover-reacting—because mismatched responses either dismiss aggrieved high-arousalcustomers or inflame low-arousal ones. Negative virality thus obeys the samearousal mechanics as positive virality, with the valence of the firm's exposurereversed.## Contractuality: Engineering Authentic WOMFirms that try to *manufacture* WOM through incentives confront a credibilitytrade-off captured by **contractuality**, defined by @Lisjak_2021 as "the extent towhich a perk is perceived to be conditional on specific behaviors and contingenciesdictated by a company." Consumers prefer perks with *less salient* contractuality,and a prior negative relationship with the firm leads them to read a high-contractuality perk as a manipulative quid pro quo with ulterior motives.Low-contractuality perks foster *authentic* WOM but trade off against effectivenessand efficiency—the firm gives up the control that an explicitshare-for-reward contract would provide. The construct formalizes the intuition thatthe most valuable WOM is the kind the firm can least directly buy.Contractuality asks whether the *sender* was paid. A second credibility test askswhether the sender is doing what they say, and it becomes observable on platformsthat publish both. @liu2026socialtrading exploit a social trading platform where auser's posted opinions and actual trades are visible side by side, and study earlycryptocurrency adoption—a setting chosen because it is exactly the high-uncertainty,high-return regime where social learning should matter most.Peers' adoption raises an individual's probability of adopting, which is thebaseline social-learning result. The informative part is the interaction: the effectis markedly stronger when a peer's *trading behavior aligns with their expressedopinion*. Talk that is corroborated by observable action carries the signal; talkthat is not, does not. This is the same economics as contractuality approached fromthe other side—there, credibility is threatened by an incentive the audience infers;here, it is established by a costly action the audience can verify.Two boundary conditions sharpen it. When uncertainty or returns run high, investorslean on positive opinions *regardless* of whether the poster's trades back themup—the alignment check is a luxury that gets abandoned under exactly the conditionsthat make it most valuable. And the responsiveness is heterogeneous in a directionworth noting: older, more experienced, higher-income, more risk-tolerant usersrespond most to credible aligned signals, while younger users are more skeptical ofpositive sentiment, plausibly because they have more experience with noisy socialmedia. For a firm seeking authentic WOM about a new technology, the designimplication is architectural rather than contractual—make advocates' behaviorvisible alongside their advocacy, and the credibility follows without a contract todisclose.## Empirical Regularities and Open ProblemsA handful of well-replicated facts and one persistent confound deserve listing asthe chapter's empirical bottom line. WOM dispersion predicts higher future ratings,but its impact declines over the product life cycle, so it should be measured early[@godes2009]—WOM is both a precursor and a consequence of behavior, whichcomplicates causal reading. Contagion is detectable even in consumer packaged goods[@Du_2011], a category once thought too low-involvement for social transmission.Authorship matters: articles by business academics, psychologists, and economistsare shared more than those by physicists and biochemists, and—net of content—menrate scientific summaries as more comprehensible, interesting, and useful and sharethem more [@milkman2014]. And identifying *influential* users in a network is itselfa modeling problem with no off-the-shelf solution [@trusov2010].The persistent confound is **structural virality** itself. Many sharingexperiments—including the influential arousal studies—ask respondents how likelythey would be to share, an intention that (i) carries little social risk in the lab,inflating the estimate, and (ii) abstracts away the network through which realdiffusion runs. Two pieces of content with identical per-person share probability$v(c)$ can diffuse to radically different sizes depending on where they seed and howthe branching process resolves (@eq-skill-luck, @eq-predictability). The frontierof the field is precisely the integration we have kept separate for expositoryclarity: joining the *content* model of $v(c)$ to the *structural* model of how$v(c)$ propagates, so that designable content features and irreducible networkstochasticity are estimated in one frame rather than two.## Key Takeaways- **Word of mouth is goal-directed.** The five functions—impression management, emotion regulation, information acquisition, social bonding, persuasion—predict *what* content travels, not merely *that* sharing happens [@Berger_2014].- **Arousal, not valence, drives sharing.** High-activation states in both valences raise transmission; incidental-arousal manipulations identify the effect by moving arousal without moving content [@berger2011a; @berger2012a].- **Viral $\neq$ popular.** Broadcasts and cascades both reach many people; only cascades are structurally viral, and genuine cascades are extreme-tail rare[@Goel_2012; @Goel_2015].- **Predictability is capped.** Cascade success decomposes into skill and luck, and the irreducible luck component imposes a theoretical $R^2$ ceiling (@eq-predictability) [@Goel_2015].- **An S-curve is not contagion.** Heterogeneity and marketing effort generate the same shape; survivorship bias compounds the inference problem[@Chatterjee_1990; @Van_den_Bulte_2001].- **Seeding pays through reach, not per-tie influence**, and tie-strength effects are nonlinear and context-dependent [@hinz2011; @rajkumar2022causal].
31.4.1 Social currency
The dominant driver of posting behavior is image-related utility rather than intrinsic enjoyment of the content: Toubia and Stephen (2013), in a field experiment that randomized whether contributions were attributed to their authors, find that image-related utility drives posting more than intrinsic utility. People share to accrue social currency—to look interesting, smart, or in-the-know. A consequence worth flagging is that “useful” and “interesting” are not interchangeable currencies: useful information signals helpfulness, but the signal is valued less than the signal of being entertaining (Berger 2014).
Social currency interacts with the channel, and the interaction is identifying. Interesting content is shared more in written (asynchronous) settings but not in oral (synchronous) ones. The mechanism is time to self-edit: asynchronous communication lets people polish what they say, which amplifies self-enhancement motives, whereas synchronous conversation must be filled in real time, pushing people toward mundane, accessible topics merely to avoid silence. This is a clean example of a moderator (channel) switching a driver (interestingness) on and off.