Non-face emojis in digital marketing: Effects, contingencies, and strategic recommendations

Abstract

Non-face (NF) emojis are increasingly used to complement or substitute words in digital marketing messages, yet the effects, mechanisms, and contingencies of this communication strategy remain underexplored. In a large-scale longitudinal study of Airbnb listings, we show that NF emojis (vs. simple text) lead to an increase in eWOM volume, an effect we replicate experimentally. This effect is qualified by important boundary conditions whose underlying mechanisms are investigated in two additional experimental studies. At the message level, using multiple substitutive (vs. complementary) NF emojis reduces message evaluations and eWOM volume due to reduced processing fluency. At the source level, seller quality further moderates the interaction between emoji function and emoji number: for premium sellers, using multiple NF emojis reduces message evaluations and eWOM volume irrespective of their function due to reduced perceptions of competence. We distill these findings into detailed managerial guidelines for using NF emojis in digital marketing.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11747-022-00917-z.

Introduction

“Competition is fierce. How do you make your listing stand out? I thought that would be emojis” -Harris-Robin, Airbnb community forum.

Over 200 million people in the US buy, rent, and sell online every year (Coppola, 2021), yet digital communication comes with its challenges. When listing an apartment on Airbnb or selling a couch on Facebook marketplace, sellers battle for customer attention against thousands of competitors: New York City alone hosts over 35,000 unique Airbnb properties at any given time (Inside Airbnb, 2021). Despite this intensely competitive environment, many platforms impose a character limit (e.g., 50 for Airbnb, 140 for Etsy, 100 for Facebook Marketplace) that makes it even harder for sellers to differentiate themselves.

As the opening quote illustrates, one popular strategy to attract customer attention is to use emojis, whose usage has reached its all-time high in 2021 (Buchholz, 2021). Emojis are a form of visual textual paralanguage that “supplements or replaces written language” through symbols (Luangrath et al., 2017, p. 98). At a superordinate level, emojis can be broadly categorized as either “face emojis” (e.g., , ) that represent emotions or “non-face emojis” (e.g., , , ) that depict objects, activities, or concepts (Bai et al., 2019; Li et al., 2019). Non-face emojis (NF here on) do not convey emotions but can be used to either complement the meaning of a word (e.g., “beer tonight?”)—the “complementary” function—or replace a word to save characters (e.g., “ tonight?”)—the “substitutive” function.

While non-face emojis make up for 90% of all existing emojis (Unicode, 2020) and are used at least as frequently as their face counterparts (Brandwatch, 2021), the effects and contingencies of NF emojis are little understood. Previous marketing research has focused mainly on face emojis (Das et al., 2019; Li et al., 2019; Smith & Rose, 2020), leaving important questions unanswered. Do messages embedding NF emojis increase message evaluations and electronic word-of-mouth (eWOM) compared to simple text? Should NF emojis be used to substitute for words? How many NF emojis are too many, and does this effect depend on their function (i.e., complementary vs. substitutive)? These questions become particularly pressing when considering that customers’ decisions are also influenced by source-level factors such as seller quality (Wilson & Sherrell, 1993). The use of NF emojis may lead to different outcomes for regular and premium sellers depending on how customers expect them to communicate (Kang & Rubin, 2009). Premium sellers who are expected to communicate with competence may be evaluated more negatively when using an informal paralanguage such as NF emojis, yet no prior research investigates this possibility.

Against this background, the current research seeks to fill the theoretical vacuum of marketing research on NF emojis. Results from a large-scale longitudinal panel of 195,733 unique Airbnb properties across 26 months (2.18 M observations of monthly listings in total) and three controlled experiments generate three main insights. First, including NF emoji in digital marketing messages increases eWOM volume (Study 1) and message evaluations (Study 2). Second, the effect of NF emoji function (complementary vs. substitutive) is moderated by their number (Studies 1, 3, and 4), such that using a high number of substitutive (vs. complementary) NF emojis hurts message evaluations and eWOM volume due to reduced processing fluency (Studies 3 and 4). Third, the effect of using a high number of NF emojis is further moderated by seller quality (regular host vs. superhost in the context of Airbnb), such that using a high number of NF emojis is not beneficial to premium sellers regardless of function. This effect occurs as premium sellers are expected to display competence in the way they communicate and using multiple NF emojis reduces this perception (Study 4).

In providing these insights, our research makes key contributions to literature on emojis in marketing (Li et al., 2019; McShane et al., 2021; Smith and Rose, 2020), multimodal communication (Hull & Nelson, 2005; Tavassoli, 1998; Tavassoli & Han, 2001; Tavassoli & Lee, 2003), and digital marketing strategy (Garnefeld et al., 2021). We are the first to isolate the effects of NF emoji in marketing communication and examine linguistic features unique to NF emoji usage, namely their complementary vs. substitutive functions. In doing so, we challenge the dominant insight that the more emojis are used in a digital message, the more effective it would be (McShane et al., 2021). Instead, we find that using NF emojis to substitute words and save characters is a strategy to be used cautiously, as multiple substitutions hinder processing fluency and hurt message effectiveness. We further contribute to theories of linguistic stereotyping (Kang & Rubin, 2009; Lambert et al., 1960) by showing that seller quality moderates the interplay between emoji function and emoji number.

Beyond its theoretical relevance, our research provides actionable insights for both sellers and digital marketplace providers. For sellers, our data surfaces an insufficient understanding of the benefits of using NF emojis depending on seller quality, with premium sellers using more NF emojis than they should. We address this issue by providing a detailed set of practical recommendations on NF emoji use for both regular and premium sellers. For digital marketplace providers, we note the existence of different policies on whether and how to use emojis due to limited understanding of their effect on customer evaluations and downstream business outcomes. For instance, Facebook Marketplace and Etsy allow for emojis in their listing titles, whereas Airbnb officially does not (Airbnb Content Policy, 2021), but does not strictly enforce this ban. Our findings implicate policy changes that allow sellers (and particularly regular sellers) to use NF emojis in their listings.

Theoretical development

Face vs. non-face emojis

Emojis are a form of visual paralanguage that supplements or replaces written language through symbols representing facial expressions, objects, and activities (Luangrath et al., 2017). While symbols such as paralinguistic marks (e.g., !!!) place emphasis on emotions and concepts, emojis differ in that they represent emotions and concepts (Luangrath et al., 2017). In their inception, emojis were created to compensate for the lack of non-verbal cues in computer-mediated communication (Bai et al., 2019). First-generation emojis thus depict various facial expressions that convey corresponding emotions, either through human faces (e.g., ) or anthropomorphized animal faces (e.g., ). We refer to these symbols as face emojis. As an ever-evolving visual paralanguage, however, emojis now include more than 3,460 unique symbols (Buchholz, 2021) that are used not only to express emotions but also to depict objects (e.g., a beer ) and activities (e.g., running ) associated with specific meanings. We refer to these symbols as non-face (NF) emojis.

Previous marketing research has predominantly focused on the effects of face emojis on message evaluations and customer behavior (Table 1). For example, Das et al. (2019) found that ads containing smiley emojis increased positive affect and purchase intention. Consistent with the Emotion as Social Information Model (van Kleef, 2009), Li et al. (2019) found that when service employees used human face emojis in a chat box, customers inferred higher warmth and lower competence. This in turn triggered an inference-making process that led to differential levels of service satisfaction depending on whether communal or exchange relationship norms were salient. Smith and Rose (2020) provided further evidence that unconscious emotional contagion underlies the effect of face emojis displaying positive affect. In summary, previous research consistently shows that face emojis serve an affective function by conveying the emotions associated to specific facial expressions.

Table 1.

Marketing studies on emoji usage

Study	Type of Emoji	Context	Key Measures Investigated	Methodology	Key Findings
Das et al. (2019)	Face emojis	Advertising (retail)	Affect, Purchase intentions	Experimental	Face emojis increase positive affect and purchase intention, moderated by product type (hedonic vs utilitarian)
Li et al. (2019)	Face emojis	Service	Service satisfaction, Purchase behavior, WOM intentions	Mixed methods	Face emojis increase warmth but lower competence, moderated by relationship norm (communal vs. exchange)
Smith and Rose (2020)	Face emojis	Service	Affect, Relationship strength	Experimental	Face emojis increase affect via emotional contagion, moderated by relationship norm (communal vs exchange)
McShane et al. (2021)	No distinction	Advertising (social media)	Intentions to engage, Twitter likes and retweets	Mixed methods	Emojis presence increases brand engagement mediated by perceived playfulness, moderated by nature of interplay between emojis and text
Ma and Wang (2021)	Face emojis	Service failure	Customer satisfaction, Repurchasing intentions	Experimental	Emojis displaying negative emotions increase satisfaction and repurchasing intentions via perceived sincerity, moderated by relationship norm (communal vs. exchange)
Urumutta Hewage et al. (2021)	Face emojis	Advertising (social media)	Emoji evaluation, Social media engagement	Experimental	Asymmetric (vs. symmetric) face emojis receive more positive evaluations via human expression resemblance and emotional expression strength
Wu et al. (2022)	Face emojis	Online reviews	Perceived review helpfulness	Experimental	Face emojis depicting unambiguous emotions increase review helpfulness via increased processing fluency, moderated by user expertise
Our Study	Non-Face emojis	Advertising (digital markets)	Ad evaluations, eWOM volume	Mixed methods	NF emojis increase eWOM volume relative to text. The effect of emoji function (complementary vs. substitutive) is moderated by emoji number and seller quality and seller quality

The main effect of non-face emojis relative to text

Perhaps due to their more recent history, NF emojis have received limited research attention in marketing. Human–computer interaction scholars found that when a message is ambiguous (e.g., got a ticket), NF emojis can help disambiguate the message (e.g., got a ticket vs. got a ticket ) and facilitate message comprehension (Riordan, 2017a, b), yet extant scholarship offers limited insight as to whether NF emojis influence business-relevant outcomes. Therefore, our first aim is to investigate whether NF emojis lead to more positive message evaluations and increased eWOM—the latter defined as any “consumer-generated, consumption-related communication that employs digital tools and is directed primarily to other consumers [e.g., online reviews] (Babić Rosario et al., 2020, 425).” We chose to investigate eWOM as a consequential variable because it represents a reliable predictor of future sales (Babić Rosario et al., 2016).

Since NF emojis do not depict facial expression but rather symbolize objects and activities, they are unlikely to persuade customers through emotional contagion. However, their pictorial nature should still be more attention-grabbing than text alone. Research has shown that pictures attract more attention than text independent of their size (Pieters & Wedel, 2004). Similar to pictures, emojis were found to increase users’ attention to digital messages based on both self-report measures (Willoughby & Liu, 2018) and eye-tracking data (Beyersmann et al., 2022). In turn, greater attention to an ad is associated with more positive message evaluations (Maughan et al., 2007) and higher purchase intentions (Goodrich, 2011). We thus expect customers to evaluate marketing messages embedding NF emojis more positively likely because they are more attention-grabbing than text alone. Formally stated:

• H1 Relative to text alone, messages embedding NF emojis increase (a) message evaluations and (b) eWOM volume.

However, when zooming into different categories of NF emojis, we predict differential effects based on their function and the number of NF emojis embedded in the message. We describe different emoji functions and articulate our predictions next.

Multimodal processing: The effect of emoji function is moderated by number of emojis

Drawing on previous literature acknowledging that NF emojis can serve different linguistic functions (Danesi, 2017; Li et al., 2019; Luangrath et al., 2017; McShane et al., 2021), we propose a further distinction in the use of NF emojis. When NF emojis are used to complement the meaning of a word (e.g., “beer tonight?”), we define them as “complementary” NF emojis. When they are used to substitute a word to save space (e.g., “ tonight?”), we define them as “substitutive” NF emojis. We predict that the different functions of NF emojis may have different effects on message evaluations and eWOM depending on how they influence processing fluency through multimodality.

Multimodality refers to the ensemble of two or more modes of meaning representation in reference to an object or concept (Kress, 2000). Multimodality is involved when a message requires at least two language modalities to be represented (Pauwels, 2012), such as written text containing emojis or written text accompanied by spoken words. Distinct modes of representation that seemingly encode identical content (e.g., beer or ) still require distinct cognitive channels and resources to be processed (Hull & Nelson, 2005; Tavassoli, 1998). For instance, people process words sequentially and phonologically, associating meaning to the resulting sound construction, but process symbols simultaneously and visually, retrieving meaning based on pre-existing semiotic associations with that symbol (Hull & Nelson, 2005; Tavassoli & Han, 2001). Consequently, multimodal (vs. unimodal) communication leads to greater brand memory and attitude because pieces of information in different modalities are processed in parallel through different cognitive channels, reducing mutual interference (Tavassoli, 1998; Tavassoli & Fitzsimons, 2006; Tavassoli & Han, 2001; Tavassoli & Lee, 2003).

In multimodal texts, emojis can accompany words to facilitate meaning-making. Using multiple complementary NF emojis that add to the meaning of an already complete sentence should not interfere with message processing, because information presented in different modalities is processed through different cognitive channels (Tavassoli, 1998). Complementary emojis are a paralanguage that provides non-verbal information by associating a context-specific visual cue to a word, drawing the readers’ attention to specific keywords (e.g., “Clean and sunny apartment in the heart of New York). Complementary NF emojis should thus facilitate the mental representation of the words they are associated with, serving as symbolic visual reinforcements without altering the text (i.e., customers can understand the message by processing only verbal information).

Emojis can also substitute words entirely. When NF emojis substitute text, both pieces of information are incomplete on their own but collectively form a complete sentence (e.g., “Clean and apartment in the of New York”). In this case, multimodal communication may actually lead to an inferior outcome because the required conversion across modalities is cognitively taxing. When NF emojis substitute a word or phrase, the sentence would be phonologically incomplete without the missing phonetic component. To provide an illustration, when reading the sentence “Clean and apartment in the of New York,” the reader needs to translate both the sunny and heart emojis and then understand the intended meaning of the sentence. Such situations force readers to convert the NF emojis to its phonetic counterparts (e.g., subvocalizing the NF emojis). Because substitutive emojis are processed first visually and then phonologically, they require multimodal decoding. The conversion of modality consumes cognitive resources and reduces processing fluency—“the subjective feeling of ease or difficulty associated with any type of mental processing” (Graf et al., 2018, p. 394)—as reading cannot proceed until the NF emojis are converted phonetically.

When the number of substitutive emojis is as low as one, we expect the negative effect of processing disfluency to be negligible. When multiple instances of substitution occur within a brief text, however, substitutive emojis become a pictorial language themselves: they substitute for the core information of multiple lexical words that otherwise should be in the same place in a sentence. Such extensive substitution is likely cognitively taxing, attenuating the integration of information (Tavassoli, 1998). This is compounded by the fact that, because NF emojis are a comparatively new digital language few people are proficient with, extensively switching between different processing channels should consume more cognitive resources and hinder processing fluency. We thus expect the effect of emoji function to be contingent on the number of emojis embedded in the message. Formally stated:

• H2 The number of NF emojis moderates the effect of emoji function, such that, when emoji number is high (vs. low), substitutive NF emojis lead to (a) lower message evaluation and (b) decreased eWOM volume, compared to complementary NF emojis.

• H3 The negative effect of a high number of substitutive (vs. complementary) NF emojis on (a) message evaluation and (b) eWOM volume is mediated by decreased processing fluency.

Seller quality moderates the effects of emoji function and number of emojis through perceived competence

Customers decoding marketing communications evaluate not only the message content, but also any attribute associated to the message sender (Gigerenzer & Gaissmaier, 2011; Wilson & Sherrell, 1993). Reputation badges signalling seller quality are one of such attributes (Hui et al., 2016). For instance, eBay automatically upgrades sellers to Top Rated Seller (eTRS) status if they completed 100 or more transactions and maintained a defect rate below 5% within the past 12 months (eBay, 2021). Similarly, Airbnb’s Superhost badge (i.e., ) signals outstanding service performance of the host over the past 12 months (Airbnb, 2021). We thus operationalize reputation badges as an indicator of seller quality.

We predict that seller quality further moderates the interactive effect between emoji function and emoji number. If premium sellers are expected to communicate professionally, perceptions of competence may be reduced by an excessive use of NF emojis irrespective of emoji function. Supporting this prediction, reputation badges are highly diagnostic of seller quality and allow to discriminate between regular and premium sellers, creating attributions of “prestige” and expectations to behave accordingly. Sociolinguistic researchers studying how social class and language intersect acknowledge that every community allows for different levels of language prestige linked to the way a group expresses themselves (Kang & Rubin, 2009). Typically, the superior language is that endorsed by community authorities. Because emoji is a relatively new and informal paralanguage whose use is also officially banned from platforms such as Airbnb, its use should be associated to a lower prestige speech community.

For regular sellers (i.e., those who do not display a top seller badge), we thus expect the interactive effect between emoji function and emoji number (Hypothesis 2) to hold, such that using multiple substitutive emojis (but not complementary emojis) should reduce processing fluency. However, for premium sellers, using multiple emojis should reduce message evaluations and eWOM through a reduction in perceived competence, regardless of whether the emojis used are complementary or substitutive. This argument is informed by the linguistic stereotyping hypothesis, which posits that using speech varieties commonly associated with low-prestige groups (e.g., a dialect) can lead to negative attributions about the speaker (Kang & Rubin, 2009; Lambert et al., 1960). Operationally, this form of linguistic stereotyping should be observed in reduced perceptions of competence about the message sender. Prior emoji research acknowledges that using smileys can lower perceptions of competence in industries in which professional communication is expected, such as banking (Li et al., 2019; Smith & Rose, 2020). Similarly, we suggest that premium sellers expected to match their communication style to their prestige will be perceived as less competent when using an excessive number of visual symbols associated to a lower prestige group. Formally stated:

• H4 Seller quality moderates the interactive effects between emoji function and emoji number on (a) message evaluations and (b) eWOM volume. For regular sellers, there will be a significant interaction effect between emoji function and emoji number. For premium sellers, the two-way interaction effect will be attenuated such that using multiple emojis should have a negative effect irrespective of emoji function.

• H5 The interactive effect between seller quality, emoji function, and emoji number on message evaluations will be mediated by perceived competence.

In line with Hulland and Houston (2021) recommendations to demonstrate behavioral outcomes, we investigate the hypothesized effects using a large secondary dataset of Airbnb listings, followed by three experiments (Fig. 1).

Fig. 1.

Overview of empirical studies and hypotheses

Study 1: Modelling airbnb listings data

Study 1 uses field data from 2.18 million observations of monthly listings of 195,733 unique Airbnb properties to provide ecologically valid evidence for the main effects of NF emojis on eWOM volume (Hypothesis 1) and the moderating effects of emoji number and seller quality (Hypotheses 2 and 4). As one of the world’s biggest sharing economy platforms (PwC, 2015), Airbnb.com is ideal to model the effects of including NF emojis in property listings on customer eWOM volume. Figure 2 shows the conceptual model for this study.

Fig. 2.

Conceptual model of Study 1

Method

Data

We source the data from insideairbnb.com, an independent research website collecting data on Airbnb listings around the globe. The data is publicly available under CC0 1.0 Universal Public Domain Dedication license (Huang, 2022). The website scrapes information from Airbnb.com on a city’s listings at a monthly frequency. We use data from four US cities – New York City, Los Angeles, Chicago, Washington DC – which represent the top four Airbnb markets in the US (Lock, 2020). Our sample covers Airbnb listings in these four markets between January 2018 and February 2020 (26 months) and ends before WHO declared COVID-19 a global pandemic in March 2020. We examine these four cities since they ensure sufficient diversity of listings that would capture major trends in emoji usage. At the same time, these cities lie in different geographic regions within the US, ensuring that our results are not driven by local or regional trends. To ensure homogeneity across listing types, we restrict to “entire home/apartment” and “private room” listings accounting for 97% observations in these cities. Our data sample covers a panel of 2.18 M observations of monthly listings covering 195,733 unique properties across 26 months.1

For this panel, we derive a corpus of listing titles, which consists of words, symbols, and emojis that promote the property at each time period. Listing titles are the first touchpoint with potential customers on Airbnb while they search for properties for their stay. They may describe objective attributes such as property size, nearby landmarks, or price.2 At the same time, hosts may include friendly messages or symbols (e.g., emojis) in the title to invite potential customers to explore the listing.3Titles are thus one of the most important marketing communications tools available to the hosts and serve as the unit of analysis for this study.

Apart from the title, the scraper also collects additional information available on each listing page including total number of reviews received, property characteristics (such as price, neighbourhood, number of bedrooms, amenities, quality scores (ratings) from past users, cancellation policy) and host characteristics (“superhost” status, number of other properties hosted by them, experience). We also observe the full textual description of the property and host on the listing page. This information helps us measure the dependent variable and provides important controls to isolate the effect of NF emoji use.

Operationalization of independent variables

First, we categorize each emoji based on their Unicode symbol (unicode.org). Following Li et al. (2019), we categorize both textual (e.g.,:), < 3) and pictorial (e.g., , ) symbols as emojis. 29,410 Airbnb listing titles employ at least one of the 96 different NF emojis or 12 types of face emojis we found in the sample. Conditional on using an emoji, the average title uses 1.92 emojis. Table 2 shows the descriptive statistics for these variables in our data sample.

Table 2.

Descriptive statistics for emoji function and frequency in Airbnb titles (Study 1)

No. of observations	2,189,224
No. of observations with (face and NF) emojis in titles	29,410
No. of unique emojis (including emoticons)	96
No. of unique face emojis	12
Total no. of face emojis in listing titles	5,182
No. of unique non-face (NF) emojis	84
Total no. of complementary NF emojis in listing titles	47,540
Total no. of substitutive NF emojis in listing titles	3,666

Next, we categorize emojis as face or non-face, based on whether the emoji denotes facial expressions or not. In terms of taxonomical classification, 91.1% of all emojis are NF emojis and 8.9% are face emojis. Table 3 shows the most popular face and NF emojis used by Airbnb hosts in our data sample. We operationalize both face and NF emoji presence and number in our data, including number of face emojis as a control variable. Figure 3 shows the empirical distribution of NF emoji usage in our data (conditional on using at least one in the title). Figure 4 further shows the limited use of face emojis in our dataset and the stark increase in popularity of NF emojis over time.

Table 3.

Most popular Face and NF emojis used in Airbnb listings (Study 1)

Fig. 3.

Frequency of NF emoji number per listing (Study 1)

Fig. 4.

Use of face vs. NF emojis over time (Study 1)

Emoji function

We differentiate NF emojis as “substitutive” or “complementary” depending on whether they replace the focal word in the listing title or not. We perform this emoji function classification by analyzing the 29,410 listing titles that incorporate any emoji. Two independent coders manually classified emoji function in listing titles. NF emojis were coded as “substitutive” if they replaced the focal word in a sentence or as “complementary” in all other cases (intercoder reliability = 91.4%). A third coder resolved any discrepancies. For instance, “Serene. In the of Beverly Hills” would be classified as a substitutive use of the “ ” emoji, while “ Cozy studio only 1/2 block from Wrigley” would be classified as complementary use(s) of the same emoji. Since a single listing title may contain both complementary and substitutive emojis, we define two variables—complementary emoji presence (0 = absence, 1 = presence) and substitutive emoji presence (0 = absence, 1 = presence)—to represent each emoji function separately.4

Emoji number

We operationalize “emoji number” as a dummy variable to signify when a “high” number of complementary or substitutive emojis is used in a listing title (denoted as 1) compared to a “low” number (denoted as 0). Given the empirical distribution in Fig. 3, we set the threshold for the high number condition as the use of two or more NF emojis. Table 4 shows descriptive statistics for overall emoji number and by each function separately. Analogous to emoji presence, we define two variables–complementary emoji number and substitutive emoji number—to separately represent high use levels of each emoji function.5

Table 4.

Descriptive statistics for dependent, independent, and control variables (Study 1)

Variable	Mean	Std. Dev
Dependent Variable:
No. of reviews	1.3056	2.5033
Independent Variables:
NF emoji presence (dummy 0 = absence, 1 = presence)	0.0114	0.1060
Complementary NF emoji presence	0.0100	0.0994
Substitutive NF emoji presence	0.0016	0.0403
Emoji number (dummy 0 = 1, 1 = 2 +)	0.0068	0.0820
Complementary emoji number	0.0067	0.0818
Substitutive emoji number	0.0001	0.0068
Superhost (Yes = 1)	0.2445	0.4298
Listing Title Characteristics
Listing title word length	5.8957	2.0205
# Face emojis	0.0023	0.0513
# Textual attention markers	0.3150	0.7372
Time-Variant Listing Features
Rating quality score	79.252	35.324
Price ($)	146.50	126.21
Security deposit ($)	116.98	176.54
Cleaning fee ($)	58.361	62.666
Instant-bookable (Yes = 1)	0.3885	0.4874
Description word length	271.89	203.11
Description topic category: 1	0.1160	0.3202
Description topic category: 2	0.3630	0.4809
Description topic category: 3	0.0376	0.1902
Description topic category: 4	0.0072	0.0847
Description topic category: 5	0.1250	0.3307
Description topic category: 6	0.1572	0.3640
Description topic category: 7	0.1939	0.3954
Accommodation capacity	3.2509	2.2083
Number of amenities	23.482	11.277
Cancellation policy: Moderate	0.2727	0.4453
Cancellation policy: Strict	0.4328	0.4955
Exogenous Listing Features
Listing age (months)	25.966	21.655
Time since last scrape (days)	32.456	18.547
Host Characteristics
# Host verifications	5.2415	2.0057
Is professional host (Yes = 1)	0.4155	0.4928
Host experience (years)	3.9797	2.1427

Seller quality

To operationalize our second boundary condition, we categorize message sources on Airbnb based on whether they are regular hosts or superhosts. We use the badge “superhost” to proxy for high service quality, in line with previous research on top seller badges in online marketplaces (Garnefeld et al., 2021; Gigerenzer & Gaissmaier, 2011). Superhosts comprise 25% of the sample.

Operationalization of dependent variable

Our main outcome of interest is the number of reviews received by the property listing between two successive web-scrapes. Listing reviews include consumption experiences from past clients and serve as a source of recommendation or information regarding the property for other customers.6 Online reviews are thus a widely recognized type of eWOM volume (Babić Rosario et al., 2020), which has a significant impact on future sales and further word-of-mouth (Godes & Mayzlin, 2004; Moe & Trusov, 2011; Xiong & Bharadwaj, 2014; for a meta-analysis, see Babić Rosario et al., 2016).7 Our main dependent variable thus measures incremental eWOM volume received by a property listing. To operationalize this variable, we assume listing, host characteristics, and other controls at time ‘t’ prevailed for the time between scrapes ‘t-1’ and ‘t’, resulting in new reviews at time ‘t’. We thus construct the number of new reviews as the difference between cumulative reviews received at scrapes “t” and “t-1.8 Figure 5 shows the average number of reviews per day across cities and time, indicating seasonality in travel behavior. The average listing receives 1.3 reviews between two successive web-scrapes.

Fig. 5.

Average number of reviews per day across cities and time (Study 1)

Other measures

Additional listing and host characteristics may simultaneously affect demand. We classify these control variables into four categories: (1) listing title characteristics, (2) time-variant listing features, (3) exogenous listing features, and (4) host characteristics. The first two categories, listing title and features, are decision variables controlled by the host, and can be easily varied month-to-month. Therefore, these decision variables could be potentially endogenous (see discussion for endogeneity under Model specification). The last two, host characteristics and exogenous listing features, are exogenous features that are beyond hosts’ purview. We provide brief descriptions for each of these below. Appendix A presents detailed description of independent, dependent, and control variables used in the analysis.

Listing title characteristics

Apart from NF emojis, listing titles may include other characteristics that attract customer attention. Listing title length measures the amount of information conveyed in the title. We also capture number of face emojis and textual symbols (e.g., !, *, ~ , |) that could be used in listing titles.

Time-variant listing features

We measure other characteristics that could affect listing demand and downstream eWOM volume. Property characteristics such as its price, rating quality, accommodation capacity and amenities directly correlate with the popularity of the listing. Additional features such as its cancellation policy and whether it is instantly bookable may increase ease of booking.

In addition to objective features, hosts also include detailed textual descriptions that can belong to a mixture of topics (e.g., a welcome note, property description, host description etc.). We capture volume of information with the word length of the description and control for the possibility that some of the topics may be correlated with both emoji usage and eWOM volume. For instance, listing titles such as “ Cozy studio only 1/2 block from Wrigley” could include welcome notes that increase warmth and are positively correlated with eWOM volume. To control for these confounders, we perform Latent Dirichlet Allocation (LDA) to identify and classify underlying topics that occur in listing text description (Berger et al., 2020). Methodologically, LDA classifies documents consisting of multiple topics in two steps, first from document to topic, and second from topic to word. We classify the textual description into 7 abstract topics (details in Web Appendix 1) and note the most probable topic assigned to each textual description. These measures adequately control for any listing description characteristics that co-occur with emoji usage.

Exogenous listing features

Property characteristics such as property type and neighbourhood (property location) are time-invariant property features. Similarly, listing age measures the time since the property first appeared on Airbnb and automatically updates with each scrape. Time since last scrape measures the time since the listing’s Airbnb page was last observed by the web scraper. Finally, external drivers like seasonality (e.g., holiday season, summer months) may also influence the rise or fall in bookings. These are exogenous controls since they cannot be changed by the host.

Host characteristics

Apart from seller quality in terms of regular host or superhost, we measure additional host characteristics including number of independent identity verifications, their experience of being an Airbnb host (in years), and whether they are a professional Airbnb host (i.e., host multiple properties on Airbnb). Identity verifications are mandated by Airbnb policy to maintain host integrity (and dissuade scams). The last two variables measure host experience. As these factors reduce risk and uncertainty, their potential influence on customer behavior warrants their inclusion.

Table 4 shows the descriptive statistics for the above variables.

Model specification

Our empirical model analyzes the relationship between NF emoji usage and number of reviews, while controlling for endogeneity and other concurrent factors. We choose the log IV GMM model specification due to superior log-likelihood and model fit. Alternate specifications are available as robustness checks in Web Appendix 2.

Formally, we model the log of number of reviews for listing $i$ and month t as follows:

$${\log \left(1,+,N,u,m,b,e,r,O,f,R,e,v,i,e,w,s\right)}_{\mathit{it}}={\mathrm{\beta }}_0+{\mathrm{\beta }}_{1}CompNFEmojiPresenc{e}_{\mathit{it}}+{\mathrm{\beta }}_{2}SubNFEmojiPresenc{e}_{\mathit{it}}+{\mathrm{\beta }}_{3}CompNFEmojiPresenc{e}_{\mathit{it}}\ast CompEmojiNumbe{r}_{\mathit{it}}+{\mathrm{\beta }}_{4}SubNFEmojiPresenc{e}_{\mathit{it}}\ast SubEmojiNumbe{r}_{\mathit{it}}+{\mathrm{\beta }}_{5}CompNFEmojiPresenc{e}_{\mathit{it}}\ast Superhos{t}_{\mathit{it}}+{\mathrm{\beta }}_{6}SubNFEmojiPresenc{e}_{\mathit{it}}\ast Superhos{t}_{\mathit{it}}+{\mathrm{\beta }}_{7}CompNFEmojiPresenc{e}_{\mathit{it}}\ast CompEmojiNumbe{r}_{\mathit{it}}\ast Superhos{t}_{\mathit{it}}+{\mathrm{\beta }}_{8}SubNFEmojiPresenc{e}_{\mathit{it}}\ast SubEmojiNumbe{r}_{\mathit{it}}\ast Superhos{t}_{\mathit{it}}+{\mathrm{\beta }}_{9}Superhos{t}_{\mathit{it}}+\mathrm{\gamma }{X}_{\mathit{it}}+\delta Z_{\mathit{it}}+{\mathrm{ϵ}}_{\mathit{it}}$$ 1

where $NFEmojiPresenc{e}_{\mathit{it}}$ is whether an NF emoji is present in the title for listing i in month t, as defined in the previous section and Table 4. We explore the role of function of NF emojis by analyzing “complementary” and “substitutive” NF emoji presence. Therefore, ${\mathrm{\beta }}_1$ (${\mathrm{\beta }}_2$) tests the average performance of complementary (substitutive) emojis in driving listing eWOM volume relative to text. Further, we test the effect of high number for complementary and substitutive emojis (Hypothesis 2) using the interaction of complementary (substitutive) emoji presence with ${\mathit{CompEmojiNumber}}_{\mathit{it}}$ $\left({\mathit{SubEmojiNumber}}_{\mathit{it}}\right)$. Here, we would interpret ${\mathrm{\beta }}_3$ (${\mathrm{\beta }}_4)$ as the average difference between high and low number of complementary (substitutive) NF emojis.9 Finally, we will also test a managerially relevant three-way interaction of complementary and substitutive NF emoji usage with emoji number and seller quality ($Superhos{t}_{\mathit{it}})$. This interaction will indicate if there is significant heterogeneity in the moderation effect of emoji number based on seller quality. Therefore, the set of parameters from ${\mathrm{\beta }}_5$ to ${\mathrm{\beta }}_8$ will be interpreted as differences between Airbnb superhosts and regular hosts (Hypothesis 4) for the corresponding variables. ${\mathrm{\beta }}_9$ captures the standalone effect of being a superhost on eWOM volume.

Apart from the IVs measuring the effect of NF emojis in listing titles on eWOM volume, Eq. (1) also allows for multiple controls. We denote $X_{\mathit{it}}$ as decision variables such as listing titles characteristics or time-variant listing features which could be potentially endogenous. $Z_{\mathit{it}}$ includes other exogenous controls for host characteristics, exogenous listing features and fixed effects described in Table 4. We shortly describe both below.

Controls and fixed effects

We include a rich set of exogenous controls $Z_{\mathit{it}}$ in our model to control for the confounding factors mentioned above. We use neighborhood fixed effects to control for systematic differences in listing demand (and consequently downstream eWOM volume) across neighborhoods within the same city (e.g., demand differs between Manhattan and Harlem listings within New York City). Similarly, demand would rise during holidays which may also vary by city and property type. For instance, demand for 2 + bedroom entire homes increases in Los Angeles during summer vacations, which would in turn increase reviews for these listings. Therefore, we also control for city- and property-type-specific time fixed effects. We thus analyze the effect of NF emoji usage on eWOM volume by comparing listings within the same city, neighborhood, property type, and month and controlling for listing amenities and host characteristics.

Endogeneity

In our analysis, endogeneity concerns may still arise if one or more of our focal independent variables (complementary/substitutive emoji presence, number and superhost status) or the regressors in $X_{\mathit{it}}$ are correlated with the error term in Eq. (1). One possible reason is due to demand shocks unobservable to the researcher, which influence listing features (Papies et al., 2017; Rutz & Watson, 2019; Wooldridge, 2010). Hosts may anticipate these shocks and respond to them by suitably adjusting any or all of listing features offered including usage and type of NF emojis, listing title and descriptions, prices, services and other amenities. This, in turn, increases booking rates and consequently number of reviews left on the platform (i.e., eWOM volume). Therefore, our set of potentially endogenous regressors include our independent variables as well as all time-variant decision variables within the hosts’ purview. We represent these variables as the endogenous vector $W_{\mathit{it}}$ in Eq. (2).

$$W_{\mathit{it}}=\{{\mathit{CompNFPresence}}_{\mathit{it}},{\mathit{SubNFPresence}}_{\mathit{it}},{\mathit{CompNFNumber}}_{\mathit{it}},{\mathit{SubNFNumber}}_{\mathit{it}},{\mathit{Superhost}}_{\mathit{it}},X_{\mathit{it}}\}$$ 2

To counter this, we instrument current month independent and decision variables $W_{\mathit{it}}$ with their previous month counterparts and host experience. These are valid instruments since last month’s prices, emoji usage, and other decision variables would be correlated to current month levels due to underlying cost structure (e.g., property type, neighborhood, etc.), listing quality, and host experience. At the same time, past period decisions would be uncorrelated with current period demand. For instance, listing prices in May are not predictive of demand spikes in June. Customer stockpiling is also not a threat in the context of travel accommodation. Given the validity of our instruments, we should recover causal estimates of NF emoji presence and number from the model analysis.

Estimation

Our final model includes 24 endogenous decision variables for listing title and listing features as well as the independent variables in Eq. (1). The first-stage equation and the GMM moment condition for endogenous variable set $W_{\mathit{it}}$ is shown in Eq. (3).

$$\begin{array}{c}{W}_{\mathit{it}}=\mathrm{\Phi }{W}_{it-1}+\pi Z_{\mathit{it}}+u_{\mathit{it}}\\ E\left([,W_{it-1},{Z_{\mathit{it}}]ϵ}_{\mathit{it}}\right)=0\end{array}$$ 3

This is estimated via two-step feasible Generalized Method of Moments (GMM) procedure proposed by Hansen (1982). The two-step estimator first estimates the system of equations defined by Eq. (1) and (3) using an identity weighting matrix.10 In the second step, we use an updated weighting matrix that yields efficient and consistent estimates of standard errors (Wooldridge, 2002). This is similar to the control function approach proposed by Petrin and Train (2010) and Danaher et al. (2015). Estimation is conducted on Stata with the ivreghdfe command (Baum et al., 2003; Correia, 2018).

Results

Table 5 summarizes the results of our model estimation (full model results are presented in Appendix B). The models present the evolution of the focal parameters ${\mathrm{\beta }}_1\mathrm{to}{\mathrm{\beta }}_9$ in Eq. (1) above. Across models, we correct for endogeneity of host decision and independent variables using GMM estimation with the instruments. The Durbin-Wu-Hausman distance test statistic of endogeneity (Baum et al., 2007) across models shows that the instruments significantly change the parameters, validating our estimation approach.

Table 5.

Main model results (Study 1)

Dependent Var: log(1 + Number of reviews)	Model 1	Model 2	Model 3	Model 4
Independent Variables1:
Complementary NF emoji presence (${\mathrm{\beta }}_1)$		0.215*** (0.026)	0.200*** (0.033)	0.234*** (0.038)
Substitutive NF emoji presence (${\mathrm{\beta }}_2)$		0.126* (0.051)	0.145** (0.052)	0.194** (0.075)
Complementary NF emoji presence*Complementary Number (${\mathrm{\beta }}_3)$			0.022 (0.043)	0.071 (0.047)
Substitutive NF emoji presence*Substitutive Number (${\mathrm{\beta }}_4)$			-0.432** (0.147)	-0.514*** (0.131)
Complementary NF emoji presence* Superhost (${\mathrm{\beta }}_5)$				-0.074 (0.065)
Substitutive NF emoji presence*Superhost (${\mathrm{\beta }}_6)$				-0.124 (0.098)
Complementary NF emoji presence*Complementary Number*Superhost (${\mathrm{\beta }}_7)$				-0.145* (0.073)
Substitutive NF emoji presence*Substitutive Number*Superhost (${\mathrm{\beta }}_8)$				0.193 (0.287)
Superhost (${\mathrm{\beta }}_9)$	0.369*** (0.001)	0.369*** (0.008)	0.369*** (0.008)	0.372*** (0.008)
Control Variables
Listing Title Characteristics1	✓	✓	✓	✓
Time-Variant Listing Features1	✓	✓	✓	✓
Exogenous Listing Features	✓	✓	✓	✓
Host Characteristics	✓	✓	✓	✓
Fixed-Effects:
City-Property Type-Month	✓	✓	✓	✓
Neighborhood	✓	✓	✓	✓
Observations	2,189,224	2,189,224	2,189,224	2,189,224
Durbin-Wu-Hausman test of endogeneity	65.3***	66.9***	68.4***	69.8***
Within-R2	0.248	0.249	0.249	0.249
BIC	3,967,425	3,964,611	3,964,570	3,964,173
Likelihood Ratio Test Statistic (compared to Model 1)		2,842***	2,913***	3,369***

Numbers in parentheses denote standard errors. Standard errors clustered by neighbourhood

Significance codes—***:0.001, **0.01, *0.05

1 – Second stage IV GMM results reported for endogenous variables

Model 1 is the base model including all control variables on listing title, host, and other listing characteristics. We also include fixed effects to control for city- and property type-specific seasonality (via city-property type-month fixed effects), and neighborhood-specific demand conditions (via neighborhood fixed effects). Standard errors are clustered by neighborhood to allow for correlations across time and listings within the same neighborhood. Note that, since the DV is in log-scale, we can interpret the percentage change in reviews per unit change in the independent variable as ${100\ast (e}^{\beta }-1)$. In model 1, the effect of seller quality (“superhost”) is highly significant (${\mathrm{\beta }}_9$ = 0.369, p < 0.001) indicating the important role played by seller quality. That is, listings with superhost badges receive 44.7% (95% CI = [42.4%, 46.9%]) higher eWOM volume than listings without. We describe additional control variables specific to the Airbnb context in Appendix A.

The effect of NF emoji presence

Model 2 disentangles the effects of complementary and substitutive emojis relative to text only. Both complementary (${\beta }_1$ = 0.215, p < 0.001) and substitutive (${\beta }_2$ = 0.126, p < 0.05) NF emoji presences have, on average, a positive effect on eWOM volume relative to text only. To interpret in managerially relevant terms, complementary emoji presence in listing titles increases eWOM volume by 24%, (95% CI = [17.7%, 30.4%]), while substitutive emoji presence increases it by 13.5%, (95% CI = [2.2%, 24.7%]) on average. We note that the difference between the effect of complementary and substitutive NF emoji presence is insignificant (mean diff. = 10.5%, n.s.). These results support Hypothesis 1.

The interactive effect of emoji function and number

Model 3 investigates the moderating role of NF emoji number. Similar to model 2, the effect of having at least one complementary (${\beta }_1$ = 0.200, p < 0.001) or one substitutive NF emoji (${\beta }_2$ = 0.145, p < 0.01) in the listing (i.e., emoji presence) is similar (mean diff. = 0.054, n.s.). Using two or more complementary NF emojis has no additional effect on eWOM volume (${\beta }_3$ = 0.020, n.s.) relative to one complementary emoji. On the other hand, using two or more substitutive NF emojis significantly reduces eWOM volume (${\beta }_4$ = − 0.432, p < 0.01) relative to one substitutive emoji. Therefore, the total effect of using multiple substitutive emojis relative to text is (${{\beta }_2+\beta }_4=-0.287,p<.05)$, which is negative and significant. Comparing complementary and substitutive NF emoji usage in the high number condition, the net difference between their effects is significant (${{\beta }_1+\beta }_3-{{\beta }_2-\beta }_4$ = 0.51, p < 0.001). This result is equivalent to a 49.8%, 95% CI = [28.1%, 71.5%] increase in eWOM volume when using high number of complementary emojis relative to a high number of substitutive emojis and supports Hypothesis 2. We provide a visual representation of Model 3 results in Fig. 6.

Fig. 6.

Percentage change in eWOM due to NF emoji usage by function and number (Study 1)

Three-way moderation by seller quality

Finally, Model 4 shows a managerially relevant source of heterogeneity in the above moderation effect. We interact NF emoji number (low vs high) and emoji function (complementary vs. substitutive) with the host’s “superhost” status. Comparing ${\beta }_1\mathrm{to}{\beta }_4$ between models 3 and 4, the results of model 4 are mostly driven by regular hosts who constitute 75.3% of the sample. Regular hosts strongly benefit from NF emoji presence regardless of their function being complementary (${\beta }_1$ = 0.234, p < 0.001) or substitutive (${\beta }_2$ = 0.194, p < 0.01). Embedding multiple complementary emojis does not provide additional benefits (${\beta }_3=0.071,n.s.)$ while embedding multiple substitutive emojis significantly reduces the effect (${\beta }_4$ = $-$ 0.514, p < 0.001). This, in essence, makes using multiple substitutive emojis less attractive relative to text (${{\beta }_2+\beta }_4=-0.320,p<.01)$ for regular hosts.

Looking at superhosts, coefficients ${\beta }_5$ to ${\beta }_8$ represent the incremental deviation between regular and premium quality hosts for the emoji function and number variables captured in ${\mathrm{\beta }}_1$ to ${\mathrm{\beta }}_4$ respectively. Complementary (${\beta }_5=-0.074,n.s.)$ and substitutive (${\beta }_6=-0.124,n.s.)$ NF emoji presences do not differ significantly from regular hosts, suggesting that, on average, superhosts and regular hosts experience similar benefits from single NF emoji presence. Comparing within function, there are no relative benefits of using a low number of complementary vs. substitutive NF emojis for superhosts (${{\beta }_1+\beta }_5-{{\beta }_2-\beta }_6$ = 0.089, n.s.).

When superhosts use a high number of complementary emojis, however, eWOM volume decreases significantly (${\mathrm{\beta }}_7=-0.145,p<0.05)$ relative to regular hosts. In the high number, substitutive NF emoji condition, superhosts experience a similar drop in eWOM volume as regular hosts (${\mathrm{\beta }}_8=0.193,n.s.)$. Therefore, using two or more NF emojis is not beneficial for superhosts irrespective of function. Comparing complementary and substitutive NF emoji usage in the high number condition for superhosts, the net difference between their effects is also no longer significant (${{\beta }_1+\beta }_3+{{\beta }_5+\beta }_7-{{\beta }_2-\beta }_4-{{\beta }_6-\beta }_8$ = 0.34, n.s.). These results support Hypothesis 4.

Robustness

We perform multiple robustness checks to show consistency of our results across different modelling assumptions and subsets. Since number of reviews is a count variable, we test alternate model specifications with Poisson and negative binomial specifications (Web Appendix 2). As additional robustness checks, we also allow for random effects with a hierarchical linear model (HLM) specification and fit an alternate quadratic specification after recoding emoji number as a continuous variable (Web Appendix 2). Last, we test the robustness of our Log IV GMM model results for different property-types and quality ratings (Web Appendix 3).

Discussion

Across more than 2.18 million observations of monthly listings, the results of Study 1 provide ecologically valid evidence for a generalized positive effect of both complementary and substitutive NF emojis on eWOM volume, relative to text only (Model 2).

Crucial to the hypotheses, further probing (Model 3) suggests that emoji number influences the effectiveness of substitutive emojis, such that two or more substitutive emojis nullify the effect of using NF emojis (relative to text). Finally, we find a significant three-way interaction with seller quality (Model 4), wherein regular hosts enjoy greater benefits than superhosts when they use two or more complementary NF emojis in their listing titles. These results are robust to functional form specifications and across different data subsets, providing converging evidence in support of Hypotheses 1, 2 and 4.

Study 2: The effect of NF emojis vs. text

Study 2 is a causal replication of the effect of NF emojis relative to text alone (Hypothesis 1).

Method

One-hundred-and-seventy-eight US residents (M_age = 32.28, SD = 11.38, 49.4% female) were recruited from Prolific in exchange for £0.40 and randomly assigned to single-factor (text only vs. complementary emoji vs. substitutive emoji) between-subjects design. Participants were instructed to evaluate a banner ad for the fictitious company LINK, a digital startup commercializing a stargazing app capable of connecting stars into constellations through augmented reality. In the text condition, the ad read “LINK: Discover the stars.” In the complementary emoji condition, a star emoji was included in addition to the focal word (LINK: Discover the stars ⭐). In the substitutive emoji condition, the focal word “stars” was replaced with the corresponding NF emoji (LINK: Discover the ⭐). Following random assignment, participants completed a 3-item, 7-point bipolar measure of message evaluations (“What is your evaluation of the ad you just saw? bad-good/unconvincing-convincing/unfavorable-favorable”; α = 0.93) adapted from Thompson and Malaviya (2013). Additional measures were included for exploratory purposes but are not discussed further, with the exception of attention (“This ad: attracts my attention/ is attention-grabbing/ is too interesting to miss; α = 0.95) which was included to validate the intuition that NF emojis are more attention-grabbing than text alone.

Results

Message evaluations

A one-way ANOVA revealed a significant difference in message evaluations across conditions (F(2, 175) = 3.60, p = 0.029). Planned contrasts showed that, compared to text only (M_text = 3.28, SD = 1.59, N = 60), having either complementary (M = 4.06, SD = 1.73, N = 59, t (175) = 2.54, p = 0.012, d = 0.47) or substitutive NF emojis (M = 3.90, SD = 1.74, N = 59, t (175) = 2.02, p = 0.044, d = 0.37) significantly increased message evaluations. The difference between complementary and substitutive emojis was not significant (p = 0.611).

Mediation analysis

We used Hayes’s (2013) PROCESS bootstrapping protocols (Model 4, 5,000 iterations, multicategorical: X1 = complementary, X2 = substitutive, reference group = text) to confirm that messages containing NF emojis are evaluated more positively than messages with text alone because they are more attention-grabbing. Results confirmed the significance of the indirect effects passing through attention for both X1 (β = 0.16, BootSE = 0.08, 95% CI = [0.01 to 0.32]) and X2 (β = 0.22, BootSE = 0.09, 95% CI = [0.07 to 0.40]). All indirect effects are standardized and standard errors are bootstrapped.

Discussion

The results of Study 2 provide causal evidence for a positive main effect of one NF emoji (either complementary or substitutive) on message evaluations, compared to text only. Hence, we replicate the positive main effects of complementary and substitutive emojis vs. text in Study 1 (${\mathrm{\beta }}_1$ and ${\mathrm{\beta }}_2$ in Model 2). Having established the superior effectiveness of NF emojis relative to text, we now delve deeper into what boundary conditions make complementary NF emojis more or less effective relative to substitutive NF emojis.

Study 3: The interactive effect of emoji function and emoji number

Study 3 aims at providing causal evidence for the interactive effects between emoji function and emoji number on message evaluations, mediated by processing fluency. By forcing multimodal processing of images and text, substitutive NF emojis run the risk of reducing processing fluency, a risk necessarily greater the more words are substituted. We thus expect that messages embedding multiple substitutive (vs. complementary) NF emojis will be evaluated more negatively (Hypothesis 2) because they are less fluent to process (Hypothesis 3).

Method

Study 3 follows a 2 (emoji function: complementary vs. substitutive) × 2 (emoji number: low vs. high) between-subjects design. Three-hundred US residents were recruited from Amazon Mechanical Turk and compensated $0.30 (M_age = 40.54, SD = 12.95, 48% female). Participants were asked to evaluate an Airbnb listing (original wording: Sunny apartment close to the beach). In the complementary condition, emojis were added to highlight key attributes, using one emoji for the low number condition (Sunny apartment close to the beach ) and two for the high number condition ( Sunny apartment close to the beach ). In the substitutive condition, emojis were added to replace the text highlighting key attributes, again using one emoji for the low number condition (Sunny apartment close to the ) and two for the high number condition ( apartment close to the ). Next, participants indicated on a 7-point bipolar scale whether the number of emojis used in the ad was low or high. This served as the manipulation check for emoji number. Participants then completed a 3-item, 7-point bipolar scale for message evaluations (“What is your evaluation of the ad you just saw? bad-good/unconvincing-convincing/unfavorable- favorable”; α = 0.94) adapted from Thompson and Malaviya (2013) and measures of processing fluency (α = 0.92) from Graf et al. (2018). All multi-item measures were averaged to form a single index for each measure.

Results

Manipulation check

The low NF number conditions were perceived as having fewer emojis than the high number conditions (M_low = 2.83, SD = 1.67, N = 149 vs. M_high = 4.32, SD = 1.52, N = 151; t(298) = 8.06, p < 0.001, d = 0.93), and the effect was not moderated by emoji function (F(1, 296) = 0.80, p = 0.371).

Message evaluations

A two-way ANOVA revealed a main effect of emoji function (F(1, 296) = 6.03, p = 0.015), no main effect of emoji number (F(1, 296) = 1.09, p = 0.297), and a significant interaction between emoji function and emoji number (F(1, 296) = 9.19, p = 0.003). When using one emoji, there was no significant difference on message evaluations based on emoji function (M_{low_complementary} = 5.14, SD = 1.24 vs. M_{low_substitutive} = 5.23, SD = 1.63, F(1, 296) = 0.17, p = 0.685). When using two emojis, substitutive (vs. complementary) emojis led to significantly lower message evaluations (M_{high_complementary} = 5.47, SD = 1.27 vs. M_{high_substitutive} = 4.55, SD = 1.61, F(1, 296) = 15.14, p < 0.001). These results support Hypothesis 2.

Processing fluency

A two-way ANOVA revealed a main effect of emoji function (F(1, 296) = 9.15, p = 0.003), a marginal main effect of emoji number (F(1, 296) = 3.61, p = 0.058), and a significant interaction between emoji function and emoji number (F(1, 296) = 8.91 p = 0.003). When using one emoji, complementary and substitutive emojis did not significantly differ in terms of processing fluency (M_{low_complementary} = 6.04, SD = 0.85 vs. M_{low_substitutive} = 6.03, SD = 1.12, F(1, 296) = 0.001, p = 0.978). However, using two substitutive (vs. complementary) emojis decreased processing fluency (M_{high_complementary} = 6.17, SD = 0.90 vs. M_{high_substitutive} = 5.43, SD = 1.36, F(1, 296) = 18.17, p < 0.001). Results are summarized in Fig. 7.

Fig. 7.

Interactive effect between emoji function and emoji number (Study 3)

Moderated mediation analysis

Hayes’s (2013) PROCESS bootstrapping protocols (model 8, 5,000 iterations, Y = ad evaluation, X = emoji function, W = emoji number, M = processing fluency) confirmed significant indexes of moderated mediation for processing fluency (B = -0.25, SE = 0.09, 95% CI = [− 0.44, − 0.08]), providing support for Hypothesis 3. Specifically, when using multiple emojis, substitutive (vs. complementary) emojis had a negative indirect effect on message evaluation through reduced processing fluency (B =  − 0.26, SE = 0.07, 95% CI = [− 0.40, − 0.13]). When using one emoji, this negative indirect effect ceased to be significant (B =  − 0.002, SE = 0.06, 95% CI = [− 0.12, 0.11]).

Discussion

The findings of Study 3 provide causal support for emoji number as a boundary condition to the effectiveness of substitutive emojis (Hypothesis 2). As substitutive emojis provide the advantage of conveying semantic meaning at the cost of a single character, one may conclude that they represent an efficient way to convey information on digital platforms imposing character limits. Study 3 however cautions against this intuition by showing that communication efficiency does not necessarily translate into effectiveness: a high number of substitutive (vs. complementary) emojis reduces processing fluency (Hypothesis 3), lowering message evaluations.

Study 4: The moderating role of seller quality

Study 4 has three key objectives. First, to provide causal evidence for the 3-way interaction between emoji function, emoji number, and seller quality (Hypothesis 4). Second, to shed light into the mechanisms underlying the 3-way interaction, notably the mediating role of perceived competence (Hypothesis 5). Third, to rule out the possibility that the reduction in processing fluency observed in Study 3 is due to the placement of emojis at the beginning and at the end of the message.

Method

Study 4 follows a 2 (emoji function: complementary vs. substitutive) × 2 (emoji number: low vs. high) × 2 (seller quality: regular vs. premium) between-subjects design. Six hundred and twenty-six US residents were recruited from Prolific in exchange for £0.30. Thirteen participants speeding through the task were excluded from the analysis, leaving a final sample of 613 subjects (M_age = 38.29, SD = 12.98, 58.2% female).

Participants were asked to evaluate the listing for a rental property appearing on the Airbnb website among other listings. The stimuli were developed to replicate the Airbnb website’s look and feel to ensure ecological validity. The focal listing title “Dream house in the heart of the city, 5 min from the trains” was then adapted depending on the experimental condition. In the complementary function conditions, emojis appeared in conjunction with words with the same meaning. In the substitutive function conditions, emojis substituted those words. In the low emoji number conditions, we used one emoji. In the high emoji number conditions, we used three emojis. To manipulate seller quality, we added the Airbnb “superhost” badge to the premium seller conditions, and did not include this badge in the regular seller conditions. All stimuli are available in Web Appendix 4. Following random assignment, participants completed the same items for message evaluations (α = 0.92) as in Study 3. Next, they completed processing fluency measures (α = 0.92) as in Study 3, followed by two items measuring perceptions of competence (“The advertisement you just saw conveys: expertise/competence”; r = 0.78) drawn from Li et al. (2019). These latter measures represented the focal mediators from Hypothesis 5. Last, participants provided sociodemographic information. All multi-item measures were averaged to form a single index for each measure.

Results

Message evaluations

A three-way ANOVA revealed no significant main effect of emoji function (F(1, 605) = 2.18, p = 0.140) and seller quality (F(1, 605) = 2.53, p = 0.112), a significant main effect of emoji number (F(1, 605) = 35.58, p < 0.001), no significant interactions for emoji function × seller quality (F(1, 605) = 0.24, p = 0.625) or emoji number × seller quality (F(1, 605) = 0.86, p = 0.356), but a significant emoji function × emoji number interaction (F(1, 605) = 5.76, p = 0.017). Crucially, the 3-way interaction was significant (F(1, 605) = 7.05, p = 0.008). For regular sellers, results replicated the pattern found in Study 3: when using one emoji, there was no significant difference on message evaluations based on emoji function (M_{low_complementary} = 5.37, SD = 1.24 vs. M_{low_substitutive} = 5.69, SD = 1.04, F(1, 605) = 2.30, p = 0.130), but using multiple substitutive (vs. complementary) emojis led to significantly lower message evaluations (M_{high_complementary} = 5.19, SD = 1.23 vs. M_{high_substitutive} = 4.46, SD = 1.57, F(1, 605) = 11.99, p < 0.001). For premium sellers, however, this 2-way interaction disappeared, such that using one NF emoji led to comparably positive message evaluations (M_{low_complementary} = 5.67, SD = 1.10 vs. M_{low_substitutive} = 5.54, SD = 1.15, F(1, 605) = 0.40, p = 0.530) but using two or more emojis decreased message evaluations (M_{high_complementary} = 5.12, SD = 1.48 vs. M_{high_substitutive} = 5.04, SD = 1.31, F(1, 605) = 0.14, p = 0.711), irrespective of emoji function.

Processing fluency

A three-way ANOVA revealed a significant main effect of both emoji function (F(1, 605) = 21.29, p < 0.001) and emoji number (F(1, 605) = 61.23, p < 0.001), no significant main effect of seller quality (F(1, 605) = 0.30, p = 0.584), no significant interactions for emoji function × seller quality (F(1, 605) = 0.86, p = 0.353) or emoji number × seller quality (F(1, 605) = 1.54, p = 0.214), and, replicating the effects of Study 3, a significant emoji function × emoji number interaction (F(1, 605) = 17.80, p < 0.001). The 3-way moderation by seller quality was non-significant (F(1, 605) = 2.53, p = 0.112), so we collapsed across seller quality to examine the two-way interaction between emoji function and emoji number. Similar to Study 3, using one emoji made no difference on processing fluency based on the type of emojis used (M_{low_complementary} = 5.96, SD = 1.16 vs. M_{low_substitutive} = 5.91, SD = 1.06, F(1, 609) = 0.08, p = 0.776), but using multiple substitutive (vs. complementary) emojis led to significantly lower processing fluency (M_{high_complementary} = 5.57, SD = 1.26 vs. M_{high_substitutive} = 4.64, SD = 1.72, F(1, 609) = 37.94, p < 0.001).

Competence

A three-way ANOVA revealed significant main effects of emoji function (F(1, 605) = 9.61, p = 0.002), emoji number (F(1, 605) = 41.52, p < 0.001), and seller quality (F(1, 605) = 10.71, p < 0.001) on perceptions of sellers’ competence. There were no significant interactions for emoji function × seller quality (F(1, 605) = 0.43, p = 0.513) or emoji number × seller quality (F(1, 605) = 0.19, p = 0.662), but a significant emoji function × emoji number interaction (F(1, 605) = 8.28, p = 0.004). Importantly, the 3-way interaction was significant (F(1, 605) = 7.35, p = 0.007). For regular sellers, using one emoji made no difference on perceived competence based on the type of emoji used (M_{low_complementary} = 4.97, SD = 1.01 vs. M_{low_substitutive} = 5.14, SD = 0.98, F(1, 605) = 0.81, p = 0.367), but using multiple substitutive (vs. complementary) emojis led to significantly lower message evaluations (M_{high_complementary} = 4.85, SD = 1.17 vs. M_{high_substitutive} = 3.97, SD = 1.45, F(1, 605) = 21.22, p < 0.001). For premium sellers, the 2-way interaction disappeared, replicating the results pattern for message evaluations (M_{low_complementary} = 5.43, SD = 1.00 vs. M_{low_substitutive} = 5.22, SD = 1.10, F(1, 605) = 1.35, p = 0.246; M_{high_complementary} = 4.88, SD = 1.32 vs. M_{high_substitutive} = 4.64, SD = 1.20, F(1, 605) = 1.75, p = 0.186). That is, using multiple emojis reduced perceived competence irrespective of their complementary or substitutive function (F(1, 605) = 18.62, p < 0.001). Figure 8 summarizes these results.

Fig. 8.

Interactive effect between emoji function, emoji number, and seller quality (Study 4)

Moderated mediation analysis

Hayes’s (2013) PROCESS Model 11 (5,000 iterations, Y = ad evaluation, X = emoji function, W = emoji number, Z = seller quality, M₁ = processing fluency; M₂ = perceived competence) confirmed that perceived competence significantly mediated the 3-way interaction effect on message evaluation (Index of moderated mediation = 0.71, SE = 0.26, 95% CI = [0.20, 1.24]), supporting Hypothesis 5. Processing fluency did not mediate the 3-way interaction (Index of moderated mediation = 0.10, SE = 0.07, 95% CI = [− 0.03, 0.26]). PROCESS Model 7 (5,000 iterations, Y = ad evaluation, X = emoji function, W = emoji number, M = processing fluency) confirmed that processing fluency still significantly mediated the 2-way interaction between emoji function and emoji number on message evaluation (Index of moderated mediation =  − 0.47, SE = 0.12, 95% CI = [− 0.71, − 0.25]), providing converging support to Hypothesis 3.

Discussion

The results of Study 4 provide causal evidence for the interactive effects surfaced in Study 1, in particular for what concerns the moderating role of seller quality. Importantly, Study 4 sheds light on the mechanisms underlying these effects: the 3-way interaction between emoji function, emoji number, and seller quality is carried over to message evaluations through perceived competence. Processing fluency remains the key mechanism underlying the 2-way interaction between emoji function and emoji number, such that multiple instances of substitutive emojis reduce processing fluency. When including seller quality into the equation, however, perceived competence better explains the 3-way interaction on message evaluations: superhosts, likely expected to be professional in the way they communicate, suffer a substantial loss of perceived competence when using two or more (vs. one) NF emoji, regardless of emoji function. This finding confirms the notion that using emojis in contexts where professionalism is expected should be discouraged (Li et al., 2019), and provides converging evidence for Hypotheses 4 and 5.

General discussion

Key findings

Despite over 200 million Americans shopping online every year (Coppola, 2021), we know comparatively very little about how emergent forms of e-communication influence their message evaluations and eWOM volume. While the use of non-face (NF) emojis in 2021 has reached all-time high (Buchholz, 2021), extant marketing research on emojis is still in its dawning and predominantly examines the effects of face emojis on customer behavior (Das et al., 2019; Li et al., 2019; Smith & Rose, 2020). Against this background, the current research is the first to investigate the effects, mechanisms, and contingencies of NF emojis on customer behavior in online marketplaces. NF emojis, which we further distinguish based on their complementary or substitutive function, are more effective than text alone, yet their effectiveness is contingent on both message-level and source-level boundary conditions. At the message level, the differential effects of emoji function depend on their number: using multiple substitutive emojis in a message reduces processing fluency and, consequently, message evaluations and eWOM volume. At the source level, the advantage of NF emojis over text is contingent on seller quality, such that the use of multiple NF emojis is less beneficial for premium sellers, as they are expected to communicate more competently. Taken together, our findings offer important contributions to marketing literature and actionable implications for practitioners.

Theoretical contributions

We offer three contributions to literature on marketing communications and digital marketing strategy. First, we extend previous work on textual paralanguage (Luangrath et al., 2017) and face emojis (Das et al., 2019; Li et al., 2019; Smith & Rose, 2020) by being the first to formalize and test the effect of NF emojis on message evaluations and eWOM volume. Providing evidence for this effect is crucial for digital marketing as NF emojis represent 90% of all existing emojis and are used as frequently as face emojis. Using a large-scale dataset from Airbnb, we find that NF emojis have the third-largest impact on eWOM volume next to seller quality and ease-of-booking. Our empirical analysis is also the first to show that NF emojis can be more effective than face emojis in driving eWOM volume in digital marketplaces (see coefficients of face emojis compared to ${\beta }_1$ and ${\beta }_2$ in Appendix B), validating their real-world impact.

As the second contribution, we extend the emerging research on the linguistic functions of emojis (Danesi, 2017; Riordan, 2017a). While the use of emojis to complement or substitute text has been theorized in the past (Danesi, 2017; Li et al., 2019), no prior study has systematically investigated their differential effects. Our research addresses this gap while providing further insights on the message-level contingencies and underlying mechanisms. Using multiple substitutive NF emojis may save characters but ultimately reduces processing fluency, providing an important boundary condition to our effects as well as an extension of the multimodal communication literature (Paivio & Csapo, 1973; Tavassoli, 1998; Tavassoli & Lee, 2003). Previous research on multimodal communication studied the interplay of two pieces of information that are each complete on their own, finding a superior effect of multimodality versus unimodality due to lack of mutual interference. Our research examines the interplay of two incomplete pieces of information that jointly form a complete sentence (i.e., sentence embedding substitutive emojis). In such a condition, multimodality has an adverse effect because shifting between modalities reduces processing fluency, especially when using multiple substitutive emojis. Importantly, this finding refutes lay intuitions in marketing communications that embedding multiple emojis in a text will always have positive effects (McShane et al., 2021).

Third, this study shows how seller quality can trump positive message effects, specifically those of NF emojis. This finding runs opposite to lay understandings of combinatorial effects between a persuasive message and a high-quality source. One may reasonably expect that the presence of status badges signifying high seller quality overwhelms the heuristic power of NF emojis. However, the mechanism underlying this effect is not one of visual interference but one of linguistic stereotyping (Kang & Rubin, 2009; Lambert et al., 1960): the seller quality conveyed by a badge creates a high expectation on how premium sellers should communicate, limiting the benefits they could derive from using the informal NF emojis. In contrast, using NF emojis is a viable communication tactic for sellers who cannot sport high seller quality when promoting their offerings.

Managerial implications

The customer-based insights we distill are relevant to both small entrepreneurs communicating through digital channels (Hamilton, 2016) as well as marketing managers. Our field data shows that sellers who employ NF emojis in their property titles use an average of 2, yet our findings (based on Model 4 of Study 1) suggest that this practice may be detrimental based on emoji function and seller quality. Figure 9 and the guidelines below explain what type and how many NF emojis to embed in digital marketing communications based on seller quality, and suggest communication policy changes for providers of digital marketplaces.

Fig. 9.

Effect of NF emojis on eWOM volume by seller quality, emoji function, and emoji number (Study 1)

Recommendations for regular sellers

For regular sellers who cannot display status badges or other markers of quality, using multiple complementary emojis is the preferential strategy to increase message evaluations and eWOM volume. Regular users need actionable communication strategies to stand out from the competition, and NF emojis just help doing so. When saving characters is not of the essence, using one complementary emoji can increase eWOM volume by + 26.3% on average (95% CI = [16.9%, 35.8%]) and using multiple ones further increases this effect to + 35.6% (95% CI = [26.8%, 44.5%]). In situations where the seller instead needs to save precious characters, using one substitutive NF emoji can increase message evaluations and yield a + 21.4% average increase in eWOM volume (95% CI = [3.6%, 39.2%]). However, our findings caution that this strategy is effective only when used with parsimony: replacing multiple words with substitutive emojis is significantly worse than using a single substitutive NF emoji or even plain text, hindering processing fluency and harming message evaluations, with an average 27.4% reduction in eWOM volume relative to plain text (95% CI = [− 42.4%, − 12.4%]).

Recommendations for premium sellers

Our findings are particularly valuable for premium sellers as they surface a suboptimal current practice and explain how to correct it. In our Airbnb dataset, we observed a clear tendency of premium sellers to embed a high number of NF emojis in marketing communications: despite representing only 25% of the total sample, they are responsible for 54% of all multiple substitutive NF emoji observations. For premium sellers, however, using multiple emojis reduces perceived competence (Study 4). Our combined field and experimental results suggest refraining from their excessive use irrespective of emoji function, as the benefit of using multiple complementary NF emojis also diminishes significantly compared to regular sellers. We recommend that premium sellers use one complementary NF emoji to benefit from an average + 17.3% increase in eWOM volume (95% CI = [4.8%, 39.8%]). The total effect for using a single substitutive NF emoji is positive, but not significantly different from plain text (+ 7.3%, 95% CI = [− 5.8%, 20.3%]).

Recommendations for online marketplace providers

Our work also offers policy implications for providers of online marketplaces. Consider Airbnb and their content policy (Airbnb Content Policy, 2021), which clearly indicates that “listing titles that include symbols or emojis” are a policy violation that can lead to account suspension or permanent deactivation. Our findings suggest that for the majority of regular users in digital marketplaces (e.g., non-superhost), NF emojis are actually an effective differentiation tool in a rather competitive environment. Babić Rosario et al. (2020) suggest that one strategy to create more opportunities for eWOM volume creation is to leverage communication formats. The results of our experimental studies confirm this intuition and further demonstrate how this communication tactic yields greater results for regular than superhost users. We thus suggest that Airbnb and other e-commerce platforms that currently do not endorse the use of emojis update their content policies. E-platforms have a history of revising their user policies when the existing policies are sub-optimal. For instance, Twitter recognized the restrictions of 140 characters messages and doubled it up to 280. We believe similar policy changes regarding emojis would be beneficial to regular users. Second, our data shows that there is value in educating sellers. Conditional on using emojis in a listing, both regular and superhost Airbnb sellers use an average of 2 emojis. Regular sellers should reduce this number when they use substitutive emojis, whereas premium sellers should refrain from using more than one NF emojis irrespective of emoji function as overuse reduces perceptions of competence. We thus recommend that digital platforms such as Airbnb increase awareness of when and how to use NF emojis in sellers’ marketing communications.

Limitations and future research

Our research has some limitations that inspire avenues for further research. First, NF emojis are a heterogeneous category that includes a large number of different symbols. In our studies, we model the aggregate effect of NF emojis in a large secondary dataset and design our experimental stimuli with some of the most frequently used NF emojis (see Table 3; Unicode, 2021). The effects of complementary vs. substitutive emojis should hold regardless of the specific emojis used, as long as their implications on processing fluency remain the same. However, NF emojis may trigger heterogeneous semantic associations that affect competence inference differently. For example, some NF emojis which represent adorable symbols (e.g., teddy bears) may further decrease perceived competence compared to other NF emojis. Beyond distinguishing their function, future research is warranted to further classify NF emojis into discrete semantic categories and study their persuasive effects.

Second, we suggest that complementary NF emoji usage in real-life scenarios could also vary in terms of emoji position (start, middle, or end of the sentence) and semantic congruence between the emoji and the associated text. While the experimental stimuli placed NF emojis in a variety of locations and the effects held across both studies, the heterogeneous nature of emojis warrants further research on the particular effects of emoji position and semantic congruence of complementary emojis on message evaluations and eWOM.

Last, the prevalence of NF emojis in our dataset (more than 90% of all emojis) may be partially driven by the specific context investigated, where NF emojis are used mainly to complement or substitute the meaning of words in a one-way communication context. Digital communications on social media may employ different conversational norms and balance the use of face and NF emojis. For example, an online chat between customers and service agents (i.e., two-way communication) may involve more face emojis than Airbnb listings, although both may involve NF emojis. Future research is warranted to understand interaction effects between face and NF emojis.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Appendix A

Table 6.

Table 6.

Dependent, independent, and control variables

Variable	Description & Notes
Dependent Variables:
No. of reviews	No. of reviews since previous scrape
Independent Variables:
NF emoji presence	Whether NF emoji is present in Airbnb listing title (0/1)
Complementary/Substitutive NF emoji presence	Whether NF emoji with complementary/substitutive function is present in Airbnb listing title (0/1)
Emoji number	Whether two or more NF emojis are used in complementary or substitutive function in Airbnb listing title (0/1)
Complementary/Substitutive emoji number	Whether two or more NF emojis are used in complementary/substitutive function in Airbnb listing title (0/1)
Superhost (Yes = 1)	Whether host is classified as an Airbnb “superhost”
Listing Title Characteristics
Listing title word length	Length of Airbnb listing title
# Face Emojis	Number of face emojis used in Airbnb listing title
# Textual attention markers	Number of textual symbols (!, *, ~ , |,., ·, $, #, ^, + , %, &, @, (,), [,]) used in listing title to attract attention
Time-Variant Listing Features:
Rating quality score	Average review quality score at the time of the scrape
Price ($)	Per-nightly price
Security deposit ($)	Security deposit (0 if missing)
Cleaning fee ($)	Cleaning Fee (0 if missing)
Instant-bookable (Yes = 1)	Whether the listing is available for “Instant Booking”
Number of amenities	Number of amenities provided by listing
Description word length	Number of words in other textual description (e.g., summary of the listing and space, host description, etc.)
Description topic category	Most probable topic discussed in listing description (categorical variable with 7 topic categories)
Accommodation capacity	Total number of guests accommodated in listing
Cancellation policy	Flexible – full refund up to 24 h before check-in Moderate– full refund for cancellations within 5 days of check-in Strict—full refund for cancellations within 14 days of check-in
Exogenous Listing Features
Time since last scrape	Number of days elapsed between current and previous scrape
Listing age	Time since first review for Airbnb listing (months)
Property type	Entire place 0–1 bedroom – Studio or 1-bedroom entire place Entire place 2 + bedroom – Entire place with 2 + bedrooms Private room – Private room within a larger house/apartment
City	City of listing – New York City, Los Angeles, Chicago or Washington D.C
Neighborhood	Neighborhood of listing
Month	Month-year (Jan 2018 to Feb 2020)
Host Characteristics
# Host verifications	No. of verifications for Airbnb host – include email, phone, reviews, Govt. ID, Facebook, Google
Professional host (Yes = 1)	If the host owns more than 1 property
Host experience (years)	Time since host became an Airbnb host (till date of scrape)

Appendix B

Table 7.

Table 7.

Full model results

Dependent Var: log(1 + Number of reviews)	Model 1	Model 2	Model 3	Model 4
Independent Variables 1 :
Complementary NF emoji presence (${\mathrm{\beta }}_1)$		0.215***	0.200***	0.234***
		(0.026)	(0.033)	(0.038)
Substitutive NF emoji presence (${\mathrm{\beta }}_2)$		0.126* (0.051)	0.145** (0.052)	0.194** (0.075)
Complementary NF emoji presence*Complementary Number (${\mathrm{\beta }}_3)$			0.022 (0.043)	0.071 (0.047)
Substitutive NF emoji presence*Substitutive Number (${\mathrm{\beta }}_4)$			-0.432** (0.147)	-0.514*** (0.131)
Complementary NF emoji presence* Superhost (${\mathrm{\beta }}_5)$				-0.074 (0.065)
Substitutive NF emoji presence*Superhost (${\mathrm{\beta }}_6)$				-0.124 (0.098)
Complementary NF emoji presence*Complementary Number*Superhost (${\mathrm{\beta }}_7)$				-0.145* (0.073)
Substitutive NF emoji presence*Substitutive Number*Superhost (${\mathrm{\beta }}_8)$				0.193 (0.287)
Superhost (${\mathrm{\beta }}_9)$	0.369*** (0.001)	0.369*** (0.008)	0.369*** (0.008)	0.372*** (0.008)
Listing Title Characteristics 1 :
# Face Emojis	0.029	0.003	0.002	-0.008
	(0.044)	(0.049)	(0.049)	(0.051)
# Textual attention markers	0.014***	0.014***	0.014***	0.014***
	(0.002)	(0.002)	(0.002)	(0.002)
Listing title word length	0.009***	0.009***	0.009***	0.009***
	(0.002)	(0.002)	(0.002)	(0.002)
Listing title word length^2	-0.001***	-0.001***	-0.001***	-0.001***
	(1.3e-4)	(1.3e-4)	(1.3e-4)	(1.3e-4)
Time-Variant Listing Features 1 :
Rating Quality Score	0.003***	0.003***	0.003***	0.003***
	(1.1e-4)	(1.1e-4)	(1.1e-4)	(1.1e-4)
Price*10 (10 s of $)	-0.006***	-0.006***	-0.006***	-0.006***
	(3.3e-4)	(4.1e-5)	(4.1e-5)	(4.1e-5)
Security Deposit*10 (10 s of $)	-0.001***	-0.001***	-0.001***	-0.001***
	(1.6e-4)	(2.5e-5)	(2.5e-5)	(2.5e-5)
Cleaning Fee*10 (10 s of $)	-0.012***	-0.012***	-0.012***	-0.012***
	(0.001)	(8.3e-5)	(8.3e-5)	(8.3e-5)
Instant Bookable	0.228***	0.227***	0.227***	0.227***
	(0.007)	(0.001)	(0.001)	(0.001)
Listing description word length	4.0e-4***	4.1e-4***	4.1e-4***	4.1e-4***
	(2.7e-5)	(2.7e-5)	(2.7e-5)	(2.7e-5)
Listing description word length^2	-1.4e-7***	-1.4e-7***	-1.4e-7***	-1.4e-7***
	(2.9e-8)	(2.8e-8)	(2.8e-8)	(2.8e-8)
Description topic category: 2	0.013	0.012	0.012	0.011
	(0.007)	(0.007)	(0.007)	(0.007)
Description topic category: 3	-0.011	-0.013	-0.013	-0.014
	(0.027)	(0.007)	(0.027)	(0.027)
Description topic category: 4	-0.241***	-0.236***	-0.236***	-0.236***
	(0.018)	(0.018)	(0.018)	(0.018)
Description topic category: 5	0.048***	0.047***	0.047***	0.046***
	(0.009)	(0.009)	(0.009)	(0.009)
Description topic category: 6	0.100***	0.098***	0.098***	0.098***
	(0.007)	(0.008)	(0.008)	(0.008)
Description topic category: 7	0.035***	0.034***	0.034***	0.034***
	(0.008)	(0.008)	(0.008)	(0.008)
Accommodation capacity	0.040*** (0.002)	0.046*** (3.3e-4)	0.046*** (3.3e-4)	0.046*** (3.3e-4)
Number of amenities	0.007***	0.008***	0.008***	0.008***
	(3.1e-4)	(5.1e-5)	(5.1e-5)	(5.1e-5)
Cancellation Policy: Moderate	0.074***	0.074***	0.074***	0.074***
	(0.006)	(0.006)	(0.006)	(0.006)
Cancellation Policy: Strict	0.040***	0.040***	0.040***	0.040***
	(0.007)	(0.007)	(0.007)	(0.007)
Exogenous Listing Features
Listing age	-0.003***	-0.003***	-0.003***	-0.003***
	(1.3e-4)	(1.3e-4)	(1.3e-4)	(1.3e-4)
Time since last scrape (days)	0.001***	0.001***	0.001***	0.001***
	(7.3e-5)	(7.3e-5)	(7.3e-5)	(7.3e-5)
Host Characteristics:
# Host Verifications	0.001	2.7e-4	2.7e-4	2.6e-4
	(0.001)	(0.001)	(0.001)	(0.001)
Is Professional Host	0.012	0.010	0.010	0.009
	(0.011)	(0.011)	(0.011)	(0.011)
Fixed-Effects:
City-Property Type-Month	✓	✓	✓	✓
Neighborhood	✓	✓	✓	✓
Observations	2,189,224	2,189,224	2,189,224	2,189,224
Durbin-Wu-Hausman test of endogeneity	65.3***	66.9***	68.4***	69.8***
Within-R2	0.248	0.249	0.249	0.249
BIC	3,967,425	3,964,611	3,964,570	3,964,173
Likelihood Ratio Test Statistic (compared to Model 1)		2842***	2913***	3369***

Numbers in parentheses denote standard errors. Standard errors clustered by neighbourhood.

Significance codes—***:0.001, **0.01, *0.05.

1 – Second stage IV GMM results reported for endogenous variables

Models and model Fit We test four main specifications. Second-stage IV GMM results are reported. Durbin-Wu-Hausman test of endogeneity indicates the requirement for correcting for endogenous regressors. Model 1 tests the base models with no IVs (except for superhost status), Model 2 introduces complementary and substitutive NF emoji usage in listing titles. Model 3 tests for effects of low and high number of complementary and substitutive NF emojis. Finally, model 4 tests for the three-way interaction between function, number, and seller quality. Model fit improves over the four models (see R² and BIC in Table 3). We reject the likelihood ratio tests for null fit for models 2–4 relative to model 1.

Control variables We find that the effect of control variables is consistent across Models 1–4. Therefore, we discuss significant patterns for control variables in Model 1 (along with any deviations) below. Note that since the DV is in log-scale, we can interpret percentage change in reviews per unit change in IV as ${100\ast (e}^{\beta }-1)$.

1. Listing Title Characteristics: We find that, in line with Li et al. (2019), face emojis have a directionally positive effect on eWOM volume though insignificant (Model 1: $\beta =0.029$, n.s.). Similarly, textual attention markers (e.g., *, ~ , etc.) also positively affect eWOM volume (mean = 0.014, p < 0.01), presumably by increasing attention to the listing title. We also find a non-linear inverse-U role of listing title world length. Listings with 4–5 words have the highest effect on eWOM volume, while listing titles with more than 9 words deter customers. This underlines the role played by attention in attracting customers to potential travel listings.

2. Time-Variant Listing Features: We control for listing features that may simultaneously affect eWOM volume. First, we control for the listing quality which has an expected positive affect on eWOM volume ($\beta =0.003$, p < 0.001). Price (and analogously security deposit and cleaning fees) has a negative effect. A $100 increase in per-night price leads to a 5.8% fall in eWOM volume. Similarly, “instant-bookable” feature ($\beta =0.228$, p < 0.01), number of amenities ($\beta =0.007$, p < 0.001), accommodation capacity ($\beta =0.040$, p < 0.01), moderate ($\beta =0.074$, p < 0.001) and strict ($\beta =0.040$, p < 0.001) cancellation policies all have expected positive effects on eWOM volume. These results highlight the role of objective listing features in building eWOM volume in digital platforms.

   Listing descriptions pose another source of information for potential customers. Description word length shows a non-linear inverse-U role with eWOM volume. Optimally, listings should aim for 1400-1500 words (with maximum benefit at 1438 words). Therefore, customers prefer information but not too verbose. Second, listing descriptions can belong to multiple categories. Our analysis shows that descriptions focussing primarily on topic 6, i.e., on amenities and stay details have maximal effect on eWOM volume, again highlighting the role of information.

3. Exogenous Listing Features: Customers prefer newer listings ($\beta =-0.003$, p < 0.01), indicating preference for well-maintained properties. We also control for time between two successive web-scrapes to compare changes in listing bookings within the same time period.

4. Host Characteristics: Once we control for listing features and superhost status, we find that additional host characteristics such as number of host verifications or number of other listings owned by host (“professional host”) do not further affect eWOM volume.

Data Availability

The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request. The raw data for Study 1 is available on insideairbnb.com under Creative Commons license.

Declarations

Conflict of Interest

The authors declare that they have no conflict of interest.