Serving with a Smile on Airbnb: Analyzing the Economic Returns and Behavioral Underpinnings of the Host’s Smile

Abstract

Non-informational cues, such as facial expressions, can significantly influence judgments and interpersonal impressions. While past research has explored how smiling affects business outcomes in offline or in-store contexts, relatively less is known about how smiling influences consumer choice in e-commerce settings when there is no face-to-face interaction. In this article, we use a longitudinal Airbnb dataset and a facial attribute classifier to quantify the effect of a smile in the host’s profile photo on property demand and identify factors that influence when a host’s smile is likely to have the biggest effect. A smile in the host’s profile photo increases property demand by 3.5% on average. This effect is moderated by a variety of host and property characteristics that provide evidence for the role of uncertainty underlying why smiling increases demand. Specifically, when there is greater uncertainty regarding either the quality of the accommodations or the interaction with the host, a host’s smile will have a greater effect on demand. Online experiments confirm this pattern, offering further support for uncertainty perceptions driving the effect of smiling on increased Airbnb demand, and show that the effect of smiling on demand generalizes beyond Airbnb.

e-commerce, sharing economy, smile effect, machine learning, non-verbal cues, Airbnb demand

Smiling is a fundamental positive emotional expression that is known to affect interpersonal relationships and resulting behaviors (Bugental 1986; Wang et al. 2017) and typically has positive effects on a wide array of business outcomes (Barger and Grandey 2006; Söderlund and Rosengren 2008). While the effect of smiling has mainly been studied during physical in-person interactions, we explore a situation where direct interaction is absent: viewing a profile picture on the peer-to-peer home-sharing platform, Airbnb. Given Airbnb’s increasing importance in the marketplace, recent studies have investigated various influences of property demand, primarily focusing on the impact of informational cues (i.e., those that provide specific information), such as the property’s price and average guest ratings (Fradkin, Grewal, and Holtz 2018). Relatively less is understood about how non-informational cues (i.e., those that do not provide specific information, and instead operate by influencing perceptions or emotions) may affect consumer choice in this domain. We focus on exploring cues that convey the host’s emotions and traits. Specifically, we examine when and why a host’s smile in their profile picture affects demand for an Airbnb property.

Using observational Airbnb demand data spanning 8 months (study 1), we find that a smile in the host’s profile picture increases demand for the property, by approximately 3.5% on average. More importantly, we explore moderators to this effect, to demonstrate the conditions under which smiling is likely to have the biggest impact on demand. Building on prior research showing that the quality of the stay and the interaction with the host are the two primary considerations when evaluating an Airbnb property (Brochado, Troilo, and Shah 2017), we propose that smiling has a bigger effect on demand when there is greater uncertainty about the quality of the stay and/or the interaction with the host. This occurs because smiling creates a positive halo effect for the host, which increases perceptions of the host’s warmth and competence, thereby mitigating uncertainty and increasing demand.

Consistent with our proposed process, we examine a variety of host and property characteristics that we theorize will affect the degree of uncertainty associated with guests’ evaluations of an Airbnb property. Specifically, we investigate the impact of the host’s gender, the host’s level of experience, whether the accommodation is shared or private, the crime rate of the neighborhood, and the availability of parking (a proxy for safety). We predict that each of these variables will affect consumer certainty regarding the quality of the accommodations, interaction with the host, or both, which will moderate the effect of a host’s smile on property demand. For properties where consumers feel greater uncertainty, a host’s smile will reduce uncertainty to a greater extent, and in turn, have a greater effect on demand.

We further validate our hypothesis that a host’s smile increases demand, as well as the moderating role of uncertainty, in a post-test (study 2) and three controlled follow-up experiments (studies 3–5). Study 2 measures the extent to which each of the moderating variables is associated with greater uncertainty regarding the quality of the stay, the interaction with the host, or both. Study 3 is a controlled experiment testing the effect of smiling on the likelihood of booking a property, as well as downstream measures of uncertainty and host perceptions. We find that smiling increases demand by improving perceptions of the host’s warmth and competence, thus reducing the uncertainty associated with the stay and, in particular, the uncertainty regarding interacting with the host. Study 4 tests one of our fundamental moderators, the host’s gender, and finds that a host’s smile increases demand for an Airbnb property, and that the effect is greater for male hosts than female hosts. Study 5 extends this result to different hospitality contexts and finds the effect of smiling on demand generalizes to both a family-owned boutique hotel and a larger branded hotel.

Our research context provides a unique opportunity to explore how smiling influences customer choice in a real-world e-commerce setting, where face-to-face interactions are absent. The rich data, coupled with the machine-learning detection of smiling in a picture, allows us to quantify the effect of smiling at scale; indeed, the size of such effects may not be detectable nor quantified in a laboratory experiment. We also leverage behavioral insights to test various moderators at scale, such as local crime rates and the host’s experience, that would be difficult to vary effectively and naturalistically in a controlled setting. Further, we verify the causality of the effect of smiling in this context with a series of controlled experiments. Taken together, we explore when and why smiling may be most effective, as well as pinpoint the financial returns that a smiling host picture can yield, given the particular characteristics of the listing or the host.

We thus offer direct guidance to practitioners in the hospitality industry and provide notable insights to marketers more broadly about when smiling may be most effective. For example, by quantifying the effect of a smile in a host’s profile photo and identifying moderators to this effect, we underscore the importance of non-verbal cues in digital interactions and suggest when and for whom these cues are likely to matter most. This has important implications for how individuals present themselves online, especially in contexts where forming personal connections is pivotal or where there is inherent uncertainty about the quality of the service provider, such as with peer-to-peer accommodation services. Furthermore, the generalizability of the effect of smiling to hospitality contexts beyond Airbnb, as study 5 highlights, opens the door for further exploration into how these findings can be applied across various digital platforms and industries where consumer uncertainty might impact service or product providers to different extents.

THEORETICAL BACKGROUND

Extensive work has explored how non-informational cues can influence judgments and decisions in real or virtual interactions. Non-informational cues refer to those that do not provide explicit information about the properties or attributes of a stimulus; instead, they appeal to intuitive reasoning, often involving emotions, humor, or aesthetics. For instance, an individual’s attractiveness can influence people’s affective states, subsequently shaping their attitudes (Petty and Cacciopo 1984). One important category of non-informational cues in shaping impressions during social interactions are non-verbal communications, such as gestures, body language, manner of speech, and facial expressions. For instance, people view teachers who fidget as less effective at their jobs (Ambady and Rosenthal 1993), people who fold their arms as less ambitious but more extroverted (Gifford 1994), and people who speak quickly as more persuasive (Cesario and Higgins 2008). Facial expressions, in particular, can affect a wide range of interpersonal trait inferences. Anger and disgust expressions signal dominance, for example, whereas fearful and sad expressions signal the opposite (Knutson 1996).

Smiling is perhaps the most universal non-verbal cue for conveying positive interpersonal traits (Kraus and Chen 2013). Smiles have been shown to trigger social interaction, reduce perceptions of aggressiveness and physical dominance, and influence the emotions and actions of those nearby (Kraut and Johnston 1979). For example, smiling can be contagious, making others feel happy and react by smiling back (O’Doherty et al. 2003). Smiling can also serve as a social cue to convey altruism, signal the intent to cooperate (Bernstein et al. 2010), or express intrinsic motivation (Cheng, Mukhopadhyay, and Williams 2020), which can have important economic implications. Participants invested more money when their partners smiled, even if it meant lower monetary returns for themselves (Krumhuber et al. 2007; Scharlemann et al. 2001). Overall, individuals who smile are commonly perceived as more sociable, honest, pleasant, polite, kind, and warm (Bernstein et al. 2010; Mueser et al. 1984).

As a result of these positive interpersonal effects, smiling has been shown to positively impact a wide array of outcomes in service encounters, particularly those requiring direct interaction. Smiling heightens perceptions of warmth and improves the mood and emotional state of both people in the interaction, which together increase both the customer’s and the employee’s satisfaction (Barger and Grandey 2006; Söderlund and Rosengren 2008). Displays of positive emotions by sales employees generally improve employee evaluations and service quality ratings (Groth, Hennig-Thurau, and Walsh 2009), elevate customers’ post encounter-mood (Pugh 2001; Tsai and Huang 2002), enhance consumers’ consumption experiences (Barger and Grandey 2006), and ultimately increase willingness to return to the store (Tsai 2001). The effect of smiling on service performance is strongest when familiarity between the customer and employee is low (Gabriel, Acosta, and Grandey 2015), suggesting smiling is important in transactions among strangers.

The present article extends this literature by exploring the effect of smiling on consumer choice in a context with unique challenges and barriers for prospective customers. Our focus is on Airbnb, a leading, fast-growing sharing economy platform (5-year growth rate of 34.5% from 2016 to 2021; Seeking Alpha 2022), which enables individual hosts to list their homes and monetize the excess capacity of their properties (Sundararajan 2016). Airbnb transactions involve staying at a stranger’s property and corresponding with them to coordinate check-in and check-out, answering questions throughout the stay, and in some cases, meeting them in person or even staying in the property together. To navigate this decision, Airbnb guests can access an online listing featuring both informational cues (e.g., property features, location, price) and non-informational cues (e.g., the host’s profile photo). The majority of research to date exploring Airbnb demand has examined how consumers use informational cues to navigate this uncertainty and their resulting choices (Guttentag et al. 2018; Jiang 2019; Visser, Erasmus, and Miller 2017).

Our focus is on exploring the impact of a straightforward and actionable non-informational cue—a host’s smile in their profile picture—on demand for an Airbnb property, by leveraging a large dataset and machine learning techniques. We both quantify the effect of a host’s smile and demonstrate the conditions under which it is likely to have the greatest impact. While a significant majority of hosts on Airbnb feature smiles in their profile pictures, a notable proportion (25%) do not. The decision to present a picture with or without a smile has the potential to shape trait inferences that influence a guest’s likelihood of choosing to stay in a given property. We go beyond assessing the overall effect of smiling on demand to investigate key moderators and better understand when a host’s smile is most likely to have a meaningful effect.

While not directly explored in past literature, recent work suggests that smiling may affect business outcomes in online marketplaces where opportunities for person-to-person emotional contagion may be less likely. First, an unpublished paper exploring the crowdfunding space found that including a smiling photo of the creator increased contributions by 5% compared to when no face was disclosed (while a neutral expression decreased contributions; Kim and Park 2017). However, the authors do not control for other potential differences between projects with smiling creators and those with neutral or no pictures, nor do they offer evidence of why this effect occurred. A second paper found that the intensity of smiles affects crowdfunding contributions, with slight smiles leading to greater contributions than broad smiles (Wang et al. 2017). However, the authors only compared broad smiles to narrow smiles, rather than to neutral expressions, so the effect of smiling overall cannot be isolated. Moreover, the crowdfunding context differs in important ways from the sharing economy, where uncertainty related to staying in an unknown property is absent, and traits associated with competence and business acumen are more highly valued. Prior work has also found a positive correlation between positive expressions of a host and Airbnb property occupancy or interest in booking. However, they did not establish a causal link, investigate moderators, or offer support for an underlying mechanism (Deng and Ravichandran 2020).

Our main finding is that a host’s smile in their Airbnb profile picture will increase the likelihood of a consumer booking their property by 3.5% on average, although this increase depends on a variety of host and property characteristics. We propose that this increase in demand is driven by decreasing the uncertainty associated with Airbnb stays. Generally, people are averse to quality uncertainty, which can arise due to noisy signals and a lack of prior usage experience (Erdem and Keane 1996). Because each listing is unique and customers lack prior experience with specific options, uncertainty when booking an Airbnb is particularly high (Mao and Lyu 2017). The risk of staying in a stranger’s home and potentially having a negative interpersonal experience further compounds this uncertainty. In highly uncertain and unfamiliar contexts, consumers may prioritize avoiding uncertainty in their decision process (Bolton and Drew 1991; Campbell and Kirmani 2000). Indeed, consumers who exhibit higher levels of uncertainty avoidance tend to harbor more negative attitudes toward booking Airbnb properties (Lee, Nur Erdogan, and Hong 2021).

In the context of Airbnb, uncertainty may arise from two primary sources: (a) the interaction with the host and (b) the quality of the accommodations (e.g., the comfort and cleanliness). Text analyses of reviews also corroborate that consumers primarily evaluate an Airbnb property based on these two key dimensions (Bridges and Vásquez 2018; Brochado, Troilo, and Shah 2017; Cheng and Jin 2019; Guttentag et al. 2018). The quality and safety of the stay are a major consideration, and the interpersonal interactions significantly influence the satisfaction of staying in a stranger’s home as well, given the potentially extensive interactions between the guest and host. Non-informational cues that help to reduce uncertainty either regarding the quality of the stay or interactions with the host should thus improve perceptions of Airbnb properties. For instance, professional images that are brighter and clearer or that display more content (Zhang et al. 2021) have been shown to improve Airbnb property demand.

We propose that other non-informational cues, like a host’s smile in their profile picture, can similarly reduce uncertainty and increase demand. This occurs because smiling creates a positive halo effect that improves overall perceptions of the host, including assessments of their warmth and competence. Warmth perceptions involve judgments of others’ intentions, such as how kind, friendly, and well-meaning they are (Aaker, Vohs, and Mogilner 2010). Warmth is comprised of two sub-dimensions: sociability (i.e., friendliness; Brambilla et al. 2021) and morality (i.e., trustworthiness and ethicality/honesty; Brambilla and Leach 2014; Goodwin 2015). Because the primary assessment in interpersonal judgments is anticipating others’ intentions toward them, both dimensions of warmth are crucial for shaping overall impressions (Cuddy, Fiske, and Glick 2007). Competence perceptions relate to the ability to act on one’s intentions, including one’s capability, effectiveness, and power (Hoegg and Lewis 2011). Together, the warmth and competence dimensions explain much of the variation in how people perceive each other. While at a group level, warmth and competence tend to go in the opposite direction (Judd et al., 2005), at the individual level, they tend to be positively, though modestly, correlated (Fiske, Cuddy, and Glick 2007). Consistent with this notion, image characteristics of Airbnb host’s profile pictures, such as their smile, have been shown to influence perceived trustworthiness (Ert and Fleischer 2020), and hypothetical online studies have found that the perceived trustworthiness and positive facial expressions of the host correlate with interest in booking the property (Ert, Fleischer, and Magen 2016; Fagerstrom et al. 2017).

Given that warmth and competence are modestly correlated at the individual level, we propose that smiling will create an overall positive halo for the host, and similarly increase perceptions of competence. These improved perceptions of the host will mitigate uncertainty regarding the Airbnb property, which will increase demand. We therefore hypothesize:

H1: A host’s smile in their profile picture will increase Airbnb demand.

H2: A host’s smile will increase perceptions of the warmth and competence of the host, which will reduce the uncertainty related to the property and increase demand for the property. Together, perceptions of the host and perceptions of certainty will mediate the effect of smiling on demand.

Moderators of the Effect of Smiling on Demand

We further predict that the effect of smiling on Airbnb demand will be greater when the baseline level of uncertainty is higher—regarding either the host interaction, accommodation quality, or both—leading to a larger increase in certainty perceptions. We explore a variety of host and property characteristics that we anticipate will affect the degree of uncertainty associated with an Airbnb listing.

First, we examine two characteristics that we propose will primarily vary in the uncertainty surrounding the interaction with the host: the host’s gender and whether the stay is shared versus private.¹ We begin with the host’s gender, as it is a fundamental attribute that generalizes to contexts beyond Airbnb. We predict that male hosts will benefit more from a smile than female hosts, due to the greater uncertainty surrounding interactions with male hosts. This uncertainty stems from the general perception that men are less warm than women. Extensive research has documented that women are perceived as friendlier, more oriented toward the community, and more trustworthy than men (Eagly and Mladinic 1989; He, Inman, and Mittal, 2008; Noble, Pomering, and Johnson 2014). Women have a more communal orientation and smile more often than men (Hall and Halberstadt 1986; LaFrance and Hecht 2000; LaFrance, Hecht, and Paluck 2003), a pattern we also find in our Airbnb data (66% of male hosts vs. 82% female hosts in our sample smiled in the profile picture). These perceptions extend to professional settings, where women are perceived as warmer than men in various professions, ranging from caregivers (Halper, Cowgill, and Kimberly et al. 2019) to surgeons (Ashton-James et al. 2019), and, in our experiment (study 4), Airbnb hosts. As a result, guests may feel more uncertain about the quality of the stay and the interaction with host when evaluating the property of a male versus a female host.

Additionally, Airbnb offers two main accommodation options: a shared stay, where guests stay in the same home as the host and/or other guests (typically in a private room but not always); or a private stay, where guests have the entire place to themselves, including a private entrance. While interaction with the host occurs in both types of stays, the interaction is expected to be greater in shared stays, which increases the uncertainty surrounding an unpleasant interaction. We also predict that there will be more uncertainty regarding the quality of the accommodations in a shared arrangement, since the guest has less control over the common spaces (e.g., the bathroom or kitchen may be seen as less clean or comfortable when also used by others). Since smiles enhance perceptions of the host’s warmth and competence, a host’s smile will serve to decrease this uncertainty to a greater extent compared with more certain properties, and in turn increase demand to a greater extent. We thus hypothesize:

H3a: The host’s gender will moderate the effect of a host’s smile on Airbnb demand. A host’s smile will increase demand more for male hosts than for female hosts.

H3b: The nature of the stay (shared vs. private) will moderate the effect of a host’s smile on Airbnb demand. A host’s smile will increase demand more for shared stays than for private stays.

We also investigate characteristics that we predict would primarily heighten the uncertainty related to the quality of the accommodations, and therefore moderate the effect of smiling on demand: whether a host is experienced, the crime rate of the neighborhood, and the availability of parking. First, we explore the amount of experience that a host has on the Airbnb platform. When evaluating less (as opposed to more) experienced hosts, consumers will exhibit greater uncertainty regarding the quality of the accommodations. There also may be more uncertainty surrounding the interactions with a less established host. The next two factors, crime rate and parking, serve as proxies for perceived safety, which we contend directly influence judgments of certainty regarding the stay; a lack of parking is indeed correlated with higher local crime rates in our dataset. We propose that properties with higher local crime rates and properties without parking will be perceived to be more uncertain and will thus benefit more from a host’s smile. We therefore hypothesize:

H4a: The host’s experience will moderate the effect of a host’s smile on Airbnb demand. A host’s smile will increase demand more for less experienced hosts (fewer months on the platform) than for more experienced hosts (more months on the platform).

H4b: The local crime rate will moderate the effect of a host’s smile on Airbnb demand. A host’s smile will increase demand more for properties with higher local crime rates than lower crime rates.

H4c: The availability of parking will moderate the effect of a host’s smile on Airbnb demand. A host’s smile will increase demand more for properties that lack parking than properties that have parking.

Next, we examine the effect of a host’s smile on property demand, as well as the underlying mechanism and proposed moderators. We begin by using a large-scale data set of Airbnb bookings to estimate an econometric model, leveraging machine learning to automatically determine whether the host is smiling in their profile picture. We then test for causality in a controlled experimental setting, explore the underlying mechanism, and extend this result to non-Airbnb hospitality contexts. Data for the online experiments can be found at https://osf.io/7gr8e/?view_only=614532cf390c4ea58c66a8856eb47e11.

RESEARCH CONTEXT AND EMPRIRICAL DATA

Airbnb is a leading sharing economy platform for hosts to list their properties and for guests to find short-term accommodations. In 2019, more than two million guests stayed in more than seven million Airbnb listings worldwide.

Data Description

We combine multiple datasets related to Airbnb property bookings and characteristics (AirDNA), neighborhood information (Walkscore.com, Zillow, and American Community Survey [ACS]), as well as Airbnb host profile photos (web scraping). The datasets are merged using the Airbnb Host ID (merging property data and profile photos), property address, and zip code (merging property data and neighborhood data). Table 1 presents the main variables, their data sources, and definitions. We include these variables in our econometric model to control for potential confounds. Web appendix section 7.1 describes the sample construction.

TABLE 1

DATA DESCRIPTIONS

Data source	Variables	Definition
AirDNA: listing-level panel data on Airbnb property bookings and property characteristics	Property demand	The number of days the property is booked (reservation days) in a month. Log transformed.
	Property nightly rate	The average nightly rate in each month.
	Time-invariant characteristics	The number of bedrooms, property type (e.g., house or apartment), listing type (e.g., private vs. shared place), and property visual appeal (property image quality)
	Time-variant characteristics	The accumulative count of guest reviews, the number of posted property photos, the length of time (in months) since a property was listed, or the cancellation policy.
Walkscore.com: data on local accessibility	Walk score	A higher score indicates that the area is more walkable.
	Transit score	A higher score indicates better access to public transit.
	Drive time	Average time driving from the property to downtown.
Zillow: home value	Home value	The average home value across homes in that zip code with the same size as the focal property.
ACS	Home median earning	The median of earning and percentage of bachelor’s degrees in zip code, conditioned on host gender and age.
5-Year estimates	Bachelor
Airbnb host profile pictures: deep learning model predictions.	Smile	A binary variable that equals 1 if the host smiled in the profile photo, and 0 if otherwise.
	Gender	A categorical variable which equaled 1 (0) if the host was predicted to be male (female).a
	Age	The predicted age of a host as number of years.

a

For the sake of conceptual clarity, we excluded hosts who were couples (approximately 10% of our sample). In a supplementary analysis, we included couples in our sample and obtained comparable results. Specifically, a host’s smile does not affect demand for couples, possibly because the fact that the hosts are a couple is sufficient to induce warmth perceptions and reduce uncertainty (web appendix 7.3).

Facial Analysis Using a Deep Learning Model

As described in table 1, we build a deep learning model to predict the age, gender, and smile of each host from host profile pictures (figure 1). The workflow steps to analyze the faces of Airbnb host photos are summarized in figure 2. We briefly describe how we extract these variables at scale. A detailed description and technical notes are in web appendix section 1.

FIGURE 1

AN EXAMPLE OF AIRBNB HOST PAGE

NOTE.—A screen shot of part of the content of the host page where the host’s profile photo is shown.

FIGURE 2

A SKETCH OF FACIAL ANALYSIS: ARCHITECTURE OF RESNET-50 AND WORKFLOW

We employ the ResNet-50 model developed by Cao et al. (2018), which is optimized for facial recognition using 3.3 million face images of more than 9,000 participants. Figure 2 presents the architecture of ResNet-50 and the description of a residual unit. To transfer ResNet-50 to our tasks, we fine-tune the model’s parameters on a training set of images for which the key facial attributes—smile, gender, age––were labeled. We consolidate the IMDb-WIKI image database (Rothe, Timofte and Van Gool 2015) for the age and gender classification task and multiple datasets for the smile detection task: Cohn–Kanade (Kanade, Cohn, and Tian 2000), GENKI-4K (Whitehill et al. 2009), and CelebFaces Attributes (Liu et al. 2015). We use 80% of the images for fine-tuning the ResNet-50 model and the remaining 20% (unseen samples) for evaluating the model. Using transfer learning to build a custom facial analysis model, rather than an off-the-shelf model, allows for flexibility in addressing biases and limitations in pre-trained models, ensuring superior performance.

Addressing Overfitting

Deep learning models explore complex non-linear structures to achieve a high degree of accuracy in prediction. Yet, the complexity of the model structure may cause overfitting on the training sample and reduce the model’s generalizability to unseen samples. Three optimization tactics help us deal with this potential issue. First, overfitting tends to arise when the model is too complex relative to the size of the training sample. Our model is trained on a large sample to alleviate this problem: 3.3 million images were used in Cao et al. (2018) and more than 800,000 images were used in the fine-tuning step. Second, we employ real-time data augmentation to introduce random variations in the training samples, a technique that has been shown to effectively reduce overfitting (Krizhevsky, Sutskever, and Hinton 2012). Third, L2 regularization helps reduce overfitting by preventing the model from assigning large weights (Goodfellow, Bengio, and Aaron 2016).

Model Classification Performance

On the 20% hold-out sample, our model achieves a high degree of accuracy: 98.8% for predicting gender, 92.4% for detecting a smile, and 5.75 as a mean absolute error for predicting age.²

STUDY 1: EMPIRICAL ANALYSES ON A LARGE-SCALE DATA SET

Study 1 investigates hypotheses 1, 3a, 3b, and 4a–c. We estimate an econometric model on observational Airbnb property demand data spanning eight months (January to August) in 2016. The panel data consist of properties of 9,248 unique hosts located in 359 zip codes across seven U.S. cities—Austin, Boston, Los Angeles, New York City, San Diego, San Francisco, and Seattle. The property demand is measured by the number of days booked in a month. Because this is a count measure, we log transform it to compute the dependent variable (DV) of the econometric model.

To account for the potential differences in the property characteristics between properties with versus without a smile in the host profile picture, we first apply inverse probability of treatment weighting (IPTW; see Giorcelli 2019) to construct a matched sample of smiling and non-smiling groups in which the distribution of measured host, property, and neighborhood covariates is independent of smiling condition. The IPTW approach contains two primary steps. One, estimate each property’s smiling probability as a function of observed covariates, then construct the weights for each property as the inverse of its smiling probability (see web appendix section 6.1 for the full list of covariates and a description of the method). Two, in the matched sample, we estimate a set of econometrics models using the constructed weights.

We predict that a host’s smile will increase Airbnb property demand, and a variety of factors may moderate this increase. However, due to the nature of observational data, we cannot pin down causality. Follow-up online studies confirm the reported pattern and establish a causal link between smiling and property demand to support the findings in study 1.³

Main Econometrics Specification on Airbnb Property Demand

In the matched sample, we estimated the following:

$$Y_{it}=\mathit{Intercept}+\beta ·\mathit{Smil}{e}_i+\lambda ·\mathit{Control}{s}_{it}+{\mathit{Seasonality}}_{c,t}+\mathit{Zipcod}{e}_i+{\epsilon }_{it}$$

(1)

where $Y_{it}=\mathit{log}(\# \mathit{Researvation} \mathit{Days}+1)$ indicates the demand for property i in time (month) t, and the binary variable $\mathit{Smil}{e}_i$ is 1 if the host of property i has a profile photo with a smile feature presented. Hence, the key parameter $\beta$ captures the average change in the property demand for hosts who smiled in the profile photos, compared with those who did not. $\mathit{Control}{s}_{it}$ is a set of control variables associated with property i in time t—both time-variant and time-invariant property characteristics—as summarized in table 2. The rich set of control variables (table 2) helps mitigate the concern that other variables correlate with both whether a host smiles in their profile picture and other factors that affect the property’s demand. For example, higher quality or more professional hosts may be more likely to smile in their photo and also attract greater demand for their property due to better service, rather than because of smiling. To address such concerns, we include variables that capture the quality of a host—Super Host (whether a host reaches Super Host status, which is a high standard status that Airbnb provides to recognize a host as “the people who are most dedicated to providing outstanding hospitality”) and Listed Month (the number of months since a property was listed). Lastly, we included two sets of fixed effects: (1) ${\mathit{Seasonality}}_{c,t}$ is the time fixed effects at the city–month pair level, allowing each city to have its own seasonality pattern, and (2) $\mathit{Zipcod}{e}_i$ is the zip code (the zip code of property i) fixed effects, capturing the location- and neighborhood-specific unobserved factors that affect Airbnb property demand. Lastly, ${\epsilon }_{it}$ is a random shock term in property i’s demand in time t, assumed to be i.i.d. normally distributed. We log transformed count data (such as the number of booked days and number of guest reviews), as well as the property nightly rate.

TABLE 2

SUMMARY STATISTICS: GROUPED BY HOST GENDER

Variables	(1)		(2)		(3)
	Male hosts		Female hosts		All hosts
	Mean	SD	Mean	SD	Mean	SD
Airbnb host facial attributes (measured from profile photos)
Smile	0.66	0.47	0.82	0.38	0.75	0.43
Host age	36.55	9.89	33.65	9.82	34.93	9.95
#Photographed faces	1.31	0.84	1.43	1.08	1.38	0.98
Airbnb property performance
# Reservation days	6.52	10.03	6.13	9.64	6.30	9.82
Listing nightly rate	240.31	314.88	250.74	291.15	247.24	301.66
# Non-blocked days	19.03	13.49	18.55	13.48	18.76	13.49
Property neighborhood information
Walk score	83.14	24.32	82.55	24.49	82.81	24.42
Transit score	76.73	23.48	76.43	23.25	76.56	23.36
Drive time (minutes)	14.75	10.03	14.62	10.14	14.68	10.10
Home value (1000 USD)	711.65	409.22	700.70	417.03	705.51	413.64
Median-Home Earning (1000 USD)	56.77	24.74	44.45	15.55	49.87	21.03
Airbnb property characteristics
# Bedrooms	1.36	1.01	1.37	0.95	1.37	0.97
House	0.29	0.45	0.32	0.46	0.30	0.46
# Guest reviews	39.53	54.66	34.82	52.29	36.85	53.38
Property visual appealing	0.41	0.25	0.41	0.26	0.41	0.26
# Property photos	17.71	13.10	18.27	13.53	18.03	13.35
Flexible cancellationa	0.047	0.21	0.040	0.20	0.043	0.20
Security deposit	147.24	366.71	147.73	326.85	147.51	344.97
Cleaning fee	36.94	55.94	44.22	61.59	41.02	59.28
Max guests	3.64	2.69	3.51	2.43	3.57	2.55
Minimum stay	4.47	14.91	4.74	11.26	4.62	12.96
Super host	0.22	0.41	0.19	0.39	0.20	0.40
Business ready	0.14	0.35	0.13	0.33	0.13	0.34
# Listed month	31.42	18.22	31.50	19.04	31.50	18.67
Parking	0.48	0.50	0.50	0.50	0.49	0.50
Pool	0.08	0.28	0.08	0.27	0.08	0.28
Beach	0.01	0.10	0.02	0.13	0.01	0.12
Internet	0.97	0.17	0.98	0.15	0.98	0.16
TV	0.76	0.43	0.74	0.44	0.75	0.43
Essentials	0.57	0.49	0.60	0.49	0.59	0.49
Laptop-friendly	0.38	0.48	0.41	0.49	0.39	0.49
Family-friendly	0.19	0.39	0.20	0.40	0.19	0.39

a

If the host allows a flexible cancelation, then guests can cancel until 24 hours before check-in for a full refund.

Table 3 reports the empirical results of estimating equation (1) on the Airbnb property demand data. Confirming our hypothesis (hypothesis 1), the estimated coefficient of Smile, 0.035 (p < .01), suggests that, on average, presenting a smile in the Airbnb host profile photo increases property demand by 3.5% (= exp^(0.0358) –1). For an average property,⁴ this effect is equivalent to additional annual revenue of $672.80 (which equals 6.30 booked days/month × $247.20/day × 12 months/year × 3.5%). Note that the coefficient for Male suggests that, on average, male hosts earned 4.5% less than female hosts (= 1 – exp^(–0.0464)). This discrepancy might be explained by female hosts being better received on Airbnb, a platform focused on lodging and home-related services. Therefore, smiling could be a tool to enhance equity on Airbnb.

TABLE 3

IMPACT OF A HOST’S SMILE ON AIRBNB PROPERTY DEMAND

	Main effect
Variables	Coefficients	SE
Smile	0.0358***	(0.00966)
Male	–0.0464***	(0.0137)
log NightlyRate	–0.119***	(0.0190)
log # nonBlockedDays	0.513***	(0.0117)
log number of reviews	0.347***	(0.00730)
log number of photos	0.00810	(0.0121)
Bedrooms	0.0148	(0.00916)
Cleaning fee	0.000324*	(0.000144)
Max guests	0.0197*	(0.00939)
Minimum stay	–0.00911***	(0.000844)
Security deposit	–0.0000401*	(0.0000177)
log # ListedMonth	–0.395***	(0.0136)
Super host	0.00984	(0.0145)
Business ready	0.148***	(0.0277)
House	–0.0427**	(0.0131)
Walk score	0.000263	(0.000332)
Transit score	0.00246**	(0.000855)
Drive time (minutes)	–0.00224	(0.00136)
Median home earning	0.00313***	(0.000692)
Home value	0.0000769***	(0.0000191)
Bachelor	–0.00372***	(0.000798)
# Photographed faces	–0.0162	(0.00987)
Age	–0.000870	(0.000492)
Internet	0.123***	(0.0361)
TV	–0.00327	(0.0172)
Parking	0.0324*	(0.0138)
Essentials	0.112***	(0.0196)
Laptop-friendly	0.0159	(0.0129)
Pool	–0.134***	(0.0279)
Beach	0.0984*	(0.0408)
Family-friendly	–0.0291	(0.0177)
Flexible cancellation	0.140***	(0.0318)
Property visual appealing	0.0830***	(0.0219)
Intercept	0.993***	(0.116)
Fixed effect	Zip code
Seasonality	City–month
Observations	73,754
R2	0.6893

NOTE.—DV: Logged number of reservation days in a month.

*

p < .05,

**

p < .01,

***

p < .001.

In the next two sections, we further explore the heterogeneity in the effect of a host’s smile across groups of Airbnb hosts and properties. As discussed in the earlier section, we investigate a set of host and property characteristics that might strengthen or weaken the effect of a host’s smile, depending on the degree of uncertainty regarding the interaction with the host or the quality of the stay.

Interacting a Host’s Smile with Characteristics That Vary in Uncertainty Regarding the Host Interaction

We test two characteristics—gender and shared/private stay—that we theorize will vary in the degree of uncertainty regarding the interaction with the host. We let equation (2) specify the regression that examines the differential effects of presenting a smile across characteristics:

$$Y_{it}=\mathit{Intercept}+\beta ·\mathit{Smil}{e}_i+\gamma \mathit{Cha}{r}_i+\delta ·\left(\mathit{Smil}{e}_i\times \mathit{Cha}{r}_i\right)+\lambda ·\mathit{Control}{s}_{it}+{\mathit{Seasonality}}_{c,t}+\mathit{Zipcod}{e}_i+{\epsilon }_{it}$$

(2)

where $\mathit{Cha}{r}_i$ serves as one of the following two moderators: $\mathit{Mal}{e}_i$⁠, which equals 1 (rep. 0) if i is male (rep. female); $\mathit{Stay} \mathit{alon}{e}_i$⁠, which equals 1 if the listing type was an entire space and 0 if otherwise. If the property is rented out as an entire space, then the guests have the whole space to themselves (58.2% of our data). Note shared stays may be shared with either the host and/or another guest(s)—the dataset does not differentiate between these possibilities. However, properties that are not listed as an entire place have a greater chance that the guest may be staying with the host, which comes with greater uncertainty regarding both the interaction with the host and quality of the accommodation, as they would not have full control of the spaces.

In equation (2), our DV of interest is $\mathrm{\delta }$—the coefficient of the interaction term $\mathrm{S}\mathit{mil}{e}_i\times \mathit{Cha}{r}_i$⁠, which captures the differential impact of a smile on $Y_{it}$ for male hosts compared with female hosts and for entire (private) stays compared with shared stays.

Table 4 reports the estimation results, which confirm our hypotheses in hypothesis 3. In column (1), the estimates of the main effect of Smile and its interaction effect, Male × Smile, suggest a larger effect of a smile on property demand for male hosts, compared with the effect for female hosts (hypothesis 3a). Specifically, a smile in the host’s profile photo did not have a significant effect on property demand for female hosts (⁠$\mathrm{\beta }$ = –0.00229, p > .1). The key coefficient $\mathrm{\delta }$ is positive and statistically significant (⁠$\mathrm{\delta }$= 0.0830, p < .001), implying that a smile in their profile photo would increase demand by 8.7% (= exp^(0.083) –1) for male hosts. This translates to additional annual revenue of $1,625⁵ for male hosts compared to female hosts. We posit that this gender gap in the effect of smiling on demand may be driven by a difference in the baseline levels of warmth perception between the two gender groups (Eagly and Mladinic, 1989), which leads to higher baseline levels of uncertainty regarding the interaction with male hosts than female hosts.

TABLE 4

IMPACT OF A HOST’S SMILE ON AIRBNB PROPERTY DEMAND: TESTING MODERATORS REGARDING HOST INTERACTION UNCERTAINTY

Variables	(1)		(2)
	Moderated by host gender		Moderated by nature of stay
	Coef.	SE	Coef.	SE
Smile	–0.00229	(0.0142)	0.0762***	(0.0196)
Smile × male	0.0830***	(0.0203)
Smile × stay alone			–0.0602*	(0.0286)
Male	–0.0910***	(0.0215)	–0.0496***	(0.0137)
Stay alone			0.352***	(0.0234)
log NightlyRate	–0.119***	(0.0190)	–0.251***	(0.0207)
log # nonBlockedDays	0.513***	(0.0117)	0.524***	(0.0117)
log number of reviews	0.347***	(0.00728)	0.333***	(0.00749)
log umber of photos	0.00893	(0.0121)	0.0107	(0.0118)
Bedrooms	0.0141	(0.00918)	0.0456***	(0.00949)
Cleaning fee	0.000321*	(0.000144)	0.000485***	(0.000147)
Max guests	0.0200*	(0.00940)	0.00511	(0.00914)
Minimum stay	–0.00911***	(0.000846)	–0.00990***	(0.000903)
Security deposit	–0.0000399*	(0.0000177)	–0.0000439*	(0.0000177)
log # ListedMonth	–0.393***	(0.0136)	–0.385***	(0.0134)
Super host	0.0115	(0.0143)	0.0268	(0.0146)
Business ready	0.148***	(0.0277)	0.103***	(0.0269)
House	–0.0424**	(0.0131)	0.0150	(0.0136)
Walk score	0.000287	(0.000330)	0.000327	(0.000326)
Transit score	0.00239**	(0.000856)	0.00223**	(0.000852)
Drive time (minutes)	–0.00232	(0.00137)	–0.00294*	(0.00138)
Median home earning	0.00318***	(0.000692)	0.00352***	(0.000692)
Home value	0.0000777***	(0.0000190)	0.0000679***	(0.0000190)
Bachelor	–0.00379***	(0.000799)	–0.00405***	(0.000798)
# Photographed faces	–0.0153	(0.00977)	–0.0244*	(0.00994)
Age	–0.000696	(0.000500)	–0.00107*	(0.000489)
Internet	0.124***	(0.0361)	0.133***	(0.0353)
TV	–0.00234	(0.0172)	–0.0244	(0.0166)
Parking	0.0335*	(0.0138)	0.0403**	(0.0135)
Essentials	0.114***	(0.0196)	0.109***	(0.0193)
Laptop-friendly	0.0159	(0.0129)	0.00140	(0.0126)
Pool	–0.127***	(0.0273)	–0.0934**	(0.0285)
Beach	0.0992*	(0.0408)	0.113**	(0.0408)
Family-friendly	–0.0281	(0.0178)	–0.0512**	(0.0177)
Flexible cancellation	0.140***	(0.0318)	0.141***	(0.0317)
Property visual appealing	0.0788***	(0.0216)	0.0897***	(0.0219)
Intercept	1.005***	(0.116)	1.484***	(0.117)
Fixed effect	Zip code		Zip Code
Seasonality	City–month		City–month
Observations	73,754		73,754
R2	0.6894		0.6932

NOTE.—DV: logged number of reservation days in a month.

*

p < .05,

**

p < .01,

***

p < .001.

Moving to column (2), consistent with hypothesis 3b, we find that the coefficient of Smile × Stay alone is significant and negative (⁠$\mathrm{\delta }$ = –0.0602, p < .05; column (2); correspondingly, the coefficient of Smile was positive and significant). This suggests that the effect of a host’s smile has a greater effect, specifically, 5.8% greater, when the guest shares the space with others than when the guest stays in the property alone, since the former implies a higher chance of interacting with the host. Staying in a shared place versus in an entire place suggests a greater degree of uncertainty in two regards: a potential unpleasant interaction and lower control over the common spaces. Shared stays thus benefit more from a host’s smile and increased feelings of certainty, as compared to private stays.

Interacting a Host’s Smile with Accommodation Quality Characteristics

We test three accommodation characteristics that we theorize will affect the degree of uncertainty regarding the accommodation quality: the host’s level of experience, the local crime rate, and access to parking. We construct the following three moderators and incorporate them into equation (2). First, we let Host Experience denote whether the host was experienced or not on Airbnb, which was operationalized as whether the host has been active more than 7 months (the sample mean) in the previous year. Second, from the local crime rate, we construct a variable to capture how risky it is to stay at a property. CrimeGrade (https://crimegrade.org/) evaluates lcoal risk based on a range of crimes, including violent crime rate (assault, robbery, etc.), property crime rate (vehicle theft, burglary, etc.), and other crime rates (kidnapping, drug crimes, etc.). CrimeGrade reports crime frequency as a measure of the risk or safety of a neighborhood. A crime occurs every 11 hours on average across the properties in our sample. To account for the fact that the incidence of crime may increase with the population in the area, we compute a normalized crime frequency, Crime Hours (normalized by the population of the neighbhorhood) to capture the local risk level for Airbnb properties. Areas with higher Crime Hours have less-frequent occurrences of crimes and hence are less risky.

Finally, each property describes the parking situation as a feature on its property page. We let the dummy variable Parking equal 1 if the property provided free parking on the property. We find that 49% of the properties in our data had access to parking. Parking may provide better privacy and is rated by Airbnb guests in the United States as a top property amenity.⁶ The fact that a property has free parking may also suggest that the surrounding neighborhood is safe—and indeed, our data show that access to parking was positively correlated with a low local crime rate; we find that the average crime frequency for properties without parking was 77% higher than that for properties with parking (t = 10.8, p < .001).

We replicate the moderator model as specified in equation (2). Table 5 reports the estimation results, which tests hypotheses 4a–c. In column (1), the negative and significant coefficient of Smile × Host Experience reveals that a smile in the profile picture improved demand more for properties with less-experienced hosts than for properties with experienced hosts (coefficient of Smile × Host Experience is –0.0652, p < .01). Inexperienced or new hosts pose a greater degree of uncertainty surrounding the quality of the accommodations (Wernerfelt 1988). As a result, an inexperienced host’s smile generates an extra 6.3% (=1 – exp^(–0.0652)) increase in demand relative to an experienced host’s smile.

TABLE 5

IMPACT OF A HOST’S SMILE ON AIRBNB PROPERTY DEMAND: TESTING MODERATORS REGARDING ACCOMMODATION QUALITY UNCERTAINTY

	(1)		(2)		(3)
	Moderated by host experience		Moderated by local crime rate factor		Moderated by access to parking
VARIABLES	Coefficients	SE	Coefficients	SE	Coefficients	SE
Smile	0.0634***	(0.0145)	0.0383***	(0.00971)	0.0582***	(0.0139)
Smile × host experience	–0.0652*	(0.0258)
Smile × local crime rate factor			–2.778***	(0.600)
Smile × parking					–0.0388*	(0.0197)
Male	–0.0460***	(0.0137)	–0.0442**	(0.0137)	–0.0466***	(0.0137)
Host experience	0.0832***	(0.0252)
Local crime rate factor			–1.588*	(0.655)
log NightlyRate	–0.115***	(0.0185)	–0.118***	(0.0190)	–0.119***	(0.0190)
log # nonBlockedDays	0.513***	(0.0117)	0.513***	(0.0117)	0.513***	(0.0117)
log number of reviews	0.348***	(0.00730)	0.347***	(0.00730)	0.347***	(0.00731)
log number of photos	0.00995	(0.0119)	0.00838	(0.0121)	0.00867	(0.0121)
Bedrooms	0.0156	(0.00912)	0.0148	(0.00916)	0.0147	(0.00916)
Cleaning fee	0.000328*	(0.000144)	0.000311*	(0.000144)	0.000328*	(0.000144)
Max guests	0.0175*	(0.00889)	0.0197*	(0.00939)	0.0197*	(0.00939)
Minimum stay	–0.00915***	(0.000845)	–0.00911***	(0.000844)	–0.00909***	(0.000844)
Security deposit	–0.0000423*	(0.0000176)	–0.0000405*	(0.0000177)	–0.0000401*	(0.0000177)
log # ListedMonth	–0.396***	(0.0133)	–0.394***	(0.0136)	–0.395***	(0.0136)
Super host	0.00996	(0.0144)	0.00905	(0.0145)	0.0105	(0.0145)
Business ready	0.143***	(0.0270)	0.148***	(0.0277)	0.149***	(0.0278)
House	–0.0407**	(0.0130)	–0.0429**	(0.0131)	–0.0421**	(0.0131)
Walk score	0.000295	(0.000327)	0.000261	(0.000332)	0.000279	(0.000330)
Transit score	0.00245**	(0.000855)	0.00245**	(0.000855)	0.00250**	(0.000857)
Drive time (minutes)	–0.00229	(0.00136)	–0.00224	(0.00136)	–0.00220	(0.00136)
Median home earning	0.00317***	(0.000689)	0.00297***	(0.000692)	0.00309***	(0.000692)
Home value	0.0000803***	(0.0000188)	0.0000782***	(0.0000191)	0.0000770***	(0.0000191)
Bachelor	–0.00371***	(0.000800)	–0.00353***	(0.000799)	–0.00368***	(0.000799)
# Photographed faces	–0.0132	(0.00966)	–0.0160	(0.00987)	–0.0169	(0.00983)
Age	–0.000989*	(0.000488)	–0.000906	(0.000493)	–0.000880	(0.000493)
Internet	0.124***	(0.0360)	0.123***	(0.0361)	0.124***	(0.0361)
TV	–0.00576	(0.0167)	–0.00309	(0.0172)	–0.00348	(0.0172)
Parking	0.0338*	(0.0136)	0.0322*	(0.0138)	0.0526*	(0.0205)
Essentials	0.109***	(0.0191)	0.113***	(0.0196)	0.112***	(0.0196)
Laptop-friendly	0.0144	(0.0127)	0.0158	(0.0129)	0.0161	(0.0129)
Pool	–0.130***	(0.0276)	–0.134***	(0.0279)	–0.135***	(0.0278)
Beach	0.0920*	(0.0408)	0.0986*	(0.0408)	0.0979*	(0.0408)
Family-friendly	–0.0304	(0.0176)	–0.0288	(0.0178)	–0.0286	(0.0177)
Flexible cancellation	0.140***	(0.0318)	0.139***	(0.0318)	0.139***	(0.0318)
Property visual appealing	0.0783***	(0.0216)	0.0820***	(0.0220)	0.0826***	(0.0220)
Intercept	0.940***	(0.114)	0.988***	(0.116)	0.979***	(0.116)
Fixed effect	Zip code		Zip code		Zip code
Seasonality	City–month		City–month		City–month
Observations	73,754		73,754		73,754
R2	0.6896		0.6894		0.6893

NOTE.—DV: logged number of reservation days in a month.

*

p < .05,

**

p < .01,

***

p < .001.

The results in column (2) confirmed hypothesis 4b and highlight that a host’s smile improved demand more for properties located in areas with higher crime rates compared to lower crime rates (coefficient of Smile × Crime Hours was negative and significant), which are inherently less certain. Moving to column (3), we found that the effect of a host’s smile was greater when free parking on the property was not provided (coefficient Smile × Parking was negative and significant). The additional increase in demand from a host’s smile for properties without a parking was approximately 3.8% (=1 – exp^(–0.0388)). Similar to crime rates in an area, access to parking serves as a proxy for perceived safety, which may directly influence judgments of uncertainty regarding the stay. This finding supports our hypothesis 4c that if the accommodation was perceived as lower in quality-related certainty, then a smile in the profile picture may be more effective in mitigating the lack of certainty.

The empirical analysis in study 1 demonstrates a potential causal effect of a host’s smile on Airbnb property demand in the real world, as well as a differential effect based on a variety of host and property characteristics, including the host’s gender. Although we matched properties across the smiling conditions on a rich set of observed characteristics, the properties might vary on unobserved characteristics. In addition, the statistical controls may have been insufficient for establishing a full causal claim, due to potential misspecification of the assumed linear functional form on property demand. Furthermore, while the moderation analyses in study 1 offer evidence consistent with an underlying uncertainty process, they do not provide direct evidence of why smiling affects property demand. To address these issues, we complement our secondary data analysis with a series of controlled experiments.

STUDY 2: A TEST OF WHETHER UNCERTAINTY VARIES AMONG THE MODERATORS

We begin by providing further support for our theoretical account that the effect of a host’s smile on demand will be greatest when there is uncertainty about the stay, either regarding the quality of the accommodation or the interaction with the host, and offer preliminary evidence for hypothesis 2. To do so, we tested whether the moderators that we identified in study 1 do indeed vary in the degree of uncertainty in the accommodation quality and/or host interaction. Study 2 was pre-registered at https://aspredicted.org/L2N_XWW.

Method

We recruited 1,600 participants on Prolific, and 1,602 finished the survey (45% male, ages 18–90, median age = 37). We tested four of the moderators identified in study 1, and for each moderator, we tested two conditions: host gender (male vs. female), accommodation type (shared vs. private), host experience (low vs. high), and neighborhood crime rate (high vs. low). Thus, each participant was assigned to one of eight conditions in a fully between-subjects design.

All participants read, “Suppose you are looking for an AirBnB for a trip you’re planning.” For each of the moderators, the next sentence varied as follows:

• Host gender: “You come across a property run by the following host.”

• Accommodation type: “You come across a private (room in a shared/entire) apartment run by the following host.”

• Host experience: “You come across a property run by the following host, who joined the platform (a few weeks ago/several years ago).”

• Neighborhood crime rate: You come across a property run by the following host, which is located in a neighborhood with a (A/C) crime rating.*

*ratings are on a scale from A to F, where A is the safest.

Beneath the text was a property listing with a photo of a room, a price ($110/night), and a photo of a host named Sam. In all conditions, except the female host gender condition, participants viewed a picture of a non-smiling male host; in the female host gender condition, participants viewed a picture of a non-smiling female host, who was pre-tested to be similarly attractive to the male host (web appendix sections 2 and 4).

There were two pages with our primary dependent measures of interest: one page had three questions regarding the quality accommodations, and the other page had three questions regarding the interaction with the host. The questions were presented beneath the scenario and property ad, and the order of the pages was counterbalanced. On one page, participants read, “Please answer the following questions regarding the quality of the accommodations (e.g., the cleanliness, comfort, amenities, etc. of the place itself),” and responded to the following questions: “How certain are you that the quality of the accommodations will be positive?” (1 = not at all certain, 9 = very certain); “How confident are you regarding the quality of the accommodations?” (1 = not at all confident, 9 = very confident); and “How worried are you that the quality of the accommodations will not meet your expectations?” (1 = not at all worried, 9 = very worried). On the other page, participants read, “Please answer the following questions regarding the interaction with the host (e.g., the friendliness, responsiveness, trustworthiness, etc.)” and responded to the same three questions, except “quality of the accommodations” was replaced with “interaction with the host” for each.

Results

Following our preregistered analysis plan, we reverse scored both “worry” questions and averaged the three measures to form an overall accommodation-quality certainty score (α = 0.87) and host-interaction certainty score (α = 0.85). We next conducted pairwise comparisons for each of the moderators (see table 6 for all means and SDs by condition).

TABLE 6

ACCOMMODATION QUALITY AND HOST INTERACTION CERTAINTY BY CONDITION (STUDY 2)

	M (SD)
Moderator: Gender	Male	Female	F test (1, 1594)	Sig.	Effect size
Accommodations	5.61 (1.53)	5.97 (1.43)	5.61	0.018	d = 0.23
Host interaction	4.72 (1.51)	5.31 (1.50)	15.05	<0.001	d = 0.38
Moderator: Accommodation type	Shared	Private	Test	Sig.	Effect size
Accommodations	5.20 (1.45)	5.56 (1.56)	6.05	0.014	d = 0.24
Host interaction	4.32 (1.60)	4.74 (1.53)	7.32	0.014	d = 0.26
Moderator: Host experience	Low	High	Test	Sig.	Effect size
Accommodations	5.17 (1.64)	5.89 (1.35)	22.93	<0.001	d = 0.46
Host interaction	4.55 (1.51)	5.10 (1.41)	12.86	<0.001	d = 0.35
Moderator: Crime rate	High	Low	Test	Sig.	Effect size
Accommodations	5.29 (1.35)	6.27 (1.39)	41.99	<0.001	d = 0.63
Host interaction	4.88 (1.52)	5.51 (1.51)	17.26	<0.001	d = 0.41

For all moderators, we found significant differences in uncertainty regarding both accommodation quality and host interaction between conditions. Each moderator affected the two types of perceived uncertainty to a different degree; for host gender, the effect size was larger for the host interaction, while for host experience and neighborhood crime rate, the effect size was larger for the accommodation quality; the effect sizes were roughly equivalent for accommodation type.

STUDY 3: EXPERIMENTAL EVIDENCE FOR THE EFFECT OF A HOST’S SMILE ON AIRBNB PROPERTY DEMAND, UNCERTAINTY, AND HOST PERCEPTIONS

Study 2 provided support that, at baseline when the host is not smiling, the moderators identified in study 1 do indeed vary in the degree of uncertainty regarding either the quality of the accommodations, the interaction with the host, or both. We propose that when uncertainty is greater, a host’s smile has a bigger effect on Airbnb demand. Next, study 3 further tests both hypotheses 1 and 2. We seek to test whether a host’s smile affects perceptions of the host, which increases the certainty associated with the stay and, in turn, increases the likelihood of booking the property.

Whereas the econometric dataset relied on the number of days booked as the primary dependent measure for aggregate Airbnb demand, study 3 focuses on a single individual-level decision and therefore uses the likelihood of booking the property as a proxy for demand. Testing the effect on demand in an experimental setting allows us to hold constant potential confounding variables such as property characteristics and host identity to isolate the role of the host’s smile. Further, we measure the uncertainty associated with the accommodations and host, as well as the perception of the host’s warmth, competence, and trustworthiness, to provide support for our underlying process. We propose that a host’s smile creates a halo effect, which increases overall evaluations of the host. These improved perceptions decrease uncertainty, particularly regarding the interaction with the host, which increase the likelihood of booking the property. Study 3 was pre-registered at https://aspredicted.org/LM5_827.

Method

We recruited 400 participants on Prolific, and 401 finished the survey (39% male, ages 19–76, median age = 37). All participants read, “Suppose you are looking for an AirBnB for a trip you’re planning. You come across the following listing for a private apartment in the city where you’re planning to travel. Please carefully review the listing and decide how likely you are to stay at this property.” We constructed property listing photos that mimicked a real Airbnb listing, which contained a photo of a room, a price ($110/night), a property rating (4.86 stars in 79 reviews), amenities, and most importantly, the photo of a host named Sam (figure 3). Participants were randomly assigned to one of two conditions: host smile versus no host smile, which only differed in whether the host was smiling in the profile picture. The photos of the hosts and the properties were pre-tested on their attractiveness. Please refer to web appendix for details on the construction of both the listing page and the host photo, which applies to studies 3–5. At the bottom of the page, participants were asked, “In this scenario, how likely would you be to stay at this AirBnB property?” (1 = “not at all likely,” 9 = “extremely likely”).

FIGURE 3

AIRBNB PROPERTY LISTINGS (STUDY 3)

On the next two pages in counterbalanced order, participants answered questions about their uncertainty regarding (a) the quality of the accommodations and (b) interaction with the host. Participants were shown the same Airbnb listing image and description as on the first page, and were told to, “Please answer a few more questions about this property and scenario.” Participants answered the same three questions as in study 2 (certainty, confidence, and worry) about each form of uncertainty.

On the following page, participants viewed a picture of the host, smiling or not smiling, and answered nine questions about their perceptions of the host. Each question used the format, “How ___ do you think this person is?” (1 = not at all ____, 7 = very ___), which we adapted from Aaker et al. (2010) and Wang et al. (2017). Participants responded to three warmth measures (warm, friendly, kind), three competence measures (competent, capable, effective), and three trust measures (trustworthy, honest, ethical).

Results

Likelihood of booking

We began by analyzing the effect of a host’s smile on the likelihood of booking the property. As predicted, a host’s smile significantly increased the likelihood of booking the property (M_SMILE = 6.08, SD = 1.83 vs. M_NO-SMILE = 5.63, SD = 1.92, t(398) = 2.40, p = .017, d = 0.24).

Analysis by participant gender

Although not part of our preregistered analysis plan, we next explored whether the effect of smiling on demand may differ depending on the gender of the participant. It is possible that women may feel more threatened by male hosts, for example, and therefore the effect of smiling may be driven primarily by women. On the other hand, men often feel competitive with other men, so smiling may be effective at mitigating competitiveness for male guests/participants. Although we could not explore this possibility in study 1, as the Airbnb field dataset did not contain the gender of the guests, we were able to examine the effect of participant gender in this study. We conducted a two-way ANOVA, with smiling condition and participant gender as the between-subjects factors. However, the main effect of participant gender was not significant (F(1, 396) = 5.92, p = .186), and more importantly, nor was its interaction with the smiling condition (F(1, 396) = 0.35, p = .556).

Accommodation-quality and host-interaction uncertainty

Next, we analyzed the effect of a host’s smile on uncertainty. First, we averaged the certainty, confidence, and reversed-scored worry measure for both accommodation-quality (α = 0.83) and host-interaction (α = 0.87) uncertainty to create composite measures. Smiling significantly increased certainty regarding the host-interaction (M_SMILE = 6.03, SD = 1.47 vs. M_NO-SMILE = 4.86, SD = 1.62, t(398) = 7.62, p < .001, d = 0.76), but only directionally (non-significantly) increased certainty regarding the accommodation-quality (M_SMILE = 6.31, SD = 1.34 vs. M_NO-SMILE = 6.11, SD = 1.58, t(398) = 1.39, p = .166, d = 0.14). These results suggest that smiling has a bigger effect on attenuating uncertainty regarding the interaction with the host than regarding the quality of the accommodations. However, at scale, a directional effect on accommodation-quality may also have important effects on demand.

Perceptions of the host

We next ran a factor analysis on the nine host perception measures, which revealed two distinct factors (see web appendix for details). The three warmth and three trust measures loaded on one factor (α = 0.95), and the three competence measures loaded on a second factor (α = 0.92), so we averaged the measures accordingly to create two composite measures. The correlation between the competence and warmth measures was 0.70 (p < .001). Smiling increased overall perceptions of warmth/trust (M_SMILE = 4.86, SD = 0.93 vs. M_NO-SMILE = 3.37, SD = 1.12, t(398) = 14.50, p < .001, d = 1.45) and competence (M_SMILE = 4.99, SD = 0.95 vs. M_NO-SMILE = 4.29, SD = 1.17, t(398) = 6.63, p < .001, d = 0.66).

Mediation analysis

Finally, we tested whether uncertainty and host perceptions mediate the effect of a host’s smile on increasing the likelihood of booking. Our theoretical account is that smiling improves perceptions of the host, which increases certainty regarding the stay and, in turn, increases demand. Because the effect on uncertainty was biggest for uncertainty regarding the interaction with the host (vs. accommodation quality uncertainty), as well as for warmth perceptions (vs. competence perceptions), we focused on those in our mediation analyses. Our primary analysis was thus a serial mediation analysis. Specifically, we used SPSS Process Macro Model 6 with 5,000 resamples (see figure 4 for the conceptual model; Preacher and Hayes 2008). We coded the smile variable as 0 = no, 1 = yes.

FIGURE 4

PROPOSED CONCEPTUAL MODEL

As predicted, the serial mediation path was significant (b = 0.70, SE = 0.14, 95% CI [0.44, 0.96]). Specifically, we find a significant effect of host smile on perceptions of warmth (b = 1.49, SE = 0.10, p < .001), a significant effect of warmth on host-interaction certainty controlling for smiling (b = 1.00, SE = 0.06, p < .001), and a significant effect of host-interaction certainty on likelihood of booking, controlling for smiling and warmth (b = 0.47, SE = 0.07, p < .001). The remaining effect of host smile reversed, suggesting that beyond increasing perceptions of warmth, a host’s smile might have a negative effect (b = –0.53, SE = 0.20, p < .001). Alternate mediation models were significant as well, including using competence or an overall host perception measure (the average of all nine host perceptions measures) in place of warmth, the quality-accommodation certainty measure in place of the host-interaction uncertainty measure in the serial mediation, the host perception measures on their own (not as a serial mediation), or the host-interaction uncertainty measure on its own (web appendix).

Our analysis thus provides evidence that smiling creates a positive halo effect that increases perceptions of the host, increases certainty regarding the stay (and the interaction with the host in particular), which increases the likelihood of booking the property.

STUDY 4: EXPERIMENTAL EVIDENCE FOR MODERATION BY HOST GENDER

Study 3 provided evidence that smiling increases the likelihood of booking in an experimental setting that controlled for all host and property characteristics, as well as provided evidence for the underlying role of improving host perceptions and certainty regarding the stay. Because the primary focus of our article is the moderators to the effect of smiling, we next test one of our key moderators, the host’s gender, in an experiment. That is, study 4 covers hypotheses 2 and 3a. We focused on the host’s gender due to its fundamental and generalizable nature (i.e., it is not specific to Airbnb), as well as its strong moderating role in the Airbnb dataset. To test for causality and moderation by gender, we only varied the gender of the host and the presence or absence of a smile in the host’s profile photo. We also further explored the role of improving host perceptions in driving the effect of smiling on demand, and the moderation by gender. Finally, participants in study 4 evaluated two properties, rather than one. See figure 5 for the stimuli that participants saw and figure 6 for the host profile pictures in this study.

FIGURE 5

AIRBNB PROPERTY LISTINGS IN STUDY 4

FIGURE 6

HOST PROFILE PICTURES IN STUDY 4

Method

We recruited 800 participants from Amazon Mechanical Turk (MTurk). Due to issues with poor data quality at the time the survey was run (July 2020; Chmielewski and Kucker 2020), we included a series of English screeners and attention checks throughout the survey (for full details, see the web appendix), leaving 519 participants (57% male, ages 18–75, median age = 34) who passed all checks. We analyzed only the data of those who passed all checks.

Participants were assigned to one of four host profile picture conditions in a 2 (gender: male vs. female) × 2 (smile: yes vs. no) between-subjects design. After completing an initial captcha and English screener, all participants read that they were planning a trip and researching Airbnb properties. Because this study was conducted amid the Covid-19 pandemic, and concerns about staying in an Airbnb during this time could depress results, we instructed participants to imagine that the pandemic had passed, and that risk of infection was no longer a concern.

Specifically, for each property, participants read, “Suppose the Covid-19 pandemic has finally passed, and you are looking for an Airbnb for a trip you’re planning. The risks of the virus spreading are no longer a concern. You come across the following listing for a place in the city where you’re planning to travel. Please carefully review the listing and decide how likely you are to stay at this property,” and reviewed a picture of the property listing. Each participant reviewed the scenario for two properties in a counterbalanced order (and were in the same condition for both properties). For each property, participants viewed a photo of either a male or female host, which were pretested to be similarly attractive. We kept the information related to price, reviews, amenities, and capacity similar between the two properties to minimize the likelihood that comparing those attributes across properties would overwhelm any effects related to the host’s picture.

At the bottom of the page, they responded to our primary DV of interest: “In this scenario, how likely would you be to stay at this AirBnB property?” on a nine-point scale anchored 1 = “not at all likely” and 9 = “extremely likely.” On the following page, participants were asked to, “Please briefly explain what went through your mind as you made your decision,” before repeating the process for the second property listing.

Next, participants responded to follow-up questions regarding which pieces of information they considered during their decisions (with an attention check included in the list of information factors), their past Airbnb usage, and whether their responses were affected by Covid-19. On the following pages, they were shown a picture of the host of both properties they previously evaluated (one host per page with the order randomized), and they rated each host on their perceived warmth and competence, using the same measures as in study 3. Because the trust measures loaded onto the same factor as the warmth measures in study 3, we only included the three warmth and three competence measures in study 4.

Results

Likelihood of booking

We began by analyzing the effect of a host’s smile by gender on the likelihood of booking the property. We ran a repeated measures ANOVA with the property/host evaluated as the within-subjects factor (because all participants evaluated both properties/pairs of hosts), and host smile and gender as two between-subjects factors. Because there was no main effect of the property/host (F(1, 514) = 0.01, p > .9), nor was any interaction with the other factors significant (all ps > .2), we report all results as the collapsed means across the properties. For a breakdown of the individual property means, see figure 7.

FIGURE 7

EFFECT OF A HOST’S SMILE ON LIKELIHOOD OF BOOKING BY GENDER

NOTE.—Error bars represent 95% confidence intervals.

Consistent with the small but positive effect detected in our econometric analysis, the main effect of whether the host was smiling in their profile picture on the likelihood of booking was directional but not significant (M_SMILE = 6.40, SD = 1.93 vs. M_NO-SMILE = 6.10, SD = 2.13, F(1, 514) = 2.46, p = .112, d = 0.15). More important to our hypothesis that the effect of smiling on likelihood of booking is moderated by gender, we find a significant host smile by gender interaction (F(1, 514) = 4.85, p = .028). Pairwise analyses with Bonferroni correction confirm the pattern detected in study 1, that smiling increases the likelihood of booking more for male hosts than for female hosts. Among male hosts, the effect of smiling on the likelihood of booking was significant (M_SMILE = 6.54, SD = 1.66 vs. M_NO-SMILE = 5.92, SD = 1.90, p = .008, d = 0.33). Among female hosts, on the other hand, the effect of smiling on likelihood of booking was not significant (M_SMILE = 6.24, SD = 1.96 vs. M_NO-SMILE = 6.35, SD = 1.93, p > .6, d = –0.06).

Analysis by participant gender

Next, we explored whether the effect of smiling on demand may differ depending on the gender of the participant. We conducted a repeated measures ANOVA, with property/host as the within-subjects factor and host gender, smiling condition, and participant gender as the between-subjects factors. However, the main effect of participant gender was not significant, nor was its interaction with the host gender, the smiling condition, or the three-way interaction. Full means and statistics are shown in the web appendix.

Host perceptions

We next conducted factor analyses on the host perception ratings for both hosts. Whereas in study 3, the three competence measures loaded onto a separate factor, in this study, the three warmth and three competence measures loaded onto the same factor for both hosts (α = 0.92). We averaged all six measures to form an overall host perception composite. We ran a repeated measures ANOVA with the property’s host as the within-subjects factor (because all participants evaluated hosts for both properties), and host smile and gender as two between-subjects factors. For ease of exposition, we collapsed the results across the two hosts; see figure 8 for full means by property.

FIGURE 8

EFFECT OF HOST SMILE ON OVERALL HOST PERCEPTIONS BY GENDER

NOTE.—Error bars represent 95% confidence intervals.

As expected, we found a main effect of gender (F(1, 515) = 43.80, p < .001) and a main effect of smiling (F(1, 515) = 141.60, p < .001). More important to our hypothesis, we found a significant gender-by-smile interaction (F(1, 515) = 5.08, p = .025). Although smiling increased perceptions of the female hosts (M_SMILE = 5.66, SD = 0.88 vs. M_NO-SMILE = 4.84, SD = 0.91, F(1, 512) = 44.27, p < .001, d = 0.66) and male hosts (M_SMILE = 5.29, SD = 0.99 vs. M_NO-SMILE = 4.08, SD = 1.09, F(1, 512) = 99.47, p < .001, d = 0.94), the extent of the increase was significantly larger for the male hosts. Also, as expected, the baseline perception in the no smiling conditions was significantly lower for the male hosts than for the female hosts (F(1, 515) = 29.30, p < .001, d = 0.59). This is consistent with our theoretical framework suggesting that there is greater uncertainty interacting with male hosts at baseline, and a resulting greater increase from smiling.⁷

Moderated mediation analysis

Finally, we tested whether host perceptions mediate the effect of a host’s smile on increasing the likelihood of booking. Because the effect of a host’s smile is moderated by gender, we conducted a moderated mediation analysis to explore the underlying mechanism. Specifically, we averaged the likelihood of booking measures across properties, as well as the composite host measure, and used SPSS Process Macro Model 7 with 5,000 resamples (see figure 4 for the conceptual model; Preacher and Hayes 2008). We coded the variables smile (0 = no, 1 = yes) and gender (1 = female, 2 = male).

As predicted, the moderated mediation path was significant (b = 0.32, SE = 0.15, 95% CI [0.04, 0.63]). Specifically, we find a directional effect of host smile on host perceptions (b = 0.44, SE = 0.27, p = 0.104), and a significant host smile by gender interaction (b = 0.39, SE = 0.17, p = .025). Host perceptions in turn increased the likelihood of booking the property, controlling for host smile (b = 0.84, SE = 0.07, p < .001). The remaining effect of host smile reversed, suggesting that beyond improving perceptions of the host, a host’s smile might have a negative effect (b = –0.62, SE = 0.17, p <.001). Examining the indirect paths separately by gender, we find significant indirect effects of warmth for both the female hosts (b = 0.69, SE = 0.10, 95% CI [0.50, 0.90]) and male hosts (b = 1.02, SE = 0.14, 95% CI [0.75, 1.31]).

Our analysis thus provides evidence that smiling improves perceptions of the host, to a greater extent for males than for females, which in turn increases the likelihood of booking the host’s Airbnb property. This is consistent with our finding that there is greater uncertainty regarding the interaction with the host for male hosts than females.⁸ Whereas this study focused on measuring the sociability (rather than morality) dimension of warmth, as well as measures of competence, we found all measures to be highly correlated, consistent with smiling creating a general halo effect in this context.

Overall, with studies 2–4, we confirm the causal role of a host’s smile increasing the likelihood of booking, because the property and host (within each gender) did not vary between subjects, only the smile did. In particular, smiling significantly increased the likelihood of booking an Airbnb property for male hosts, but not for female hosts.

However, one question that remains is whether the effect of smiling on demand generalizes beyond Airbnb. While the primary contribution of the present article rests on garnering insights from a large real-world dataset, the question arises about whether the effects and the process are unique to the specific context we have studied. Airbnb properties have the unusual characteristic of sharing a stranger’s home, so the effects of smiling may be particularly pronounced in this domain. However, because our theory relies on the role of uncertainty and improving perceptions of the host underlying the increase in demand, we would expect smiling to increase demand for other hospitality contexts, even those that do not require sharing one’s property. Many hospitality websites now have offerings for both large hotel chains, boutique hotels, and home shares. Therefore, in study 5, we test whether a smile on an online listing increases demand for a family-owned boutique hotel and a large, branded hotel chain.

STUDY 5: TESTING THE EFFECT OF A SMILE FOR A FAMILY-OWNED BOUTIQUE HOTEL VERSUS A BRANDED HOTEL CHAIN

Study 5 tests whether smiling increases demand for a family-owned boutique hotel versus a branded hotel chain. We varied whether the host’s picture represented a co-owner of a family boutique hotel or a manager of a branded chain hotel, as well as whether he was smiling or not in the online listing photo. Since the concerns about uncertainty regarding both the quality of the accommodations and the interaction with the staff would remain in other areas of the hospitality industry, we expect that smiling will increase the likelihood of booking for both types of hotels. However, similar to a smile, a well-known brand name serves as a non-informational cue that could reduce perceptions of uncertainty, compared to a lesser-known family-owned hotel. We thus predict the effect of smiling will be smaller for the branded hotel than for the family-owned boutique hotel. Study 5 was pre-registered at https://aspredicted.org/ZZS_DMK.

Method

We recruited 801 participants on Prolific, and 799 participants (49% male, 2% other, ages 18–85, median age = 34) passed the attention checks and completed the study. Participants were assigned to one of four conditions in a 2 (hotel: family-owned vs. branded) × 2 (smiling: yes vs. no) between-subjects design. All participants read: “Suppose you’re planning a trip and looking for a hotel in the area. You come across the following listing for a (small family-owned/Hilton) hotel. Please carefully review the listing and decide how likely you are to stay at this property.” Beneath the text was a picture of a “King Suite with Mountain View” room at either the “Benton Family Inn” or “Hilton Hotels & Resorts” along with a picture of the (co-owner/branch manager) Sam Benton and the price (figure 9) for stimuli. We used the same pictures as the male host from pair 2 in study 4, and participants evaluated only one property. At the bottom of the page, participants responded to our primary dependent measure, “In this scenario, how likely would you be to stay at this hotel?” on a nine-point scale from “not at all likely” to “extremely likely.” On the following page, participants answered demographic questions.

FIGURE 9

PROPERTY LISTINGS IN STUDY 5

Results and Discussion

A two-way ANOVA with the host smile and hotel as two between-subjects factors revealed a main effect of a host’s smile (M_SMILE = 5.86, SD = 1.97 vs. M_NO-SMILE = 4.85, SD = 2.05, F(1, 795) = 53.1, p < .001, d = 0.49), and a main effect of hotel, with a higher likelihood of booking the Hilton (M_HILTON = 5.53, SD = 2.02 vs. M_FAMILY-OWNED = 5.18, SD = 2.12, F(1, 795) = 7.47, p = .006, d = 0.17). The hotel × smile interaction was not significant (F(1, 795) = 0.45, p = .501). In line with our pre-registered analysis plan, we examined the pairwise comparisons with Bonferroni correction. Smiling increased the likelihood of booking for both the family-owned (M_SMILE = 5.72, SD = 2.00 vs. M_NO-SMILE = 4.59, SD = 2.09, p < .001, d = 0.55), and the Hilton hotel (M_SMILE = 6.02, SD = 1.93 vs. M_NO-SMILE = 5.08, SD = 2.00, p < .001, d = 0.45)—the effect size for the family-owned hotel was directionally but non-significantly larger.

Study 5 replicates the effect of a smile on the likelihood of booking in a hospitality context beyond Airbnb, suggesting that the results are not limited to any idiosyncratic features of Airbnb, but rather they extend to broader hospitality contexts. Our theory predicts that smiling would increase demand more for the family-owned hotel than for the branded Hilton hotel, since the family-owned hotel is unknown and thus less certain, but the interaction was not significant.

GENERAL DISCUSSION

The present work examined the effect of a smile in the host’s profile photos on Airbnb property demand. We leveraged deep learning models to automatically detect the presence of a smile in host profile photos and used that to estimate the effect of a host’s smile on Airbnb property demand, as well as identified a variety of moderators to the effect. We conducted controlled follow-up experiments to offer further evidence of the causality of the effect of a host’s smile on the likelihood of booking an Airbnb property and explored the mechanism underlying this effect. We found that a host’s smile increases perceptions of the hosts, which increases certainty and, in turn, increases the likelihood of booking the property.

To the best of our knowledge, the present work is the first to show both computational and experimental evidence for how smiling affects business outcomes in a sharing economy setting. We contribute to the literature on the effect of non-informational cues in the growing sharing economy by building and testing consumer theories using big data and image analytics, as well as identifying moderators that affect the degree to which a smile influences demand. We also contribute to the literature on gender differences in interpersonal interactions by demonstrating that smiling may be more beneficial for men than women due to greater uncertainty surrounding interactions with men, which can, in turn, have real economic and societal consequences.

We further test for causality and explore the mechanism highlighting the role of uncertainty underlying the effect of a host’s smile, which lays the groundwork for future studies. Economics-based research often overlooks non-informational cues. However, with machine learning techniques becoming more prevalent, quantifying these cues has become more straightforward; ignoring them in robust econometric models could lead to skewed and unreliable outcomes. Moreover, rather than merely identifying the existence of an effect, we quantify the economic impact in a real-world setting. We use behavioral insights to test various moderators at scale that we would otherwise be unable to effectively and naturalistically vary in the lab, including local crime rates and the host’s experience, and provide corroborating evidence for the role of uncertainty driving the effect of smiling.

In addition, our research has managerial relevance, with direct implications for practitioners in the hospitality industry, as well as in the service industry more broadly. Smiling is widely used to improve customer impressions in sales and in-person settings, to enhance consumers’ consumption experiences and customer relationships, and to improve outcomes in advertisements (Petroshius and Crocker 1989). Our research extends this understanding to online settings by offering insight into when a smile in a profile picture is most effective, helping practitioners optimize their online visual presentation and increase customer engagement. Consistent with recent work in the crowdfunding space, we find similar effects in peer-to-peer lodging settings; in particular, consumers (Airbnb guests) also form interpersonal trait perceptions about the sellers (Airbnb hosts) based on their face photos, which in turn affects interest in the provided product or service. We therefore highlight the critical role of non-informational (visual) cues in digital marketing. Furthermore, our work offers key guidance about when the effect of smiling is likely to be greatest—in particular when uncertainty surrounding the service (either the quality of the offering or the interaction with the service providers) is greatest. Our findings thus contribute to the person-perception literature and inform broader marketing strategies about how and when a well-placed smile can best reduce transaction frictions.

Future research can extend this investigation to explore the effects of smiling in other online service domains. We find that when uncertainty is lower due to quality or interaction concerns, the effect of smiling is greater, although our studies suggest that the uncertainty related to interpersonal interaction may be the more important driver. This distinction may also explain why past research did not always find a universally positive effect of smiling in the crowdfunding space, where there is no expectation of interacting with the creator. An open question remains, however, in other domains where interaction is expected. For example, employees, such as attorneys in a law firm or physicians in a medical service, often display their individual photos on websites, as do prospective romantic partners on online dating sites. These photos often serve as an entry point in a communication channel, so a smile may be a way to reduce uncertainty regarding the anticipated service or interaction. However, these contexts differ from the sharing economy in the extent to which the interpersonal interaction (vs. other attributes) matters in the evaluation, which may lead to different effects of smiling. Further, individuals might have different preferences in each domain. In legal settings, for example, preferences regarding the nature of the interaction might be more heterogeneous than in the sharing economy, so the effect of smiling might instead be moderated by the preferences of the chooser rather than the characteristics of the expresser.

The present work explored the role of smiling versus not smiling on the perceptions of the host and the uncertainty surrounding the service, and the resulting impact on choice. Future work can take a more nuanced look at this question. Past work has highlighted that the effects of smiling may be moderated by how broad or narrow the smile is (Wang et al. 2017), or by how genuine or authentic it is perceived to be (Tsai and Huang 2002). Because the current facial attribute classifier only allowed us to classify smiles as a binary response, we were not able to explore how demand may be affected depending on the type of smile. When data become available that enable optimizing a machine learning classifier to measure the extent or authenticity of smiles from a large set of profile pictures, future research can investigate these questions at scale. For example, Wang et al. (2017) found that warmth and competence perceptions differed depending on whether the smile was broad versus narrow, but these results were in the context of a more cut-throat crowd-funding context, where lower warmth individuals may be seen as more competent, and where there is no expectation of interacting with the creator. In a hospitality context, where warmth and competence are highly correlated, Wang et al.’s (2017) results may not hold.

Moreover, although our work identified uncertainty as a key moderator to the effect of smiling on demand, it is likely not the only moderator at play. Future work can build on our findings to identify other conditions that lead smiles to have a great effect. In our econometric analysis, we identified one variable that did not fit the pattern of greater uncertainty leading to a greater effect of smiling: age. We found that a host’s smile has a greater effect for older hosts than younger hosts, which does not correspond to greater uncertainty perceptions. Older people are generally seen as warmer (Cuddy and Fiske 2002; Cuddy, Norton, and Fiske 2005), so there is less uncertainty regarding prospective interactions. The fact that age moderated the effect of smiling in this direction instead thus suggests that other explanations and moderators may operate. Future research could thus explore other variables and explanations underlying the role of smiling on choice.

One potential concern about our findings is that smiling did not necessarily increase the demand, but rather, non-smiling faces signaled a negative emotion that decreased the demand. Although this possibility would not apply to the online lab studies, where the host’s photos were selected from a database of neutral faces, we nevertheless sought to explore whether the expressions of non-smiling hosts were neutral or conveyed another emotion. To do so, we randomly selected 500 images that were predicted as “non-smiling” by our machine learning classifier, and recruited 2500 MTurkers to rate the emotional expression of the image (five raters per image) as one of the following categories: Happy, Neutral, Sad, Angry, Surprised, Disgust, Fearful, or N/A (N/A if the photo does not contain a human face; full details are given in the web appendix). Using a majority vote among the five raters, we found that 476 out of the 500 images (92.5%) were seen as neutral, and another 19 images (3.8%) received two “neutral” votes and two other votes. This suggests that it is not the signal of other negative emotions in the non-smiling photos driving the effect, but rather that smiling improves perceptions of the host over a neutral expression.

Finally, while the current article explores the effect of a host’s smile on demand for a given Airbnb property, we do not model the broader hospitality landscape, nor do we differentiate whether the increased demand arises from other Airbnb properties, or from alternatives such as hotels or friends’ homes. Theoretically, if smiling increases demand for a specific Airbnb property, one conjecture may be that the dominant strategy is to use a picture with a smile. However, given that the effect of smiling was moderated by various host and property characteristics, it may not be this simple. The question of when smiling may have greater influences on demand for both Airbnb and hotels, as well as the effect on both primary demand and relative market share is quite complex. Future research could carefully model supply and demand, and characterize the impact of smiling on demand in equilibrium.

CONCLUSION

We examined the effect of a smile in the host’s profile photos on Airbnb property demand, using both a large-scale dataset and controlled experiments. In both settings, we found that a smile in the host’s profile picture increases property demand and that this effect is moderated by various host and property characteristics that influence the degree of certainty regarding the accommodations and host interaction. Smiling creates a positive halo effect for the host, which increases perceptions of both warmth and competence, thereby mitigating uncertainty and increasing demand.

DATA COLLECTION STATEMENT

The first and third authors purchased Airbnb demand data from AirDNA for the first study in fall 2017. The first author collected Airbnb host profile face photo data in May 2017. The first author analyzed these data and discussed them with the second author (study 1). The second (joint-first) author ran studies 2–5 and analyzed the data in consultation with the first author. Study 2 was conducted in August 2023 on Prolific. Study 3 was conducted in November 2023 on Prolific. Study 4 was conducted in July 2020 on Amazon Mechanical Turk. Study 5 was conducted in September 2022 on Prolific. The data are currently stored in a project directory on Dropbox and the Open Science Framework.

Footnotes

1

Our post-test in study 2 confirms that both host gender and accommodate type vary in the degree of uncertainty regarding the interaction with the host, as well as regarding the accommodation quality. The host’s gender showed a bigger effect size for host-interaction uncertainty regarding the host interaction (relative to accommodation-quality uncertainty). However, accommodation type (shared vs. private) showed a similar effect size for both types of uncertainty. We believe grouping the accommodation type variable with the host interaction moderators makes sense conceptually, however, given that greater interaction is more likely in shared arrangements than private ones.

2

We examined the prediction performance across host and property attributes and found that there was no systematic pattern in prediction errors across hosts’ demographic groups, and that the prediction errors were not correlated with the property attributes.

3

Data and code for study 1 can be found at https://www.dropbox.com/scl/fo/3cvokpkuelky13mernpf0/APPi5P7ShdQV8bXTPxn3X3I?rlkey=3o9ilo1eqcukqnbyru8u4z1sa&dl=0.

4

In our data, properties are booked for 6.3 days in a month on average. The average property nightly rate is $247.20.

5

In our sample, on average, a property was booked 6.3 days per month and the average nightly rate was $247.2. Thus, the effect of a smile on increasing property demand translates to an incremental annual revenue of $1,625 (= 6.30 booked days/month × $247.2/day × 12 months/year × 8.7%).

6

https://news.airbnb.com/amenities-do-matter-airbnb-reveals-which-amenities-guests-search-for-most/; https://www.ipx1031.com/what-guests-want-airbnb/.

7

Hack (2014) found that smiling increased perceived warmth more for female than for male faces. Two key differences to note between his study and the present study: first, Hack selected faces that had the baseline level of warmth between male and female; second, the participants first completed tasks (reading and rating stereotype and counter-stereotype statements about women and men), which could have influenced the subsequent assessment of warmth perceptions.

8

If we treat the warmth and competence variables as separate factors, we find that warmth significantly mediates the effect, but competence does not (see web appendix for analysis). Host gender showed greater variability in uncertainty surrounding interaction with the host than accommodation quality, whereas the other moderators did not. We therefore would expect both warmth and competence to mediate in other contexts, so we instead focus on overall perceptions of the host.

Author notes

Shunyuan Zhang ([email protected]) is assistant professor of business administration, Harvard Business School, Boston, MA 02163, USA.

Elizabeth M. S. Friedman ([email protected]) is assistant professor of business, Columbia University, New York City, NY 10027, USA.

Kannan Srinivasan ([email protected]) is H.J. Heinz II professor of management, marketing and business technologies, Carnegie Mellon University, Pittsburgh, PA 15213, USA.

Ravi Dhar ([email protected]) is George Rogers Clark professor of management and marketing & director of the center for customer insights, Yale University, New Haven, CT 06511, USA.

Xupin Zhang ([email protected]) is assistant professor of information management, School of Economics and Management, East China Normal University, Shanghai 200062, China.

The authors would like to thank Jennifer Savary, Olivier Toubia, Mohin Banker, Yu Ding, and Hortense Fong for their feedback on previous drafts. Supplementary materials are included in the web appendix accompanying the online version of this article.

Author notes

Shunyuan Zhang and Elizabeth M. S. Friedman authors contributed equally.