Understanding the Impact of Inconsistency on the Helpfulness of Online Reviews

Abstract

This study investigates how review inconsistency influences perceived helpfulness in online restaurant reviews both in ratings and specific aspects of service attributes. Drawing on 106,464 Yelp reviews spanning 666 restaurants, we employed aspect-based sentiment analysis and Tobit regression to capture not only rating inconsistencies but also differences in sentiment toward décor, taste, service, and price. Results indicate that rating inconsistency negatively affects review helpfulness, suggesting that highly divergent ratings reduce credibility. However, aspect inconsistency shows mixed effects. Discrepancies in décor and taste positively influence perceived helpfulness by offering novel information, whereas service-related inconsistencies diminish review helpfulness, due to heightened consumer sensitivity to possible service failures. Reviewer expertise further strengthens the negative influence of inconsistency as readers expect experienced reviewers to provide objective feedback. These findings extend current research by shifting the analytical lens from individual reviews to sets of reviews, thereby capturing the relational dynamics that shape consumers’ perceptions of review credibility. The results also highlight the importance of analyzing review content by specific aspects to uncover nuanced effects. Practically, platforms can benefit from grouping reviews by attributes and alerting users to noteworthy inconsistencies, facilitating more informed consumer decision-making.

1. Introduction

A recent Qualtrics survey revealed that 93% of customers trust online reviews as much as personal recommendations from friends [1]. Consumers often read reviews prior to trying new services to evaluate if the offerings are likely to meet their expectations [2]. This finding highlights a significant shift in consumer behavior, indicating that online customer reviews (OCRs) are increasingly replacing traditional word-of-mouth (WOM) communications. This shift underscores the importance for marketers to carefully manage online reviews to positively shape perceptions of their corporate brand and offerings.

However, the sheer volume of available reviews means that consumers often read only a handful before making a purchase decision. To mitigate users’ searching costs, platforms strive to highlight the most valuable reviews by leveraging reader voting systems that identify and promote the most helpful content.

Previous studies have examined the factors contributing to the helpfulness of a review. These include review factors such as number of reviews, star ratings and review depth [3,4], reviewer factors like credibility [5] and profile image [6], and product-type factors. However, these studies have primarily focused on singular reviews, often overlooking the dynamics among multiple reviews. Given consumers typically consult multiple reviews rather than relying on a single one to inform their decisions [7], existing research focusing on individual reviews inherently has limitations in capturing the complexity of this process. For instance, the detailed content of a review loses its value if it merely reiterates information covered in other reviews. Conversely, if a review contradicts the content of other reviews, its reliability could be compromised [7].

Given these considerations, this study extends beyond the characteristics of individual reviews and focuses more intently on the relationships among reviews. Specifically, we investigate how inconsistency between reviews influences perceived helpfulness. While initial assumptions might suggest that inconsistency primarily detracts from helpfulness, its actual impact could be more complex and context-dependent. For instance, conflicting information on objective aspects like service failures decreases helpfulness due to cognitive dissonance, whereas disagreements on subjective aspects like taste could offer valuable diverse perspectives. Building on these considerations, this study addresses the following research questions:

(1)

How does rating inconsistency affect review helpfulness?

(2)

How does aspect-based sentiment inconsistency influence review helpfulness?

(3)

Is the relationship between inconsistency and review helpfulness moderated by reviewer expertise?

To investigate these questions, this study draws upon relevant theoretical perspectives to develop its research model. Specifically, we utilize signaling theory to explain how reviews function as signals in information asymmetric environments and cognitive dissonance theory to understand reader reactions to conflicting information.

This research offers valuable insights into online review analysis by advocating for a focus on review sets, which allows for a more comprehensive understanding of consumer behavior. Moreover, this study employs aspect-based sentiment analysis to evaluate sentiments toward specific aspects, addressing the shortcomings of past studies that only considered one-dimensional analysis. This study also investigates the moderating role of reviewer expertise on review usefulness, finding that inconsistent opinions from highly credible reviewers are perceived as less helpful. Practically, this study suggests that review platforms implement features that organize reviews by specific attributes and highlight meaningful inconsistencies, thereby facilitating more informed consumer decision-making.

2. Literature Review

2.1. Signaling Theory

While economic models often assume information symmetry, the reality in most online services is one of information asymmetry, favoring sellers [8]. This imbalance presents challenges for consumers, making it difficult for them to accurately assess the value of goods or services purchased online due to limited prepurchase information.

Signaling theory, as proposed by Spence [9], addresses situations characterized by information disparities between transaction parties. Spence proposes that the party possessing more information (the signaler) can convey valuable insights or cues (signals) to the less informed party (the receiver), thereby mitigating information asymmetry. Effective signaling is characterized by high observability of the signal and significant costs associated with signaling [8].

In the e-commerce context, where direct product evaluation is hindered by physical distance, signaling theory has gained widespread application [8,10,11,12]. Signals within online reviews can guide readers, potentially helping them avoid service failures and the negative consequences stemming from inadequate information [13]. For instance, Boateng [10] explored the Ghanaian banking sector, focusing on interactivity and online trust as mediating signals in online platforms. Similarly, Wang et al. [12] investigated how emotional content in online reviews influences perceived product quality and purchase decisions, treating various emotions as signals through the pleasure–arousal–dominance model. Building on this theoretical foundation and its established relevance in e-commerce, this study utilizes signaling theory to investigate the factors influencing review helpfulness.

2.2. Online Customer Review Helpfulness

Intangible services such as travel or dining often induce greater uncertainty among consumers compared to tangible goods, complicating judgment and decision-making processes [14,15]. Therefore, customers read OCRs to aid in their decision-making. However, unlike traditional word-of-mouth (WOM), OCR lacks nonverbal communication cues like gestures, tone, and facial expressions, making it harder to ascertain credibility [11]. As the beneficial impact of OCR on business sales has become pervasive, some businesses have engaged in unethical practices like deleting negative reviews, creating fake positive ones, or even posting negative reviews to tarnish competitors’ reputations.

To address these issues, some platforms provide indicators of review helpfulness, measured by the number of “helpful votes” a review receives from users [16]. This feature is intended to help consumers sift through the vast number of reviews by emphasizing the ones considered most valuable, thereby reducing search and evaluation costs [17,18].

Previous studies have identified various signals influencing review helpfulness, including factors related to the review itself and the reviewer [5,19,20]. Review factors pertain to the content and quality of the review. Previous research indicates that review depth, indicated by the number of words used, correlates positively with perceived helpfulness [3,5,20]. Furthermore, the extremity of a review’s sentiment has a positive effect on its helpfulness [5]. Other research has examined the impact of language style on reviews, revealing that assertive language is viewed as more helpful and creates positive perceptions of the business among travelers planning trips far in the future. In contrast, a less assertive language style is often perceived as more helpful by consumers with short-term purchase intentions [15].

Reviewer factors pertain to the perceived credibility and expertise of the reviewer, often inferred from signals such as their reviewing history, platform status (e.g., badges), or follower count [3,5,21]. Signals indicating a reviewer’s credibility, such as observable indicators of their activity level or demonstrated expertise (e.g., helpful vote counts, elite status), often require significant effort for the reviewer to acquire, thus carrying a high signal cost. Consequently, reviews authored by individuals with high levels of experience or expertise are considered more credible [5].

Despite considerable research on the attributes of individual reviews and their helpfulness, less attention has been paid to the relational dynamics among reviews. Therefore, this study advocates for a shift in focus from individual review characteristics to the relationships among reviews. Specifically, we aim to explore the concept of review inconsistency and develop a research model grounded in this relational perspective.

2.3. Review Inconsistency

According to Festinger and Carlsmith [22], cognitions can be either relevant or irrelevant to each other. When two cognitions are similar, this is referred to as consonance. Conversely, when they are different, this is known as dissonance. Dissonance can create discomfort in individuals, motivating them to avoid sources that cause this dissonance [23]. The stronger the dissonance, the stronger the tendency to reduce discomfort. Online reviews, written by numerous individuals, often contain either consistent or inconsistent information [5].

Much of the prior research examining information inconsistency in online reviews has focused on discrepancies within a single review, such as conflicts between the text valence and the numerical rating [24,25,26]. These studies generally find that such internal inconsistency harms perceived credibility, thereby reducing the review helpfulness. However, consumers rarely rely on a single review. They typically consult multiple reviews before making a decision. Therefore, this study focuses on inconsistency among reviews.

A notable form of inconsistency between reviews is the discrepancy between an individual review’s rating and the overall business rating. Ratings, typically on a 1–5 scale across most review platforms, directly reflect a consumer’s satisfaction with a product. Contextually, lower ratings are often considered particularly informative due to a common positivity bias in ratings [27].

Previous research has established that higher rating inconsistencies tend to diminish a review’s usefulness. For example, Choi and Leon [5] found that greater discrepancies between the overall rating and an individual review’s rating decrease the perceived usefulness of the review. Similarly, Baek et al. [28] argued that rating inconsistencies lower a review’s credibility and, consequently, its helpfulness. Based on these studies, the following hypothesis is proposed:

H1. 

Rating inconsistency has a negative impact on review helpfulness.

While ratings are pivotal in assessing customer satisfaction, they do not incorporate the content of reviews. Consumers often show greater interest in the detailed experiences shared by others rather than in mere ratings [29]. One of the common methods used to analyze review text data is through sentiment analysis, which measures valence. Valence, representing the positive or negative tone of a consumer’s experience, has been acknowledged as significant as it is regarded as a direct representation of review content in various studies [11,30]. In particular, negative reviews are deemed more useful since the motivation to write negative reviews often stems from direct personal experiences rather than external attributions, meaning they are perceived as being more authentic [31]. Similarly, positive reviews that mention negative content are regarded as more trustworthy.

Sentiment analysis, while useful, has limitations in representing a review’s content with a single sentiment. Services often have multiple attributes, allowing people to have mixed feelings about different aspects of the same service. For example, a review stating “The food at this restaurant was delicious, but the waiter was incredibly rude” illustrates this complexity by offering positive feedback on the food and negative feedback on the service. Simplifying this review into a single sentiment score could misrepresent the reviewer’s nuanced experience.

This methodological challenge may contribute to inconsistencies in research findings regarding the helpfulness impacts of overall review valence. For instance, Yin et al. [32] suggest that negative reviews can be more helpful, indicating that reviews expressing anxiety could be more beneficial than those expressing anger, while Pentina et al. [33] found that positive messages led to higher perceptions of review helpfulness, trustworthiness, and credibility compared to negative or two-sided messages.

To overcome the limitations of traditional sentiment analysis, some researchers argue for aspect-based sentiment analysis, highlighting the necessity of conducting sentiment analysis on an aspect-by-aspect basis rather than analyzing the entire text as a whole [34,35]. Aspect-based sentiment analysis aims to identify opinions or emotions expressed about specific aspects of an entity, offering a more detailed understanding of complex customer feedback by focusing on specific aspects or features rather than the overall sentiment of the text or review [36].

For aspect-level analysis, selecting appropriate aspects based on sufficient domain knowledge is crucial, as core aspects vary across contexts. Drawing from established hospitality and service marketing literature, this study focuses on attributes consistently identified as key drivers of customer satisfaction and behavioral intentions within the restaurant domain. Specifically, prior research highlights the importance of food quality, service quality, and the physical environment [37,38,39,40,41,42], alongside price, which significantly influences customer satisfaction and choices [38,40,43,44]. Aligning with these findings, this study examines inconsistency related to four fundamental aspects of the restaurant experience, namely taste (representing food quality), service (representing service quality dimensions), décor (representing the physical environment/ambiance), and price (representing cost and value perception). Analyzing inconsistency at this aspect level enables a more nuanced investigation of its impact on review helpfulness.

However, inconsistency at this specific aspect level has received limited research attention, leaving its impact on perceived helpfulness less understood [7,25,45]. Unlike internal inconsistency, which primarily signals a lack of coherence or credibility [24,25,26], inconsistency among reviews might be interpreted through a dual lens. While significant divergence can still raise credibility concerns, it might also offer information richness by exposing readers to a wider range of opinions and experiences.

The interpretation likely depends on the nature of the information conveyed. Drawing from signaling theory, online reviews act as signals of unobservable quality or experience [11,12,46]. From the perspective of information processing theory, signals are often most effective when they are clear and isolated [47]. When consumers evaluate objective aspects of a service, such as service execution or price (e.g., timeliness, correctness of order, staff attentiveness against expected standards), they seek clear signals about a relatively verifiable reality. Inconsistency between reviews on these aspects creates contradictory signals about this reality [45,46]. Instead of providing useful richness, this clash of signals can lead to information overload and competition for the user’s limited cognitive resources [48]. Consumers struggle to determine which signal reflects the “true” quality, leading to heightened cognitive dissonance [49]. These feelings of discomfort and distrust likely cause consumers to devalue the information, reducing the overall helpfulness of reviews containing such objective inconsistencies [25]. Therefore, the following hypothesis is proposed:

H2a. 

Price inconsistency negatively influences review helpfulness.

H2b. 

Service inconsistency negatively influences review helpfulness.

However, the processing of subjective information (e.g., taste, décor) operates differently. These aspects reflect personal preferences and experiences. This type of inconsistency does not represent contradictory signals about a single objective truth but rather diverse signals often interpreted as information richness [48]. This type of richness, focused on understanding possibilities rather than resolving factual contradictions, is less likely to be perceived merely as overload. Therefore, encountering diverse signals about subjective aspects is less likely to induce strong cognitive dissonance. Instead, it can be viewed as valuable information, providing deeper insights than consistent, homogenous reviews might offer [50]. Considering the previous research, the following hypothesis is proposed:

H2c. 

Décor inconsistency positively influences review helpfulness.

H2d. 

Taste inconsistency positively influences review helpfulness.

2.4. Reviewer Expertise as a Moderator Variable

Social information processing theory suggests that in online interactions, where non-verbal cues are absent, people lean on alternative forms of information to fill the gap [51]. This means that even with the same textual information, readers’ perceptions can vary based on such additional cues. In the realm of online reviews, these cues help in evaluating a reviewer’s credibility, including their total review count, the length of their active participation, their follower count, and the number of times their reviews have been voted as helpful [3]. Reviewers with more followers and a higher review count are seen as more knowledgeable and reliable due to their consistent interaction with the platform [21]. Additionally, reviews from individuals with a significant history of helpful contributions are viewed as more credible, as these are indicators of their expertise, making them more trustworthy in the eyes of readers [52].

Expertise is defined as having sufficient knowledge about a specific subject or skill [53]. Therefore, reviewer expertise can be defined as the experience and prior knowledge the individual has for writing a review on a particular topic.

People often believe that expert reviewers deliver unbiased evaluations because their past activities demonstrate an ability to draw on wide-ranging experiences, leading to assessments that are more objective than those from less experienced individuals [5]. When a review diverges from the consensus, it raises questions about whether it is skewing the overall perception or offering overlooked insights. Due to the objectivity expected of experts, their reviews are typically seen as standards against which other evaluations are measured. However, when an expert’s review significantly diverges from the consensus, it violates this expectation of objectivity. This violation can create cognitive dissonance for the reader, potentially diminishing the perceived helpfulness of the inconsistent expert review. This violation can occur regardless of the specific attribute or the typical direction of the inconsistency’s main effect on helpfulness. Thus, we propose the following hypothesis:

H3. 

Reviewer expertise negatively moderates the negative relationship between rating inconsistency and review helpfulness.

H4a. 

Reviewer expertise negatively moderates the negative relationship between Price inconsistency and review helpfulness.

H4b. 

Reviewer expertise negatively moderates the negative relationship between Service inconsistency and review helpfulness.

H4c. 

Reviewer expertise negatively moderates the negative relationship between Décor inconsistency and review helpfulness.

H4d. 

Reviewer expertise negatively moderates the negative relationship between Taste inconsistency and review helpfulness.

3. Research Methodology

To investigate the effect of inconsistency on the helpfulness of reviews, this study analyzes data from online reviews, specifically extracting variables from Yelp’s online review datasets. The model framework for this analysis is depicted in Figure 1 below.

3.1. Data Description

This research focuses on Yelp data, specifically targeting restaurants. The reason for choosing restaurants as the study subject is that they represent experiential goods, where quality is difficult to assess prior to experience, making consumers reliant on reviews from others. Yelp is among the most renowned review platforms in the hospitality industry, attracting 130.25 million users monthly in 2022. Yelp offers a system where users can vote on the usefulness of reviews with options like “useful”, “funny”, and “cool”. Additionally, Yelp provides visibility into a reviewer’s history, including the number of fans or friends following them, offering insights into the reviewer’s expertise. Yelp’s system facilitates the measurement of inconsistency, reviewer expertise, and helpfulness, which are key variables of this study. Hence, we intend to validate our hypotheses using Yelp data, considered representative of consumer opinions due to its vast size and diversity.

Considering the recency of reviews, data from five years, 2018 to 2022, was collected. To ensure the reliability of the data analysis, only reviews from restaurants with more than 100 reviews were included. Ultimately, this resulted in the collection of 106,464 reviews from 666 restaurants. A basic statistical analysis of the used data is presented in

Table 1. Statistical analysis.

Variables	Minimum	Maximum	Mean	Standard Deviation
Dependent Variable
Review Helpfulness	0	462	2.033	6.374
Independent Variable
Rating Inconsistency	0	3.915	0.946	0.734
Décor Inconsistency (DI)	−17.1	34.717	0.286	1.751
Taste Inconsistency (TI)	−19.450	38.304	0.564	2.277
Price Inconsistency (PI)	−11.597	17.967	0.085	0.986
Service Inconsistency (SI)	−24.533	37.725	0.379	2.250
Reviewer Expertise	0	587,933	833.731	8179.763
Control Variable
Review Subjectivity	0	1	0.574	0.127
Text Length	6	5000	578.084	506.673
Business Popularity	100	1321	392.176	222.270

.

3.2. Measurements

This study intends to use the sum of three variables provided by Yelp (useful, funny, cool) as a measure of perceived helpfulness. Our analysis examines two types of inconsistencies: rating inconsistency and aspect inconsistency. Rating inconsistency is defined as the absolute difference between a review’s rating and the restaurant’s average rating, mathematically represented as follows:

$$Rating\_inconsistenc{y}_{i,j}=\left|\overline{average\_ratin{g}_i}-review\_ratin{g}_{i,j}\right|$$

(1)

where $Rating\_inconsistenc{y}_i$ represents the rating inconsistency for the $j^{th}$ review of the $i^{th}$ restaurant, $average\_ratin{g}_i$ is the average rating of the $i^{th}$ business restaurant, $review\_ratin{g}_{i,j}$ is the rating given in the $j^{th}$ review of the $i^{th}$ business restaurant.

Aspect inconsistency is assessed by calculating the absolute difference between the sentiment score of specific aspects within a review and the average sentiment score for those aspects across all reviews of the same restaurant. Building on prior research that identifies physical environment, food quality, service, and price as critical to customer satisfaction and revisit intentions in the restaurant context, this study analyzes sentiment towards these four aspects by segmenting reviews into sentences [39,46]. Given the decision to divide sentiment inconsistency into four aspects, Hypothesis 2 comprises four sub-hypotheses: price inconsistency (H2a), service inconsistency (H2b), décor inconsistency (H2c), and taste inconsistency (H2d).

Gregoriades et al. [54] highlighted the importance of carefully handling negation sentences in natural language processing. For example, a sentence like “The food was okay, but the service was terrible” contains two distinct sentiments about different aspects. Misinterpreting such sentences as a singular sentiment could lead to inaccuracies. Therefore, this study treated sentences with negation connectors as separate statements for aspect classification to capture the intended sentiment accurately.

In traditional classification tasks, the availability of a pre-trained dataset is crucial for model training. However, the advent of zero-shot learning (ZSL) methods has revolutionized this by enabling classification with minimal data. This technique’s strength lies in its ability to generalize from seen to unseen data categories without explicit prior training on the latter [55]. In the context of this research, the BERT zero-shot model was employed to perform sentence classification, demonstrating the practical application of ZSL principles.

Given that ZSL relies on this generalized semantic understanding rather than direct training on our specific aspect labels (décor, taste, price, service, irrelevant), evaluating its classification performance on our actual data was crucial to ensure the reliability of the subsequent analysis. To establish the reference labels for this evaluation, two experienced researchers manually classified a random sample of 500 sentences. The researchers reviewed the sentences jointly. While sentences with clear aspect relevance were labeled directly, any sentences deemed ambiguous or initially prompting disagreement were discussed in detail between the two researchers. The predictions of the BERT ZSL model for these 500 sentences were then compared against these reference labels. This empirical approach yielded a promising accuracy rate of 85%, showcasing the effectiveness of zero-shot learning in leveraging limited datasets for robust classification tasks.

Therefore, we measured sentiment scores for each aspect classified through zero-shot classification. To ensure robustness and mitigate potential biases associated with any single lexicon, we conducted sentiment analysis using two distinct and widely recognized lexicon-based approaches. AFINN version 0.1 and SentiWordNet 3.0 methods involve assigning scores to individual words indicating polarity and aggregating these scores to calculate the overall sentiment of a sentence or text segment pertinent to an aspect [56,57]. AFINN provides integer valence scores often tuned for informal text, while SentiWordNet assigns positivity, negativity, and objectivity scores based on WordNet synsets, offering a potentially more nuanced semantic perspective. By employing both, we aimed to verify that our substantive findings regarding inconsistency effects are robust to the choice of sentiment lexicon. The main analysis presented in this paper utilizes scores derived from AFINN, while the results using SentiWordNet are presented in the robustness check section.

Reviewer expertise was assessed based on the number of helpfulness votes received by the reviewer in the past. Regarding the control variables, subjectivity and polarity were classified using TextBlob 0.19.0. Business popularity was determined by the total number of reviews a restaurant has received.

3.3. Model

Drawing from the insights of Mudambi and Schuff [58], this study adopts a Tobit regression model to analyze the unique characteristics of helpfulness votes, such as the censored nature of the sample. This approach is particularly pertinent due to the bounded range of votes reviews can receive on Yelp, which often results in extreme outcomes in terms of helpfulness assessments. The Tobit model is adept at managing these extreme value situations.

A critical aspect of using Yelp data is that we only have access to the number of helpfulness votes a review garners, without insight into the total number of people who have practically read the review. This lack of comprehensive viewership data leads to a potential selection bias, as not every reader of a review casts a vote. In such scenarios, the Tobit model proves to be more advantageous than methods like ordinary least squares (OLS) or generalized least squares (GLS). This is because the Tobit model is designed to account for the possibility that the probability of a review receiving a vote is correlated with other explanatory variables, thus addressing potential biases that OLS and GLS might overlook [59].

By employing the Tobit regression, this study ensures a more accurate and relevant analysis of Yelp review data, particularly in understanding the dynamics of helpfulness votes, and overcomes the limitations posed by traditional regression models in the context of censored and biased online review data. Considering all variables and interaction terms, the resulting equation is as follows:

$$\begin{gathered} Review helpfulnes{s}_i={\beta }_{1}Rating Inconsistency+{\beta }_{2}Price Inconsistency+ \\ {\beta }_{3}Service Inconsistency+{\beta }_{4}Decor Inconsistency+{\beta }_{5}Taste Inconsistency+ \\ {\beta }_{6}Reviewer expertise+{\beta }_{7}Rating Inconsistency\times Reviewer expertise+ \\ {\beta }_{8}Price Inconsistency\times Reviewer expertise+{\beta }_9 Service Inconsistency\times \\ Reviewer expertise+{\beta }_{10}Decor Inconsistency\times Reviewer expertise+ \\ {\beta }_{11}Taste Inconsistency\times Reviewer expertise+\epsilon \end{gathered}$$

(2)

4. Results

Before testing the hypotheses, we checked for multicollinearity among variables. One widely used method for checking multicollinearity is VIF analysis [60]. Generally, VIF values above 5 are considered indicative of high correlation [60]. In this study, VIF ranged from 1.00 to 1.14, indicating no multicollinearity issues among variables. These findings are summarized in

Table 2. Multicollinearity.

Variables	VIF
Independent Variable
Rating Inconsistency	1.043
Price Inconsistency (PI)	1.020
Service Inconsistency (SI)	1.030
Décor Inconsistency (DI)	1.034
Taste Inconsistency (TI)	1.021
Reviewer Expertise	1.016
Control Variable
Review Subjectivity	1.065
Text Length	1.111
Business Popularity	1.002

.

The results of testing H1, which hypothesized that rating inconsistency has negative effects on review helpfulness, shows a significant effect $(\beta =-0.461,p={0.000}^{\mathrm{*}\mathrm{*}\mathrm{*}})$. Therefore, H1 is supported. Regarding the hypotheses H2a, H2b, H2c, and H2d, which aim to verify the impact of sentiment inconsistency on review helpfulness, the results indicate that DI $(\beta =1.778,p={0.003}^{\mathrm{*}\mathrm{*}})$, TI $(\beta =1.103,p={0.033}^{\mathrm{*}})$ has a significant positive influence on review helpfulness, whereas SI $(\beta =-1.723,p={0.002}^{\mathrm{*}\mathrm{*}})$ shows a significant negative impact. However, PI does not show a significant correlation with review helpfulness. Thus, H2b, H2c, and H2d are supported while H2a is not supported.

Therefore, to examine the moderating effect of reviewer expertise on the relationship between inconsistency and review helpfulness, it was hypothesized (as per H3 and H4) that higher reviewer expertise would negatively moderate this relationship. Analysis results indicate a negative moderating effect on the relationships between rating inconsistency and review helpfulness. Therefore, H3 is supported.

The analysis of the moderating effect between sentiment inconsistency and review helpfulness revealed negative moderating effects for DI $(\beta =-0.916,p={0.027}^{\mathrm{*}})$, TI $(\beta =-6.456,p={0.000}^{\mathrm{*}\mathrm{*}\mathrm{*}})$, and SI $(\beta =-4.164,p={0.000}^{\mathrm{*}\mathrm{*}\mathrm{*}})$. Therefore, hypotheses H4b, H4c, and H4d are supported. As for the control variables, review subjectivity $(\beta =-1.211,p={0.000}^{\mathrm{*}\mathrm{*}\mathrm{*}})$ and business popularity $(\beta =-2.882,p={0.000}^{\mathrm{*}\mathrm{*}\mathrm{*}})$ were found to have a negative impact on review helpfulness. In contrast, text length positively influenced review helpfulness $(\beta =22.706,p={0.000}^{\mathrm{*}\mathrm{*}\mathrm{*}})$. The results are summarized in

Table 3. Results of hypotheses test.

	Coefficient	Std.Err	Z-Value	p-Value
First order effects
Rating Inconsistency (RI)	−0.461	0.038	−12.185	***
Price Inconsistency (PI)	0.535	0.576	0.928
Service Inconsistency (SI)	−1.723	0.549	−3.138	**
Décor Inconsistency (DI)	1.778	0.602	2.956	**
Taste Inconsistency (TI)	1.103	0.517	2.133	*
Reviewer Expertise (RE)	4.803	0.049	98.458	***
Second order effects
RI × RE	−0.621	0.050	−12.529	***
PI × RE	7.352	0.610	12.051	***
SI × RE	−4.164	0.514	−8.102	***
DI × RE	−0.916	0.415	−2.206	**
TI × RE	−6.456	0.186	−34.625	***
Control variables
Review Subjectivity	−1.211	0.228	−5.302	***
Text Length	22.706	0.266	85.331	***
Business Popularity	−2.882	0.153	−18.812	***
(Intercept):1	−2.441	0.153	−15.984	***
(Intercept):2	2.045	0.004	543.072	***

Note. * p < 0.05, ** p < 0.01, *** p < 0.001.

.

To assess the robustness of our findings to the specific sentiment lexicon employed, we re-estimated the full Tobit regression model, using aspect inconsistency scores derived from the SentiWordNet lexicon instead of AFINN. The detailed results are presented in

Table 4. Sentiment wordnet results.

	Coefficient	Std.Err	Z-Value	M3
First order effects
Rating Inconsistency (RI)	−0.452	0.038	−12.032	***
Price Inconsistency (PI)	0.045	0.573	0.078
Service Inconsistency (SI)	−1.560	0.545	−2.864	**
Décor Inconsistency (DI)	1.935	0.596	3.246	**
Taste Inconsistency (TI)	1.140	0.512	2.225	*
Reviewer Expertise (RE)	5.072	0.050	101.419	***
Second order effects
RI × RC	−0.571	0.049	−11.556	***
PI × RC	12.464	0.609	20.479	***
SI × RC	−6.012	0.324	−18.529	***
DI × RC	−5.233	0.329	−15.926	***
TI × RC	−8.157	0.361	−22.604	***
Control variables
Review Subjectivity	−1.157	0.227	−5.104	***
Text Length	22.503	0.264	85.174	***
Business Popularity	−2.841	0.152	−18.686	***
(Intercept):1	−2.427	0.152	−16.003	***
(Intercept):2	2.038	0.004	541.598	***

Note. * p < 0.05, ** p < 0.01, *** p < 0.001.

. A comparison between the results using AFINN (Table 3) and SentiWordNet (Table 4) reveals substantial consistency in the main effects and interactions related to our core hypotheses. The key control variables, text length and business popularity, also show consistent effects. While the overall pattern is stable, there are minor variations in coefficient magnitudes and the significance levels of interaction terms. However, these variations do not alter the substantive conclusions drawn regarding our hypotheses. Overall, the consistent pattern of results provides strong support for the robustness of our study’s main findings.

5. Discussion

This study investigated the impact of review inconsistency on review helpfulness, distinguishing between rating inconsistency and aspect inconsistency. The findings revealed that the presence of rating inconsistency negatively affects review helpfulness, aligning with previous research suggesting that higher rating inconsistencies reduce a review’s credibility and, consequently, its usefulness [5,28].

On the other hand, the impact of aspect inconsistency on review helpfulness revealed a more complex pattern, aligning with the nuanced theoretical expectations developed earlier based on signaling theory and cognitive dissonance theory. Specifically, inconsistencies related to the subjective aspects of décor and taste were found to positively influence review helpfulness, whereas service inconsistency had a negative impact. Price inconsistency did not have a significant effect.

These results partially align with prior research indicating that sharing experiences that differ from others can lead to negative perceptions of a review’s usefulness [49]. The difference in findings could be attributed to previous studies not examining review inconsistency by aspect.

Online review helpfulness is intrinsically linked to the information’s capacity to aid consumer decision-making. Generally, information tends to increase in helpfulness when it offers novel perspectives or details that extend beyond the prevailing consensus [50]. The positive effect of inconsistency regarding décor and taste attributes observed in this study can be interpreted as reflecting this principle. These attributes possess inherently subjective characteristics, and consumers naturally anticipate a degree of opinion heterogeneity in these areas, stemming from diverse personal tastes and experiences. Therefore, conflicting opinions on décor or taste are likely to serve the function of providing valuable new perspectives, rather than being dismissed as errors.

Furthermore, when this naturally expected diversity manifests as inconsistency across reviews, it may function as an indirect signal of authenticity. A lack of complete consensus on subjective matters can increase reader confidence that the reviews represent genuine, unmanipulated customer experiences, potentially enhancing trust in the information presented.

However, this positive interpretation of disagreement is predicated on the subjective nature of the attribute in question. When evaluating aspects where a degree of objective reality or standard is assumed (e.g., service), inconsistency acquires a different meaning. Divergence in these areas is less likely to be perceived as a beneficial novelty and more likely interpreted as indicative of an error, a problem, or fundamental unreliability, consequently diminishing helpfulness.

The pattern of results in our study directly substantiates this distinction. Inconsistency related to the subjective attributes of décor and taste exerted a positive influence on helpfulness, consistent with the value derived from novel perspectives and authenticity signals. Conversely, the negative impact observed for service inconsistency supports the notion that divergence concerning more objective operational aspects is interpreted negatively, likely stemming from perceived unreliability or the signaling of potential service failures.

In service sectors like restaurants, staff rudeness or indifference can be perceived as a core service failure, universally deemed an unpleasant experience [61]. Thus, when reading reviews, individuals specifically look for indications of service failures. Unlike subjective aspects, the objective occurrence of service failures becomes crucial, potentially making inconsistent reviews less helpful. This reasoning similarly explains why price inconsistency does not significantly impact helpfulness. While individual perceptions of price vary, most reviews include both an evaluation of the price and the actual price. Readers can thus make their own judgment about price appropriateness directly. This process suggests that price inconsistency, unlike other factors, would not have a significant influence on the helpfulness of a review.

This study demonstrates that reviewer expertise negatively moderates the relationship between inconsistency and review helpfulness across attributes. This finding complements and extends prior research on reviewer credibility by demonstrating how expertise specifically interacts with review inconsistency to affect perceived helpfulness. Higher reviewer credibility implies an expectation of objectivity and representativeness from readers. While individual reviews might reflect personal experiences or judgments, a consensus within a substantial set of reviews is considered reflective of the actual characteristics of an establishment. When a credible reviewer deviates from the consensus, it can violate readers’ expectations of alignment and objectivity, thus lowering helpfulness, suggesting that credibility can amplify, rather than buffer, the impact of inconsistency in certain contexts. This interpretation is supported by the finding that higher review subjectivity negatively impacts review helpfulness.

Consistent with previous studies, text length positively influences review helpfulness [3]. Interestingly, business popularity has a negative effect on review helpfulness. This seemingly counterintuitive finding can be understood from a consumer information processing perspective. While a large number of reviews initially signals popularity and reduces uncertainty, an excessively high volume can lead to information overload [62]. Consumers faced with hundreds or thousands of reviews can find it cognitively taxing to process additional individual reviews, especially if they perceive much of the information to be redundant. The marginal utility of information from one additional review likely diminishes significantly when a large base of reviews already exists. Consequently, consumers might perceive newer individual reviews in highly reviewed businesses as less helpful compared to reviews for less popular businesses where each piece of information carries more weight. Furthermore, under conditions of information overload, consumers can shift from systematic processing of individual reviews towards relying more on heuristic cues, such as the aggregate rating or sorting by “most helpful”, further devaluing the perceived contribution of an average new review.

6. Conclusions

This study investigated the consequences of review inconsistency for the perceived helpfulness of online restaurant reviews. Utilizing a large-scale dataset from Yelp and employing aspect-based sentiment analysis, this research moves beyond analyses centered on individual review characteristics to examine the effects emerging from the relational context among reviews.

The findings demonstrate that the influence of review inconsistency on helpfulness is not uniform, but rather is contingent upon several factors. Key results indicate differential effects based on the type of inconsistency (rating deviation vs. aspect sentiment divergence), the intrinsic characteristics of the service aspect under consideration (subjective vs. objective), and the level of reviewer expertise. Collectively, this research highlights the significance of adopting a fine-grained, aspect-level analytical perspective and accounting for reviewer attributes to accurately understand consumer information processing and evaluation when faced with conflicting user-generated content. The subsequent sections elaborate on the theoretical contributions and practical implications derived from these findings.

6.1. Theoretical Implications

This study has several implications. First, it shifts the focus from individual reviews to the utilization of review sets, addressing a gap largely overlooked in previous research. As the significance of reviews gains recognition, there is an increasing concentration on the impact of individual reviews on potential buyers. The advent of natural language processing methodologies has sparked interest in understanding how the content of consumer reviews affects perceived usefulness. However, consumers typically make purchasing decisions based on sets of reviews rather than single reviews. Therefore, accurately capturing consumer behavior necessitates analyzing sets of reviews. This research extends our understanding of how consumers judge the usefulness of reviews through the inconsistency among review sets, thereby addressing limitations in existing research.

Second, this study overcomes the limitations of prior research that assumed single reviews conveyed a unified message by focusing on the multiple aspects within a single review. Previous studies utilizing review datasets often reduced the critical textual information in reviews to a single variable like variance, which overlooks information about two-sided arguments commonly found in reviews. Recent findings indicate that inconsistencies within a single review, typically manifesting as two-sided arguments, can actually enhance the perceived usefulness of the review. By employing aspect-based sentiment analysis (ABSA), this study categorizes a single review into multiple aspects and measures the reviewer’s evaluation for each. This method supplements recent studies that focused solely on the inconsistency within individual reviews.

Third, our findings complement recent studies on the positive correlation between reviewer expertise and review helpfulness, which largely focus on how reviewer expertise enhances the trustworthiness of the review content, aiding readers in their decision-making process. Generally, reviewers with higher credibility are believed to provide more objective reviews, based on their extensive experience with various restaurants. This research contributes to that discourse by suggesting that inconsistencies in reviews written by highly credible reviewers are perceived as less useful, regardless of the attribute in question. This indicates that reviewer expertise demands objectivity, and any deviation in this regard diminishes the perceived utility of the review. Thus, our study not only supplements and extends recent findings but also underscores the need for future research to focus on the attributes of reviewers.

6.2. Practical Implications

Prior to purchasing, consumers use reviews to gather information for their decision-making. This process involves screening and evaluation costs among numerous reviews, with the reduction in these costs enhancing platform utility and loyalty. Given these insights, this study offers several implications for review platforms.

Platforms can mitigate customers’ screening costs by employing text analysis techniques to categorize and present review attributes. This study found that individuals assign different evaluations based on various aspects of restaurants. Consumers delve into numerous reviews to extract information on attributes they deem significant. To diminish preliminary search costs, platforms may facilitate quicker access to necessary information by offering summaries of average reviews per aspect. This enables consumers to form a basic impression of their desired restaurants, guiding them to either decide on visiting the restaurant or to seek further review details.

Additionally, platforms can create a tab specifically for reviews with high inconsistency, aiding consumers in their decision-making process. Our research discovered that inconsistencies in subjective attributes like décor and taste are perceived by users as new and valuable information. Therefore, by offering a feature that aggregates reviews with inconsistencies in subjective attributes, platforms can assist users in making informed decisions from various perspectives.

Furthermore, this study suggests that reviewers with higher credibility should be cautious when posting inconsistent reviews. As reviewers gain experience and extend their activity period, they are perceived as more credible by others. Credible reviewers’ information becomes significantly important through this process. However, they are also expected to write objectively. The utility of inconsistent reviews across all attributes is deemed less useful as the credibility of a reviewer increases. Therefore, when posting content that differs from other reviewers, they should provide specific information to justify and support their views. This approach allows reviewers to mitigate the reduction in review usefulness caused by inconsistency.

7. Limitation and Future Research

This study offers valuable insights by utilizing real-world review data reflecting customer experiences and opinions. However, it falls short in linking these reported experiences and evaluations to their actual purchasing actions. Review data provides insight through the perspectives of the authors, influencing potential customers’ judgments. While it is possible to extract the review authors’ intentions to revisit from the content, there is no way to determine the subsequent actions of potential customers. To address this, this study employed review helpfulness as a substitute measure, but this has its drawbacks since it does not directly correspond to purchase behavior. Future research combining data analysis and survey methods to track the actual behavior of potential customers could significantly enhance our understanding of the impact of online reviews on decision-making.

This study measured inconsistency by comparing individual review ratings and aspect sentiments against the average rating and average aspect sentiments, respectively, for a given business. It acknowledges that customers do not base their decisions solely on one review or by reading all reviews. Instead, they may apply personal criteria, such as focusing on the most recent reviews, seeking out particularly negative reviews to avoid service failures, or only considering reviews from top reviewers. These customer behaviors may diverge from the assumptions of this study. Future research could include these diverse methods of information gathering to more precisely measure inconsistency, thereby potentially yielding findings that more accurately represent real customer behavior.

The findings are based on data drawn exclusively from Yelp. While Yelp is a major review platform, its user base may not fully represent all online consumers, and its specific features (e.g., 5-star scale, “Elite” status) might influence user behavior and review dynamics in ways unique to the platform. Therefore, caution is needed when generalizing the results to other platforms with potentially different user demographics or functionalities.

Finally, the study’s focus on the U.S. restaurant industry limits the direct generalizability of the findings. Restaurants are predominantly experience goods, and the observed effects of inconsistency might differ for search goods or other service categories. Cultural variations in reviewing norms and information processing also suggest that replicating the study in different geographic contexts would be valuable to ascertain the broader applicability of our conclusions.