Analyzing the Asymmetric Effects of COVID-19 on Hotel Selection Attributes and Customer Satisfaction Through AIPA

Abstract

The COVID-19 pandemic reshaped travel patterns and customer expectations, generating profound challenges for the hotel industry. This study analyzes 50,000 TripAdvisor reviews of New York hotels to examine how customer satisfaction with hotel selection attributes shifted before and during the pandemic. BERTopic was applied to extract eight key attributes, while VADER, PRCA, and Asymmetric Impact–Performance Analysis (AIPA) were used to capture asymmetric effects and prioritize improvements. Comparative analyses by hotel classification, travel type, and customer residence reveal significant shifts in food and beverage, location, and staff, particularly among lower-tier hotels, business travelers, and international guests. The novelty of this study lies in integrating BERTopic and AIPA to overcome survey-based limitations and provide a robust, data-driven view of COVID-19’s impact on hotel satisfaction. Theoretically, it advances asymmetric satisfaction research by linking text-derived attributes with AIPA. Practically, it offers actionable guidance for hotel managers to strengthen hygiene, expand contactless services, and reallocate resources effectively in preparation for future crises. In addition, this study contributes to sustainability by showing how data-driven analysis can enhance service resilience and support the long-term socio-economic viability of the hotel industry under global crises.

1. Introduction

In December 2019, the first confirmed case of COVID-19 was reported in Wuhan, Hubei Province, China, and by 2020 it had escalated into a global pandemic. Consequently, on 11 March 2020, the World Health Organization (WHO) officially declared COVID-19 a worldwide pandemic [1,2,3]. Among the industry’s most severely affected, the hotel and lodging sector experienced dramatic losses due to declining demand and a sharp drop in guest inflows [4,5]. Specifically, International tourist arrivals plummeted by more than one billion, triggering a precipitous decline in hotel occupancy rates [6,7]. In response, the hospitality industry explored a variety of recovery strategies; for example, Marriott introduced contactless services during the pandemic to ensure a safe and hygienic travel environment [8].

Within this managerial context, prior studies have examined how hotel selection attributes influence customer satisfaction during the COVID-19 pandemic. For instance, Promnil and Polnyotee [9] analyzed 386 SME hotels in Thailand and found that recovery strategies centered on customer relations and service provision were most effective in revitalizing occupancy and profitability. Bonfanti et al. [10], based on interviews with luxury hotel managers, identified seven safety-oriented measures—such as hygiene protocols, digital innovations, and staff training—that fundamentally reshaped customer experiences during the pandemic. Similarly, Kim and Han [11] demonstrated through an importance–performance analysis that the significance of cleanliness and hygiene attributes markedly increased after COVID-19, underscoring the shift in customer priorities. Complementing these operational perspectives, Cai et al. [12] applied a CASBEE–IPA framework to Japanese ryokans, highlighting that post-pandemic customers placed greater emphasis on green and sustainable practices as part of service quality. Beyond operational adjustments in the hotel context, broader service research has consistently confirmed customer satisfaction as a central determinant of behavioral outcomes. For example, Ali et al. [13] demonstrated in a theme park setting that experiential satisfaction strongly predicted revisit intention and positive word-of-mouth, while Raza et al. [14] showed that satisfaction mediates the relationship between service quality and loyalty. These findings, although from related service contexts, underscore the critical role of satisfaction as an outcome variable, thereby justifying the need for asymmetric approaches such as AIPA to evaluate hotel attributes. However, prior studies have predominantly employed survey-based data, which are resource-intensive and constrained to attributes predefined by researchers. In contrast, online travel platforms such as TripAdvisor enable customers to express their evaluations of hotels based on actual stay experiences, producing large-scale datasets enriched with contextual information including travel purpose and place of residence [15]. Prior research indicates that these contextual factors significantly moderate customer perceptions. Hotel classification, commonly divided into economy, mid-range, and luxury, serves as a proxy for service quality, with higher-tier hotels associated with elevated customer expectations [16,17,18,19]. Travel purpose, typically business versus leisure, also influences attribute prioritization [20]. Moreover, cultural and geographical differences further shape service evaluations: Zhang et al. [21] demonstrated that regional cultural contexts alter how customers assess hotel services, while Liu et al. [22], analyzing over 400,000 TripAdvisor reviews, revealed that domestic and international guests emphasize distinct satisfaction drivers. Accordingly, it is crucial to analyze variations in attribute importance and satisfaction across customer groups segmented by classification, travel type, and residence, particularly in the contrasting contexts of the pre- and post-COVID-19 periods.

To address these gaps, this study analyzes large-scale TripAdvisor reviews covering both the pre-COVID-19 period and the entire pandemic period. By applying BERTopic, a topic modeling technique, in combination with Asymmetric Impact–Performance Analysis (AIPA), the study provides a comprehensive, data-driven perspective that complements prior survey-based research. In addition, this study contributes to sustainability by showing how crisis-driven shifts in customer priorities affect not only the socio-economic resilience and long-term viability of the hospitality sector, but also its environmental and social sustainability—for example, through enhanced hygiene practices, contactless services that reduce resource use, and strategies that support both employee well-being and community trust.

Building on these gaps, the present study formulates the following research questions.

• RQ1: How did the importance and asymmetric effects of hotel selection attributes on customer satisfaction differ before and during the COVID-19 pandemic?
• RQ2: Do the asymmetric effects of hotel selection attributes on customer satisfaction differ according to hotel characteristics, specifically classification into economy, mid-range, and luxury hotels?
• RQ3: Do the asymmetric effects of hotel selection attributes on customer satisfaction differ according to customer characteristics, specifically travel purpose (business vs. leisure) and residence (local vs. foreigner)?

2. Theoretical Background

2.1. BERTopic

Topic modeling is an algorithmic approach designed to identify latent themes within a collection of documents and to cluster semantically related topics. Among the most widely adopted techniques in this domain is Latent Dirichlet Allocation (LDA) [23]. LDA has been extensively employed to extract core themes from diverse forms of textual data, including online shopping reviews, community forum posts, and news articles [24,25,26]. However, because LDA derives topics based on word co-occurrence frequencies, it falls short in adequately capturing contextual meaning and word order. In hotel research, LDA has been predominantly used; however, its reliance on word frequency makes it inadequate for short reviews that often contain multiple attributes. BERTopic, by leveraging contextual embeddings, captures these nuances more effectively and thus provides more accurate topic extraction in hotel review analysis. More recently, BERTopic has emerged as a context-aware topic modeling technique designed to overcome these limitations. Built on the BERT architecture, BERTopic incorporates contextual information of words, thereby distinguishing itself from traditional frequency-based methods [27]. Through this context-driven approach, BERTopic has demonstrated superior performance compared to conventional topic modeling techniques [27,28,29].

As illustrated in Figure 1, the BERTopic algorithm comprises three sequential stages: document embedding extraction, dimensionality reduction and clustering, and topic generation based on class-based TF-IDF (c-TF-IDF). In the first stage, pre-trained BERT language models are employed to generate document embeddings. In the second stage, the dimensionality of embeddings is reduced using Uniform Manifold Approximation and Projection (UMAP), after which semantically similar documents are clustered. While HDBSCAN is the default clustering algorithm in BERTopic, we adopt K-means for this study. HDBSCAN has been shown to classify a substantial portion of short texts as “noise,” excluding them from further analysis. de Groot et al. [30] demonstrated that replacing HDBSCAN with K-means in BERTopic significantly reduced the number of discarded documents, while maintaining comparable topic coherence and diversity. Given that TripAdvisor reviews are short, sentence-level texts, this adjustment ensures comprehensive topic coverage and preserves important customer perceptions. In the final stage, topics are derived from each cluster using c-TF-IDF. Accordingly, this study collects customer reviews from TripAdvisor through web crawling and applies the BERTopic algorithm to extract the hotel selection attributes most salient to customers.

2.2. Three-Factor Theory

The Three-factor theory posits that the impact of product or service attributes on customer satisfaction varies according to their performance level [31]. As illustrated in Figure 2, attributes are classified into three categories: exciting factors, performance factors, and basic factors, each exerting either symmetric or asymmetric effects on customer satisfaction. Exciting factors generate satisfaction when fulfilled but do not necessarily cause dissatisfaction when absent. Performance factors exhibit symmetric characteristics, producing satisfaction when fulfilled and dissatisfaction when unfulfilled. Basic factors, by contrast, do not enhance satisfaction even when fulfilled but lead to dissatisfaction if unmet. Thus, while basic and exciting factors exert asymmetric effects, performance factors influence customer satisfaction symmetrically.

In the hospitality field, the Three-Factor Theory has been widely employed to analyze customer satisfaction drivers. Xu and Li [17] applied it to hotel reviews and showed that cleanliness and staff service often function as basic factors, while leisure amenities act as excitement factors. Bi et al. [16] demonstrated that room quality and service operate as performance factors, whereas hygiene-related aspects are crucial basic factors. More recent studies have extended the framework to peer-to-peer accommodations such as Airbnb [15,25], confirming that customer perceptions of accommodation attributes can be consistently categorized into basic, performance, and excitement factors. These findings support the applicability of the Three-Factor Theory in hospitality contexts and motivate its adoption in the present study.

A representative method for analyzing the characteristics of such attributes is the Penalty-Reward Contrast Analysis (PRCA), proposed by Brandt [32]. Developed based on Kano et al. [31] theory, PRCA aims to examine the asymmetric effects of product and service attributes on customer satisfaction and to identify the key attributes that enhance satisfaction. The method generally consists of two stages. In the first stage, the performance level of each attribute is divided into high performance and low performance, generating two dummy variables. In the second stage, these dummy variables are employed as independent variables, while customer satisfaction is set as the dependent variable in a multiple regression analysis. By comparing the coefficients of high and low performance, attributes can be classified into excitement factors, performance factors, or basic factors in accordance with the Three-Factor Theory [33,34]. Specifically, this study adopts the IA (Impact Asymmetry) value, following the approaches of Albayrak and Caber [16,35]. The IA value is calculated by subtracting the absolute value of the low-performance unstandardized regression coefficient from the high-performance coefficient and dividing the result by the sum of the two coefficients. An IA value exceeding 0.1 indicates an excitement factor, values between −0.1 and 0.1 indicate a performance factor, and values below −0.1 indicate a basic factor.

2.3. AIPA

Asymmetric Impact-Performance Analysis (AIPA) is a methodological approach designed to identify the attributes influencing customer satisfaction, determine their relative priorities, and thereby guide the formulation of service improvement strategies [36]. Figure 3 illustrates an example of AIPA, where the X-axis represents the performance of each attribute and the Y-axis denotes the IA index derived from PRCA. The two horizontal lines within the figure serve as thresholds for classifying attributes into excitement factors, performance factors, and basic factors. Specifically, attributes with IA values between 0.1 and 1 are categorized as excitement factors, those between −0.1 and 0.1 as performance factors, and those between −1 and −0.1 as basic factors [35]. The vertical line, in turn, distinguishes high-performance from low-performance attributes, calculated as the arithmetic mean of performance values across all attributes. Based on this framework, attributes are classified into six categories: High-performance Excitement (HE), High-performance Performance (HP), High-performance Basic (HB), Low-performance Excitement (LE), Low-performance Performance (LP), and Low-performance Basic (LB). To maximize customer satisfaction while minimizing resource expenditure, managerial attention should first be directed toward improving the quality of low-performance attributes, followed by maintaining the standards of high-performance attributes. When attributes exhibit the same performance level (either high or low), resource allocation priorities should follow the order of basic factors, performance factors, and excitement factors. Consequently, the overall improvement priorities is LB, LP, LE, HB, HP, and HE. AIPA extends conventional IPA by combining the IA (impact-asymmetry) index from PRCA with attribute performance, operationalizing the Three-Factor Theory. By contrasting high- vs. low-performance effects, PRCA reveals whether an attribute functions as basic, performance, or excitement; plotting IA against mean performance then positions attributes into actionable AIPA cells (HE, HP, HB, LE, LP, LB). This multidimensional approach moves beyond binary sentiment or symmetric assumptions and yields realistic, management-relevant priorities, which is crucial for diagnosing pre- and post-pandemic shifts in hotel attributes. While conventional IPA positions attributes only by their relative importance and performance, it does not account for the asymmetric nature of customer satisfaction. By contrast, AIPA integrates the Three-Factor Theory, enabling managers to identify whether attributes function as basic, performance, or excitement factors. This distinction provides clearer managerial guidance: basic factors must be maintained to avoid dissatisfaction, performance factors should be continuously optimized, and excitement factors can be enhanced to generate delight. Consequently, AIPA offers more nuanced and actionable priorities for resource allocation than traditional IPA.

Furthermore, prior studies have predominantly applied AIPA to survey-based, predefined attributes, restricting analysis to factors determined in advance by researchers and potentially overlooking attributes that matter most to customers. In contrast, this study employs BERTopic to automatically extract salient hotel attributes directly from large-scale online reviews and then applies AIPA to these empirically derived factors. This integration eliminates the constraints of pre-specified attributes and ensures that the asymmetric effects reflect customers’ authentic perceptions. The integration of BERTopic with AIPA has rarely been explored in hospitality research, and this study highlights its potential to link data-driven attribute discovery with asymmetric impact analysis, thereby providing fresh theoretical and managerial insights.

In the field of hospitality and tourism, AIPA has been widely employed to investigate the asymmetric relationship between customer satisfaction and hotel selection attributes. Albayrak and Caber [35], for example, conducted a survey of 163,090 customers using Turkish travel agencies and applied AIPA to identify the priority order for improving hotel selection attributes. Similarly, Dueñas et al. [37] applied AIPA to analyze regional differences among 409 tourists visiting Ecuador. Albayrak and Caber [38], focusing on the growing popularity of rock climbing, examined 473 tourists and employed AIPA to assess the asymmetric effects of destination attributes on overall satisfaction. Their analysis revealed that tourists traveling for rock climbing perceived basic facilities and accommodation as belonging to the LE category. In another study, Yuan et al. [39] applied AIPA to 604 tourists in Savannah, USA, to analyze the asymmetric relationship between urban tourism attributes and overall satisfaction.

While these studies provide valuable insights, they primarily relied on survey-based data collection, which is limited to factors predefined by researchers and poses challenges in obtaining large-scale datasets. To address these limitations, the present study employs web crawling to collect a substantial volume of hotel reviews and ratings from TripAdvisor. BERTopic is then applied to the collected reviews to extract hotel selection attributes, which subsequently serve as the basis for conducting AIPA.

3. Methodology

To overcome the limitations of survey-based approaches, this study collected comprehensive information, customer ratings, and TripAdvisor review data on hotels located in New York. The collected reviews were analyzed using BERTopic, an advanced topic modeling technique, to extract the hotel selection attributes perceived as most significant by customers. To conduct an in-depth examination of the impact of COVID-19, a segmented AIPA was performed based on hotel classification, travel type, and customer residence. As illustrated in Figure 4, the research framework was designed as a five-step analytical procedure.

In the first stage, hotel information, customer ratings, and review data for New York hotels were collected from the TripAdvisor platform using web crawling. In the second stage, review sentences were segmented and preprocessed, after which BERTopic was applied to extract topics. In the third stage, sentiment scores were computed using VADER, followed by the execution of PRCA in the fourth stage. Finally, in the fifth stage, AIPA was conducted based on the sentiment scores and PRCA results. A refined comparative analysis was then performed across the pre- and post-COVID-19 periods, segmented by hotel classification, travel type, and customer residence.

3.1. Data Collection

As shown in Figure 5, the data collection stage involved extracting comprehensive information on New York hotels from TripAdvisor. The collected variables included hotel name, hotel classification, customer reviews, overall satisfaction ratings on a five-point scale, review date, as well as travel type and place of residence reported by customers. In addition, to improve data quality, we applied data-cleaning procedures by removing incomplete, duplicated, abnormally short, or suspicious reviews that could potentially represent fake or spam content. These steps help reduce noise in the dataset and ensure that the subsequent analysis reflects more reliable customer feedback. TripAdvisor was chosen as the data source because it is among the world’s largest online travel review platforms [22]. As of November 2015, it reported 350 million unique monthly visitors and hosted more than 320 million reviews covering over 6.2 million accommodation providers, restaurants, and attractions across 48 international markets. This extensive scale and diversity ensure a robust dataset for analyzing hotel selection attributes and customer satisfaction. In the United States, a national emergency was declared in March 2020 due to COVID-19 and was lifted in April 2023 [40]. Accordingly, this study gathered data spanning the post-COVID-19 period (March 2020 to April 2023) using the Python programming language, while also collecting additional data from March 2016 to April 2019 to establish a valid basis for comparison with the pre-COVID-19 period. Specifically, the data collection employed Python 3.9 with Selenium WebDriver to automate navigation of dynamic TripAdvisor pages and BeautifulSoup4 for HTML parsing. Review texts, ratings, travel type, and residence information were systematically extracted and stored in structured datasets, ensuring reproducibility and comprehensive coverage.

3.2. BERTopic Analysis

Following the methodology of Albayrak et al. [41], this study employed the SpaCy package in Python to segment review data into individual sentences, as illustrated in Figure 6. This approach was adopted because a single review often contains multiple sentences addressing different aspects [41,42,43], and sentence-level analysis is considered more effective for extracting precise topics. The segmented sentences underwent preprocessing, including lowercasing, tokenization, removal of special characters, and elimination of stopwords. To determine the optimal number of topics, the Normalized Pointwise Mutual Information (NPMI) metric was utilized, which evaluates the semantic coherence of topics, with higher values indicating stronger cohesion [44]. Based on this evaluation, the optimal number of topics was established, after which BERTopic was applied. The extracted keywords were then assigned appropriate thematic labels with reference to prior research.

3.3. Sentiment Analysis

In this study, sentiment scores for each sentence were calculated using VADER (Valence Aware Dictionary and sEntiment Reasoner). VADER is a lexicon- and rule-based sentiment analysis tool, producing scores that range from −1 to +1, where values closer to −1 indicate negative sentiment, 0 represents neutrality, and values approaching +1 indicate positive sentiment [45]. Since TripAdvisor measures customer satisfaction on a five-point scale, this study followed the approach of Shang et al. [43] and converted sentiment scores accordingly, as shown in

Table 1. Sentiment score conversion [41].

−1	−0.9	−0.8	−0.7	−0.6	−0.5	−0.4	−0.3	−0.2	−0.1	0	0.1	0.2	0.3	0.4	0.5	0.6	0.7	0.8	0.9	1
Awful (1)					Bad (2)					Neutral (3)	Good (4)					Perfect (5)
“<−0.55”					“≥−0.55 and <0”					“=0”	“>0 and ≤0.55”					“>0.55”

. Specifically, scores below −0.55 were converted to 1, scores between −0.55 and 0 to 2, a score of 0 to 3, scores above 0 but up to 0.55 to 4, and scores exceeding 0.55 to 5.

Accordingly, Figure 7 presents an example based on Table 1, illustrating the results of sentiment analysis conducted on four sentences contained in Review 1, with the corresponding sentiment scores converted into a five-point scale.

Finally, the segmented sentences were reassembled into their original review units, with an example of this procedure illustrated in Figure 8. As shown, Review 1 consists of four sentences: Sentence 1 corresponds to Topic 2 with a sentiment score of 5, Sentence 2 to Topic 1 with a score of 3, Sentence 3 to Topic 2 with a score of 3, and Sentence 4 to Topic N with a score of 2. Since Sentences 1 and 3 both pertain to Topic 2, the average sentiment score for these sentences was calculated, following the approach of Shang et al. [43]. Consequently, the final sentiment score for Topic 2 was determined as 4, while the final scores for Topic 1 and Topic N were calculated as 3 and 2, respectively.

3.4. PRCA

In the PRCA stage, binary dummy variables for High Performance and Low Performance are first generated for each hotel attribute. Subsequently, multiple regression analysis is conducted to examine the impact of these dummy variables on overall satisfaction.

Step 1: Dummy variable generation

In Step 1, dummy variables are generated based on the sentiment scores of hotel attribute topics obtained through sentiment analysis in Figure 8. For example, when the sentiment score is 4 or 5, High Performance is assigned a value of 1 and Low Performance a value of 0. Conversely, when the sentiment score is 1 or 2, High Performance is set to 0 and Low Performance to 1. If the sentiment score is 3, both variables are coded as 0.

Step 2: Regression Analysis

In this stage, the High Performance and Low Performance dummy variables generated for each topic are used as independent variables, while the overall satisfaction rating serves as the dependent variable. Multiple regression analyses are conducted by hotel classification. The unstandardized regression coefficients obtained from this analysis are then employed to classify each hotel selection attribute into one of the Three-Factor Theory categories—basic factors, performance factors, or excitement factors.

Step 3: Attribute Classification

Finally, using the unstandardized regression coefficients of High Performance and Low Performance obtained from the multiple regression analyses by hotel classification, each hotel selection attribute is categorized into excitement, performance, or basic factors. Following the approaches of Albayrak and Caber [35] and Bi et al. [16], this study employs the IA value, which reflects the asymmetric influence between the two coefficients. Specifically, the IA value is calculated by subtracting the absolute value of the Low Performance coefficient from the High Performance coefficient and dividing the result by the sum of the two coefficients. Attributes with IA values greater than 0.1 are classified as excitement factors, those between −0.1 and 0.1 as performance factors, and those below −0.1 as basic factors.

3.5. AIPA

In the AIPA stage, the X-axis (Performance) is calculated as the average of the topic-specific five-point scale scores derived from the sentiment analysis stage, while the Y-axis (IA value) is obtained from the topic-specific IA values computed in the PRCA stage. First, a comparative analysis using AIPA is conducted on the entire dataset for the pre- and post-COVID-19 periods. Subsequently, to provide a more refined assessment of the pandemic’s impact, additional comparative analyses are performed by segmenting the data according to hotel classification, travel type, and customer residence.

4. Results

4.1. Descriptive Statistics of the Data

This study collected TripAdvisor review data spanning the post-COVID-19 period and, for comparison, the pre-COVID-19 period (March 2016 to April 2019). This design was adopted to account for the strong association between travel motivations, destination choices, and seasonality [46,47], thereby minimizing seasonal distortions through equivalent time frames. As shown in

Table 2. Descriptive statistics of pre-COVID-19 data.

Economy Hotel (1,2 Star)	Mid-Range Hotel (3 Star)	Luxury Hotel (4,5 Star)	User Region		Travel Type
			Local	Foreign	Business	Leisure
227	13,920	31,129	23,609	21,667	9450	35,826

and

Table 3. Descriptive statistics of post-COVID-19 data.

Economy Hotel (1,2 Star)	Mid-Range Hotel (3 Star)	Luxury Hotel (4,5 Star)	User Region		Travel Type
			Local	Foreign	Business	Leisure
244	2820	5778	4840	4002	2320	6522

, a total of 45,276 reviews were collected for the pre-COVID-19 period and 8842 reviews for the post-COVID-19 period, covering 414 hotels in New York City.

By hotel classification, 244 reviews were collected for economy hotels, 2820 for mid-range hotels, and 5778 for luxury hotels in the post-COVID-19 period, compared with 227, 13,920, and 31,129 reviews, respectively, in the pre-COVID-19 period. By travel type, the dataset includes 2320 business travelers and 6522 leisure travelers post-COVID-19, compared with 9450 and 35,826 reviews, respectively, pre-COVID-19. By residence, 4840 reviews were from local travelers and 4002 from foreign travelers post-COVID-19, whereas 23,609 and 21,667 reviews, respectively, were collected pre-COVID-19.

4.2. BERTopic Results

In this study, the number of topics was determined with reference to prior research, using the NPMI value as the criterion. As shown in Figure 9, the NPMI values for different topics were calculated, and the highest value of 0.153 was observed when the number of topics was set to eight. This indicates that semantic coherence is maximized when the model extracts eight topics.

Accordingly, this study set the number of topics to eight. The eight topics extracted through BERTopic and their top ten keywords are presented in

Table 4. BERTopic results.

Topic 1	Topic 2	Topic 3	Topic 4	Topic 5	Topic 6	Topic 7	Topic 8
Room	Location	F&B	Staff	Mobility	Front Desk	Public facilities	Internet
room	location	breakfast	staff	square	check	elevator	wifi
bed	trip	bar	desk	walk	luggage	lift	free
clean	visit	restaurant	friendly	subway	charge	wait	internet
bathroom	place	coffee	helpful	park	bag	floor	work
shower	great	food	front	central	card	slow	fast
comfortable	business	drink	service	station	fee	stair	charge
small	price	rooftop	welcome	broadway	key	system	speed
view	family	eat	reception	close	reservation	long	pay
floor	weekend	lounge	customer	distance	sheraton	peak	signal
noise	city	wine	guest	location	late	fast	include

. Thematic labels were assigned by manually reviewing the keywords extracted by BERTopic and aligning them with constructs established in prior hospitality and tourism studies employing topic modeling [15,16,17,20,25,29]. This approach ensured that topic labels were not determined arbitrarily but remained interpretable and consistent with established categories in the literature. Topic 1, represented by keywords such as room, bed, and clean, was labeled Rooms. Topic 2, characterized by location, trip, and visit, reflects the Location of the hotel. Topic 3, containing breakfast, bar, and restaurant, pertains to Food and Beverage. Topic 4, with keywords including staff, desk, and friendly, was identified as Staff. Topic 5, comprising square, walk, and subway, corresponds to Mobility. Topic 6, represented by check, bag, and charge, was labeled Front Desk. Topic 7, with elevator, lift, and stair, relates to Public Facilities. Finally, Topic 8, characterized by wifi, free, and internet, was categorized as Internet.

4.3. AIPA Results

The number of reviews varied considerably across subgroups such as hotel classification, travel type, and residence, the analysis was conducted in two stages. First, we performed an overall comparative analysis of AIPA results using the full dataset, ensuring that the primary findings reflect aggregate patterns. Second, subgroup analyses by hotel classification, travel type, and residence were carried out to provide additional insights beyond the aggregate comparison. During the PRCA regression stage, multicollinearity diagnostics were conducted, and all Variance Inflation Factor (VIF) values were below 5. This confirms that the classification of attributes into basic, performance, and excitement factors is statistically robust.

4.3.1. Overall Analysis Results

The PRCA regression results and IA-based categorizations are reported in

Table 5. IA-based categorization of hotel attributes (pre-COVID).

Attributes	High Performance		Low Performance		IA Value	Categorization
	B	VIF	B	VIF
Room	0.433 ***	1.04	−0.605 ***	1.031	−0.166	Basic
Location	0.404 ***	1.036	−0.734 ***	1.026	−0.289	Basic
F&B	0.368 ***	1.025	−0.361 ***	1.019	0.011	Performance
Staff	0.349 ***	1.024	−1.081 ***	1.029	−0.511	Basic
Mobility	0.245 ***	1.018	0.041	1.015	0.712	Excitement
Front	−0.010	1.008	−0.518 ***	1.014	−0.963	Basic
Public facilities	−0.094 ***	1.003	−0.421 ***	1.002	−0.636	Basic
Internet	−0.074 **	1.003	−0.607 ***	1.001	−0.782	Basic

*** p < 0.001, ** p < 0.01, R2 = 0.262, Durbin-Watson = 1.846, Dependent Variable = Overall customer satisfaction.

(pre-COVID) and

Table 6. IA-based categorization of hotel attributes (post-COVID).

Attributes	High Performance		Low Performance		IA Value	Categorization
	B	VIF	B	VIF
Room	0.560 ***	1.054	−0.966 ***	1.058	−0.266	Basic
Location	0.640 ***	1.073	−0.872 ***	1.038	−0.154	Basic
F&B	0.538 ***	1.043	−0.621 ***	1.024	−0.072	Performance
Staff	0.640 ***	1.053	−0.888 ***	1.057	−0.162	Basic
Mobility	0.360 ***	1.027	−0.013	1.013	0.928	Excitement
Front	0.003	1.013	−0.715 ***	1.032	−0.992	Basic
Public facilities	−0.203 **	1.004	−0.609 ***	1.006	−0.500	Basic
Internet	0.131	1.005	0.518	1.008	−0.597	Basic

*** p < 0.001, ** p < 0.01, R2 = 0.407, Durbin-Watson = 1.770, Dependent Variable = Overall customer satisfaction.

(post-COVID). Before the pandemic, mobility was categorized as an excitement factor, F&B as a performance factor, and all other attributes as basic factors.

After the pandemic, the classification pattern remained largely consistent: mobility continued to function as an excitement factor, F&B remained a performance factor—though its IA value approached the threshold of a basic factor—and the remaining attributes persisted as basic factors. These categorizations provide a statistically grounded foundation for interpreting the subsequent AIPA.

The overall AIPA results for the pre- and post-COVID-19 periods are presented in Figure 10. Specifically, in the post-COVID-19 period, public facilities, rooms, and front desk attributes were classified as LB; location, food and beverage, staff, and internet attributes as HB; and mobility as HE. In contrast, during the pre-COVID-19 period, public facilities, front desk, and room attributes were classified as LB; location, staff, and internet as HB; food and beverage as HP; and mobility as HE.

4.3.2. Results by Hotel Classification

The AIPA results by hotel classification are illustrated in Figure A1 (Appendix A). For economy hotels, most attributes retained similar positions across periods; however, location and staff gained importance, moving into higher categories after COVID-19. In mid-range hotels, attributes showed overall stability, with food and beverage and mobility continuing to act as strong drivers of satisfaction. For luxury hotels, only minor changes were observed, though staff and food and beverage shifted toward higher-impact categories in the post-pandemic period.

4.3.3. Results by Travel Type

Figure A2 (Appendix A) presents the results by travel type. Leisure travelers exhibited few changes, with most attributes remaining stable across periods, while food and beverage became more strongly associated with satisfaction after COVID-19. Business travelers, however, showed more noticeable shifts: location lost relative importance, whereas staff and food and beverage emerged as stronger drivers in the post-pandemic period.

4.3.4. Results by Customer Residence

The AIPA results by customer residence are shown in Figure A3 (Appendix A). For local travelers, attributes remained largely consistent, with mobility continuing as a key driver. For foreign travelers, more pronounced changes were observed: staff, in particular, increased in importance, moving into a higher-impact category after COVID-19, while food and beverage and location also gained greater weight in shaping satisfaction.

5. Discussion

This study employed an overall AIPA to examine the impact of COVID-19 on hotel attributes, followed by segmented analyses based on hotel classification, travel type, and customer residence. The results indicate that, with the exception of food and beverage, most attributes exhibited no significant changes. Prior to COVID-19, food and beverage were positioned within the HP quadrant; however, it shifted to the HB quadrant in the post-COVID-19 period. This suggests that, before the pandemic, food and beverage functioned as a high-performing attribute, generating satisfaction when expectations were met and dissatisfaction when they were unmet. After the pandemic, however, the attribute ceased to elicit satisfaction even when expectations were fulfilled, instead provoking dissatisfaction only when expectations were not met. In other words, while the performance level of food and beverage remained consistently high, its functional role shifted. From the perspective of AIPA’s resource allocation priorities, this finding highlights the necessity of more concentrated management and investment in food and beverage services in the post-pandemic era. These results align with Hameed et al. [48], which observed that although hotels enhanced the hygiene and quality of food and beverage offerings after the pandemic, customers increasingly avoided in-hotel dining for safety reasons, perceiving room service and takeout as default options. Such changes in customer perception account for the repositioning of the food and beverage attribute. These findings can also be interpreted through the lens of consumer behavior under uncertainty. In times of heightened health risks, customers placed greater emphasis on safety and reliability, even when service quality in food and beverage was maintained. This shift resonates with the literature on service resilience, which highlights the importance of adaptive capacity and flexible resource allocation in sustaining customer trust during crises.

A comparison of AIPA results for economy hotels before and after COVID-19 revealed notable changes across most attributes, with the exception of front desk and mobility.

First, the room attribute shifted from the LP quadrant (low performance; satisfaction when expectations are met, dissatisfaction when unmet) in the pre-pandemic period to the LE quadrant (low performance; satisfaction when expectations are met, no dissatisfaction when unmet) after COVID-19. This suggests that enhanced cleanliness and quality control in rooms during the pandemic led customers to experience satisfaction when expectations were exceeded, while unmet expectations did not necessarily result in dissatisfaction.

Second, the location attribute moved from the LB quadrant (low performance; satisfaction when expectations are met, no dissatisfaction when unmet) to the HP quadrant (high performance; satisfaction when met, dissatisfaction when unmet). This reflects a shift in customer priorities: whereas budget-conscious guests had relatively low expectations for location prior to the pandemic, safety concerns such as the avoidance of public transportation heightened the importance of hotel location in the post-pandemic period [49].

Third, the food and beverage attribute transitioned from the HB quadrant (high performance; no direct effect on satisfaction or dissatisfaction) to the HP quadrant (high performance; generating both satisfaction and dissatisfaction). Prior to the pandemic, food and beverage services—typically limited to complimentary breakfast in economy hotels—were not a decisive factor in guest satisfaction [50]. However, in the post-pandemic context, customers increasingly favored contactless and minimal-contact services for safety reasons [4], making the provision or absence of food and beverage services a critical determinant of both satisfaction and dissatisfaction. Hence, instead of excessive resource allocation, economy hotels should strengthen hygiene and safety in their food and beverage offerings.

Fourth, the staff attribute shifted from the HB quadrant (high performance; no direct effect on satisfaction or dissatisfaction) before COVID-19 to the HP quadrant afterward. This indicates that staff-related factors—such as mask wearing, sanitization practices, and adherence to social distancing protocols—became crucial determinants of guest satisfaction. Customers were highly sensitive to the presence or absence of these safety measures, experiencing satisfaction when expectations were met and dissatisfaction when they were not. Consequently, economy hotels must prioritize staff compliance with hygiene and safety protocols while ensuring the availability of contactless services.

In the case of mid-range hotels, the room attribute shifted from the LB quadrant (low performance; no effect on satisfaction or dissatisfaction) before COVID-19 to the HB quadrant (high performance; no effect on satisfaction or dissatisfaction) afterward. This indicates that while rooms did not serve as a decisive determinant of customer satisfaction in either period, their performance level improved substantially. In practice, many U.S. hotels introduced enhanced hygiene protocols during the pandemic, such as providing hand sanitizers in rooms and removing nonessential items like magazines and stationery that posed potential risks of cross-contamination. These measures contributed to the observed improvement in room performance. Accordingly, for mid-range hotels, rather than allocating additional resources, a strategy focused on maintaining and sustaining the enhanced quality of rooms in the post-pandemic context is deemed most appropriate.

For luxury hotels, the location attribute shifted from the HB quadrant (high performance; no direct effect on satisfaction or dissatisfaction) before COVID-19 to the HP quadrant (high performance; generating both satisfaction and dissatisfaction) afterward. Prior to the pandemic, the advantageous positioning of luxury hotels—typically situated near city centers, tourist attractions, and shopping districts—was largely taken for granted by guests [51]. However, in the post-pandemic context, with restrictions on mobility and widespread avoidance of public transportation, location emerged as a critical operational factor directly tied to safety. Consequently, location has become a pivotal attribute that simultaneously drives both satisfaction and dissatisfaction, underscoring its importance as a key area of managerial focus in the post-COVID-19 era.

The analysis by travel type revealed that for leisure travelers, most attributes showed little change except for food and beverage. Prior to COVID-19, food and beverage were positioned in the HP quadrant (high performance; generating both satisfaction and dissatisfaction), but shifted to the HB quadrant (high performance; generating only dissatisfaction) afterward. This suggests that while food and beverage offerings previously contributed directly to customer satisfaction, the shift toward takeout and delivery services in the post-pandemic era meant that their provision became taken for granted [52]. As a result, food and beverages no longer generated additional satisfaction but, when absent, served as a source of dissatisfaction.

For business travelers, most attributes remained stable, with notable changes observed in location, food and beverage, mobility, and staff. Specifically, the location attribute shifted from the HB quadrant (high performance; no effect on satisfaction or dissatisfaction) before the pandemic to the LP quadrant (low performance; generating both satisfaction and dissatisfaction) afterward. Prior to COVID-19, hotel location was largely regarded as a given for business travelers. However, with the avoidance of public transportation and mobility restrictions during the pandemic, the importance of location increased significantly. For business travelers, who prioritize accessibility and proximity for work-related purposes, location transformed into an attribute that elicits satisfaction when expectations are met and dissatisfaction when unmet.

Second, the food and beverage attribute shifted from the HP quadrant (high performance; generating both satisfaction and dissatisfaction) before COVID-19 to the HB quadrant (high performance; generating only dissatisfaction) afterward. This reflects a change in customer behavior, as guests increasingly preferred contactless interactions and avoided crowded spaces for safety reasons, rendering food and beverage services an expected baseline rather than a differentiating factor. Consequently, while their provision no longer generates additional satisfaction, their absence provokes dissatisfaction.

Third, the mobility attribute moved from the HE quadrant (high performance; generating only satisfaction) in the pre-pandemic period to the HP quadrant (high performance; generating both satisfaction and dissatisfaction) after COVID-19. Prior to the pandemic, convenient transportation enhanced satisfaction when available but did not necessarily induce dissatisfaction when lacking. However, due to restricted access to transportation options during the pandemic [53], mobility became a decisive factor influencing both satisfaction and dissatisfaction simultaneously.

Fourth, the staff attribute transitioned from the HB quadrant (high performance; generating only dissatisfaction) before COVID-19 to the HP quadrant afterward. In the pre-pandemic context, staff services did not significantly enhance satisfaction when expectations were met. In contrast, the post-pandemic expansion of contactless services and robotic assistance transformed staff into a central component of the guest experience. For customers less accustomed to technological solutions, this shift often led to inconvenience, underscoring the heightened importance of staff management as a determinant of satisfaction in the post-pandemic environment.

The analysis by customer residence revealed that for local travelers, most attributes remained stable, with the exception of food and beverage. Prior to COVID-19, food and beverage were positioned in the HP quadrant (high performance; generating both satisfaction and dissatisfaction), but shifted to the HB quadrant (high performance; generating only dissatisfaction) afterward. This shift can be explained by the fact that local travelers are more familiar with hotel policies and service practices, and thus take the availability of delivery and takeout services for granted.

For foreigner travelers, notable changes were observed in rooms, staff, and food and beverage. First, the room attribute moved from the LB quadrant (low performance; no effect on satisfaction or dissatisfaction) before the pandemic to the HB quadrant (high performance; generating only dissatisfaction) afterward. This reflects the improvement of room performance following the implementation of enhanced health protocols, including social distancing and disinfection measures, and is consistent with the findings of Mehta et al. [54]. Song et al. [55] similarly reported that guests increasingly view not only traditional services but also hygiene and safety measures as key satisfaction drivers, supporting the results of this study.

Second, the staff attribute shifted from the HB quadrant (high performance; generating only dissatisfaction) before COVID-19 to the HP quadrant (high performance; generating both satisfaction and dissatisfaction) afterward. While staff services previously did not elicit additional satisfaction when expectations were met, the expansion of contactless services and social distancing policies led to the reduction or alteration of certain services, directly affecting guest experiences. According to Bonfanti et al. [10], kiosks for check-in and check-out, smartphone-based requests, and room service ordering became widespread during the pandemic. However, for foreigner travelers unfamiliar with such digital interfaces, these changes increased the likelihood of dissatisfaction.

Third, the food and beverage attribute shifted from the HE quadrant (high performance; generating only satisfaction) before the pandemic to the HB quadrant afterward. Previously, foreigner travelers regarded local food and beverage services as opportunities for cultural exploration [56]. In the post-pandemic context, however, the normalization of delivery and takeout services transformed food and beverage into an expected offering, no longer a source of additional satisfaction but a factor that induces dissatisfaction when absent.

Similar dynamics have been observed in other service sectors. In the airline industry, travelers became highly sensitive to refund policies and cabin cleanliness, reflecting the prioritization of safety and reliability under uncertainty [57]. Likewise, in the restaurant sector, hygiene protocols and contactless services emerged as decisive factors influencing satisfaction, with full-service restaurants even reporting improved satisfaction levels due to enhanced safety practices [58]. These parallels suggest that the shift observed in hotel attributes is part of a broader trend in consumer behavior under uncertainty, highlighting the role of service resilience across industries.

6. Conclusions

6.1. Summary

This study applied the AIPA method to a large-scale dataset to analyze changes before and after COVID-19, with further comparisons segmented by hotel classification, travel type, and customer residence. The analysis identified eight key hotel selection attributes valued by customers—rooms, location, food and beverage, staff, mobility, front desk, public facilities, and internet. First, the overall analysis revealed that, with the exception of food and beverage, most attributes exhibited little change across the pre- and post-COVID-19 periods.

Second, the comparison by hotel classification revealed significant differences only in the attributes of rooms, location, food and beverage, and staff for economy hotels; rooms for mid-range hotels; and location for luxury hotels.

Third, the analysis by travel type showed that changes were observed only in food and beverage for leisure travelers, whereas for business travelers, location, food and beverage, mobility, and staff exhibited notable shifts.

Fourth, the residence-based analysis indicated that for local travelers, only food and beverage showed differences, while for foreigner travelers, changes were evident in rooms, staff, and food and beverage.

Taken together, food and beverage, location, rooms, and staff emerged as critical managerial attributes in the post-pandemic context, closely linked to heightened customer awareness of safety and the preference for contactless services. Accordingly, hotel managers must strategically reallocate resources to strengthen hygiene and safety standards, expand contactless service options, and ensure convenient transportation access to meet evolving customer expectations.

6.2. Implication

The academic implications of this study are as follows. First, this research collected and analyzed a large-scale dataset comprising reviews from customers across multiple countries and hotels. By leveraging such extensive data, the study overcomes the limitations of traditional survey methods, which are often difficult to administer and require significant time and financial resources, thereby offering substantial academic value. Moreover, this study advances methodology by integrating BERTopic with AIPA, allowing attributes to be derived directly from customer reviews and then analyzed for their asymmetric effects on satisfaction. Unlike conventional LDA, which is based on word co-occurrence frequencies and often struggles to capture contextual meaning in short reviews, BERTopic leverages contextual embeddings to identify more coherent and semantically meaningful topics. This capability is particularly important in hotel reviews, where short sentences frequently mix multiple service attributes. This novel combination provides a more data-driven and nuanced framework than conventional survey-based approaches. In addition, compared with existing approaches in hospitality research, the BERTopic–AIPA framework offers distinct advantages. Traditional survey-based IPA studies rely on predefined attributes and assume symmetric effects, whereas regression models often capture only linear relationships. By contrast, BERTopic extracts emergent attributes directly from unstructured customer reviews, and AIPA incorporates asymmetric effects, together providing richer managerial insights that more accurately reflect actual customer experiences.

Second, whereas prior studies often examined only partial periods of the COVID-19 pandemic, this study considered the entire duration of the pandemic. By conducting a detailed comparative analysis across hotel classifications, travel types, and customer residences, the study not only complements but also substantiates the findings of existing COVID-19-related research.

Third, this study conducted a fine-grained comparison based on customer characteristics such as hotel classification, travel type, and residence. The results demonstrate that satisfaction with hotel selection attributes varies according to hotel and customer type. This finding underscores the necessity for future research to segment customer groups more precisely when analyzing hotel satisfaction.

Fourth, although the dataset focuses on New York City hotels, the findings have broader global relevance. New York represents one of the most diverse and competitive urban hospitality markets in the world, and the shifts observed here provide insights that can inform hotel management practices internationally. These results highlight that the dynamics of customer satisfaction under conditions of uncertainty are not confined to a single location, but rather reflect a wider trend with implications for the global hospitality industry.

The practical implications of this study are as follows. First, in relation to hotel responses to COVID-19, the findings provide several important and actionable insights for practitioners. Specifically, they offer guidance for hotel managers in formulating appropriate strategies during health crises with high transmissibility and significant societal impact—such as SARS, H1N1, Ebola, MERS, and COVID-19. By tailoring services according to hotel classification, travel type, and customer residence, managers can draw upon the results of this study to enhance overall customer satisfaction.

Second, while Song et al. [55] reported an overall increase in hotel satisfaction after COVID-19, the present study, through a segmented analysis of customer types and a comprehensive assessment of the entire pandemic period, reveals that changes in overall satisfaction varied significantly depending on customer characteristics. Accordingly, hotels are advised to consider the prioritization of hotel selection attributes identified in this study as a means of improving overall customer satisfaction and strengthening their marketing strategies.

Third, the findings contribute to the development of service resilience by highlighting which hotel attributes require the most attention under crisis conditions. By reallocating resources to safety-sensitive attributes such as food and beverage, staff, and hygiene-related services, hotels can maintain customer trust and ensure operational continuity. Moreover, although the dataset focuses on New York City, the insights derived from a highly diverse and competitive hospitality market have broader relevance for hotel operators worldwide. Finally, the results underscore the growing importance of digital and contactless services, suggesting that hotels should invest in technological innovations to strengthen resilience and preparedness for future disruptions.

Fourth, from a sustainability perspective, the AIPA framework enables hotels to optimize resource allocation by avoiding over-investment in attributes that merely prevent dissatisfaction (basic factors) and focusing instead on excitement factors that generate disproportionate satisfaction. This approach reduces waste, enhances operational efficiency, and supports the long-term viability of the hospitality sector.

6.3. Future Recommendation and Study Limitation

The limitations of this study and directions for future research are as follows. First, this study analyzed hotel data collected exclusively from New York, United States. However, governmental responses to COVID-19 varied considerably across countries and cities. Future research should therefore incorporate regional contexts to provide a more comprehensive understanding. Moreover, although this study relied on TripAdvisor reviews, the proposed methodology is transferable to other platforms such as Booking.com or Expedia, provided that sufficient review data are available. Applying the framework across diverse regions and platforms would further test the robustness and generalizability of the findings.

Second, this study did not consider cultural or demographic moderators that may influence how customers perceive and evaluate hotel attributes. Future research should examine these moderating factors to determine whether satisfaction drivers vary across customer groups with different cultural backgrounds, ages, or other demographic characteristics. Such an approach would provide a deeper understanding of heterogeneous customer responses under conditions of uncertainty.

Third, the sample size for economy hotels in the post-COVID-19 period was relatively small. While the analysis still provides useful indications of attribute shifts, the limited number of reviews may constrain the robustness and generalizability of results for this subgroup. Future research should therefore incorporate larger samples of economy hotels or employ weighting techniques to validate these findings.

Fourth, as with other studies based on online reviews, potential biases such as self-selection and polarization may influence the results, and these factors should be considered when interpreting the findings.