Review content and customer satisfaction: a textual analysis of Yelp restaurant reviews

Online consumer reviews have become a central mechanism for evaluating experience goods in the modern economy, particularly for services like dining where quality is observable only after consumption. Ratings on platforms such as Yelp now significantly influence consumer choices and business performance (Hamzah et al., 2020; Abed et al., 2015; Fogel and Zachariah, 2017; Camilleri, 2018). Empirical studies show that a one-star Yelp rating increase can raise a restaurant’s revenue by roughly 5–9% (Luca, 2016; Anderson and Magruder, 2012), underscoring how digital reputation directly affects firm outcomes (Fang, 2022). Despite the importance of online ratings, we lack clear empirical evidence identifying which specific restaurant attributes drive these review outcomes (Yang and Luo, 2021; Liao et al., 2023). In other words, while it is well understood that online ratings affect consumer behavior, the role of tangible operational features (e.g. service offerings, amenities, ambience) in shaping customer satisfaction remains underexplored.

Most prior research on online reviews has concentrated on aggregate rating effects or platform dynamics. For example, studies have examined how review ratings impact sales (Chevalier and Mayzlin, 2006) and have investigated issues of review authenticity and manipulation (Luca and Zervas, 2016; Gallagher et al., 2019). Relatively few studies, however, have analyzed how verifiable business features map into customer satisfaction as expressed in reviews. Where such work exists, it often relies on broad categorizations (such as cuisine type or price tier) or infers attributes from review text without confirmation, leaving a gap in our understanding of how specific operational characteristics contribute to positive or negative customer feedback.

This study addresses that gap by explicitly linking structured restaurant attributes to observed review sentiment and ratings. Our approach is guided by Lancaster’s (1966) theory that consumers derive utility from a product’s individual characteristics rather than the product as a whole. Applying this perspective to restaurants, we investigate how concrete features of the dining experience (for instance, the availability of delivery or a pleasant ambience) drive customer satisfaction metrics. We incorporate the textual sentiment of Yelp reviews—quantifying the positivity or negativity of review language—in addition to the traditional star rating. By capturing nuances in tone that star ratings alone may miss, this approach provides a richer understanding of consumer evaluations for experience goods like restaurants.

Understanding which attributes customers value is essential for at least two reasons. First, consumers increasingly rely on online ratings to choose unfamiliar businesses. If we do not know what features underlie those ratings, firms and platforms might misallocate resources or mislead customers by emphasizing the wrong aspects. Second, businesses often invest in operational improvements specifically to boost online reviews. A clearer understanding of what customers actually reward can help avoid costly investments in features that have little impact on satisfaction. More broadly, identifying the determinants of review sentiment and ratings can inform the design of better feedback mechanisms and search filters in platform markets, where information asymmetry is high.

Our contributions are threefold. First, we assemble a national-scale dataset of approximately 580,000 US restaurant reviews and provide the first evidence at this scope linking concrete restaurant features to both review text sentiment and star ratings. Second, we examine non-linear effects in consumer responses by distinguishing threshold effects (the jump in satisfaction from adding a basic feature where none existed) from diminishing returns (the smaller incremental gains from additional features once the basics are in place). Third, we demonstrate that many peripheral features—such as extra amenities, entertainment options, or very broad menu offerings—yield little to no improvement in customer satisfaction once core expectations (reliable service, decent food, a pleasant environment) are met.

The remainder of the paper is organized as follows. Section 2 reviews relevant literature on customer satisfaction, online reviews, and product attributes in platform markets. Section 3 describes the dataset and the construction of feature categories. Section 4 outlines the empirical methodology and estimation strategy. Section 5 presents the regression results along with robustness checks. Section 6 concludes with implications for economic theory, managerial practice, and future research.

Customer satisfaction in the restaurant sector depends on a bundle of characteristics that collectively shape the dining experience. These typically include the quality of food, the level of service, the ambience of the environment, pricing, and convenience (Ramanathan et al., 2016; Vasani et al., 2024). Classic service quality frameworks like SERVQUAL (Parasuraman et al., 1988) identified reliability and responsiveness as critical components of customer evaluations in service industries. Subsequent models geared toward restaurants, such as DINESERV (Stevens et al., 1995), similarly underscored the importance of attentive staff and a well-maintained environment in driving satisfaction. Empirical evidence confirms that while food quality often ranks as the most important factor, service and atmosphere significantly influence the overall experience as well (Sulek and Hensley, 2004). These patterns align with Lancaster’s (1966) theory that consumers derive utility from a product’s attributes: diners evaluate their restaurant experience not just based on the food itself, but on the collection of service elements and contextual factors in which it is delivered.

A consistent theme in this literature is the presence of diminishing returns and threshold effects in customer satisfaction. Once a basic level of quality or service is provided, further enhancements tend to yield smaller marginal gains in perceived satisfaction (Ryu and Han, 2011). In many cases, customers respond positively only up to a certain threshold—ensuring core expectations are met—beyond which additional improvements have limited impact. For example, in a study of fast-food outlets, Chun and Nyam-Ochir (2020) find that food, service, and ambience each boost customer satisfaction up to the point of meeting expectations, but improvements beyond those baseline expectations produce minimal gains. This non-linearity suggests that consumers often behave as satisficers: they reward restaurants for meeting basic expectations, but do not increase their ratings proportionally once those standards are fulfilled.

Ambience is one aspect of the restaurant experience where the law of diminishing returns also becomes apparent. Bitner (1992) introduced the notion of a “servicescape” to describe how a service environment’s physical design (layout, lighting, music, decor, etc.) affects customer perceptions. Consistent with this idea, Ryu and Jang (2008) find that an appealing ambience can improve customer satisfaction and encourage repeat visits. However, these positive effects tend to plateau once a reasonable standard is achieved. In fact, extremely lavish décor may raise expectations that the food or service cannot meet, which can lead to disappointment. In practice, simple, clean, and authentic settings often satisfy diners just as effectively as opulent interiors, so long as core quality standards in food and service are met.

Amenities and convenience features (such as parking availability, Wi-Fi, or child-friendly options) often function as hygiene factors in the dining experience. According to Kano’s (1984) model, certain attributes are “must-haves” that customers expect by default and only really notice when they are lacking. Empirical studies support this idea. Gan et al. (2017) find that various amenities explain only a small fraction of the variation in Yelp star ratings after accounting for food and service quality. Rabaei et al. (2021) reach a similar conclusion, noting that amenities contribute minimally to customer satisfaction once basic needs are fulfilled and thus act primarily to prevent dissatisfaction. In other words, a restaurant without essential conveniences may be penalized in reviews, but adding more and more amenities beyond the basics offers little additional reward.

Offering an excessive array of features can even backfire due to information overload or unintended signaling effects. Consumers tend to prefer familiar, straightforward options over overly complex offerings (Szmigin and Foxall, 1998). In line with this tendency, Luca (2016) observed that restaurants advertising a very wide range of payment methods had slightly lower average ratings. One plausible explanation is that an overabundance of features can create confusion or signal a mass-market, low-specialization strategy that some customers interpret negatively. Thus, more is not always better: beyond a certain point, piling on extra options may actually detract from the perceived quality or identity of the service.

The breadth of a restaurant’s menu provides another example of the trade-off between variety and focus. Offering a diverse menu can attract a broader customer base, but an overly extensive menu might dilute perceived quality or overwhelm patrons. Behavioral research shows that too much choice can reduce satisfaction (Iyengar and Lepper, 2000). In the restaurant context, Johns et al. (2013) document that medium-sized menus tend to coincide with higher customer ratings, whereas extremely long menus correlate with lower satisfaction. This pattern suggests diminishing marginal utility from variety and possibly concerns about a lack of specialization in the kitchen. Notably, prior studies of menu breadth and its impact on ratings have often been limited to specific locales or relatively small samples. Our analysis expands on this line of inquiry by testing menu breadth effects at a national scale using a large cross-section of restaurants.

Despite the rich body of research on customer satisfaction factors, relatively few studies have directly connected the presence of specific business attributes to online review outcomes. Digital platforms like Yelp have become an important context for studying consumer behavior in the experience economy, as they help reduce information asymmetry by revealing product quality over time. Researchers have used Yelp data to measure the impact of ratings on restaurant demand (Luca, 2016; Anderson and Magruder, 2012), to detect fake or manipulated reviews (Luca and Zervas, 2016; Gallagher et al., 2019), and to explore how firms manage and respond to online reputations. However, most of this work treats either the numeric star rating or the textual review content in isolation. Few studies leverage Yelp’s structured attribute data—details like whether a restaurant offers delivery, takes reservations, has outdoor seating, and so on—to explain variation in customer satisfaction. Review text often mentions such features, but without verification of their presence, it is difficult to pin down their true effect.

Recent research has begun to address this gap. For instance, Yu et al. (2017) attempted to infer restaurant features from review text via machine learning, but that indirect approach can introduce substantial measurement error. Another notable study by Tian et al. (2021) combined star ratings with textual sentiment analysis to examine drivers of the dining experience, finding that review sentiment (the emotional tone of the text) and star ratings capture different facets of customer satisfaction and may respond differently to certain attributes. Building on these insights, our study takes a more direct approach by utilizing the verified attribute flags provided by the Yelp platform. We aggregate nearly 70 binary restaurant attributes into seven broad categories (services, ambience, amenities, alcohol, payment options, menu breadth, and entertainment). This framework allows us to test two main propositions: (1) whether having at least one feature in a given category—crossing the threshold from “none” to “some”—significantly boosts customer satisfaction, and (2) whether offering more features within that category provides any incremental benefit or instead exhibits diminishing returns. Through this empirical strategy, we evaluate how specific, tangible aspects of the restaurant experience contribute to both the sentiment expressed in customer reviews and the star ratings those reviews accompany.

We use the 2019 Yelp Open Dataset (Data Licensing, 2025), which provides a publicly available sample of Yelp reviews and business information for research purposes. The dataset spans one full year (2019), avoiding seasonal biases and any disruptions from the 2020 pandemic. After cleaning the data – removing duplicate business entries, excluding any venues flagged by Yelp for irregular activity (to avoid fake or closed listings), and dropping restaurants with fewer than five reviews – our final sample contains roughly 35,000 restaurants and approximately 580,000 reviews. These establishments are spread across several US states (Arizona, California, Florida, Illinois, Nevada, New Jersey, Pennsylvania, Texas) and one Canadian province (Alberta). This broad geographic coverage strengthens external validity by capturing a diverse range of markets and customer bases. We account for regional differences by including fixed effects for city and state in our models, so that comparisons are effectively within the same locale. The Yelp Open Dataset is intended for research use, so our data collection and usage comply with Yelp’s licensing guidelines (the data are publicly provided via Yelp’s research portal for projects like this).

It is important to note that Yelp reviews are written by self-selected customers, not a random sample of all diners. Thus, our dataset reflects the opinions of those who choose to post online reviews, which may introduce some bias. For example, individuals with extremely positive or negative experiences (or more tech-savvy/socially active customers) are more likely to contribute reviews, whereas the “silent majority” of patrons who don’t post online are not directly observed. Our findings should therefore be interpreted as conditional on the reviewing population. We cannot capture the sentiment of customers who remain offline, so if their experiences systematically differ, that is an unobserved segment. Additionally, Yelp employs an automated recommendation algorithm to filter out reviews deemed less reliable or potentially fake. The Yelp Open Dataset includes primarily these “recommended” (verified) reviews, i.e. the ones that pass Yelp’s quality criteria and appear on business pages. Focusing on recommended reviews enhances data credibility (by excluding suspected spam or biased entries), but it also means that certain legitimate reviews (for instance, those from very new users with only one review) might be omitted by Yelp’s system. In short, our sample consists of credible consumer opinions that were publicly posted and recommended by Yelp’s filters. This provides confidence in the authenticity of the data, though it narrows our window to the more active and trusted Yelp users. We acknowledge that the views of less-active diners or those who don’t use Yelp at all are not represented, which is a limitation common to studies using online review data.

Each business listing in our data comes with a rich set of binary attributes indicating the presence of various features – for example, whether the restaurant offers delivery, has outdoor seating, allows credit card payment, serves alcohol, provides live music, etc. These attributes are part of the Yelp business profile and can be considered verified features (typically provided or confirmed by the business owner or via Yelp’s processes, rather than inferred from reviews). In total, we have nearly 70 such attributes (see Table 8). Many of these features tend to co-occur: for instance, a restaurant that offers one convenience (say, table service) often offers others (like reservations), or an upscale restaurant might simultaneously provide multiple amenities. We observe high pairwise correlations among some raw features (often above 0.4) and correspondingly large variance inflation factors when including all features at once. To ensure a parsimonious analysis and avoid multicollinearity, we group these attributes into seven broader categories. Specifically, the feature families are: Services, Ambience, Amenities, Alcohol, Payment, Menu Breadth, and Entertainment. This grouping is informed by subject-matter logic and supported by a principal components analysis that suggested natural clusters of related features. Using aggregated categories both reduces noise and aligns with the economic concept that consumers evaluate bundles of features (Lancaster’s characteristics model). For each category, we construct two indicators: (1) an “Any” dummy that equals 1 if the restaurant has at least one feature in that category, and 0 if none; and (2) a “High” dummy that equals 1 if the restaurant offers more features in that category than the sample average (i.e. a high concentration of features in that group). These variables let us test threshold effects (does having any feature in a category make a difference?) and diminishing returns (do a lot of features in the same category add value, or possibly even dilute quality?). For example, over 90% of restaurants in our sample offer at least one Service convenience (such as takeout or delivery), but fewer than 20% qualify as High Ambience (meaning they offer an above-average number of ambience-related features like decor, lighting, music). This indicates that basic services are ubiquitous, whereas an intensive focus on ambience is relatively rare – a pattern that will be relevant in interpreting our results (see Table 1).

In terms of outcome variables (customer satisfaction measures), we leverage both the numerical star ratings and metrics derived from review text sentiment. For each restaurant, we compute the average star rating (mean of all 1–5 star reviews it received in 2019) – this is a conventional measure of overall satisfaction. We then extract two text-based sentiment measures from the review corpus. First, we calculate the restaurant’s average Valence Aware Dictionary and Sentiment Reasoner (VADER) sentiment score by applying the VADER lexicon to each review and then averaging those scores at the restaurant level. VADER produces a continuous sentiment polarity score from −1 (most negative) to +1 (most positive) for each review; thus, the average VADER score summarizes the overall emotional tone of the feedback for that restaurant. Second, we derive an AFINN (a sentiment lexicon used to evaluate emotional tone in text by assigning numerical values to individual words)-based positivity ratio. For each review, we use the AFINN lexicon to assign a numeric sentiment score (summing positive and negative word weights) and classify the review as positive if the net score is above zero or negative if below zero. We then compute the share (percentage) of each restaurant’s reviews that are labeled positive. This positivity share can be interpreted as the probability that a given customer’s experience was positive, according to their review. Together, these three outcomes – average sentiment, positive review fraction, and mean star rating – provide a comprehensive view of customer satisfaction. They are strongly correlated in our data (e.g. restaurants with higher star averages almost always show higher VADER sentiment scores), but there are subtle differences: star ratings involve more cognitive evaluation (comparisons and overall judgments), while text sentiment captures emotional tone and specific likes/dislikes. By examining both, we can see whether observable features influence the emotional sentiment of reviews in the same way as the explicit ratings. All reviews in the dataset are timestamped within 2019, and the text is predominantly in English. We focus on English-language content for sentiment analysis to ensure the accuracy of the VADER/AFINN tools (non-English reviews, which are a small minority, were omitted from sentiment calculations to avoid misclassification).

Figure 1 plots average VADER polarity against mean ratings. Points lie on a steep upward curve: ratings below three align with near-neutral sentiment, while ratings above four approach the +0.8 ceiling. The pattern confirms that the text metrics track the numeric scale and hints at a non-linear jump around the three-rating mark, which motivates the threshold structure of the covariates.

Figure 2 shows a frequency-weighted word cloud of review vocabulary after stop-word removal. The largest tokens—“food,” “service,” “staff,” “great,” “friendly”—signal what diners stress most. Words tied to secondary features, such as “parking” or “music,” appear smaller, mirroring the smaller coefficients found later for amenities. The visual supports the decision to treat services and ambience as core groups.

Figure 3 displays a kernel-density map of restaurant locations. Density peaks in Los Angeles, the Bay Area, New York, Chicago, Houston, and Miami, while rural tracts remain light. Such clustering implies unobserved local factors—competition, income mix, tourism—that could bias estimates if ignored. City and state fixed effects neutralize these regional shocks, forcing identification from within-market variation in feature bundles.

We acknowledge a limitation in our dataset regarding the social context of dining experiences. Yelp reviews do not record whether a customer dined alone, with a partner, or in a group, nor the occasion of the visit. This is important because group dynamics or special occasions could influence satisfaction independently of restaurant attributes. For example, a birthday dinner with friends might yield a more positive review due to the celebratory mood, or conversely, serving a large group might strain the restaurant’s service and affect the feedback. Because we lack data on party size or context, we cannot control for these factors. This unobserved heterogeneity is absorbed into the error term of our models. We assume that social context effects are idiosyncratic and not systematically correlated with the features we study (for instance, it seems unlikely that only restaurants with live music always have group diners, or similar). Thus, while social context could add noise to our sentiment and rating measures, it should not bias our estimated coefficients if its occurrence is essentially random with respect to the presence of specific features. We flag this as a caveat: part of the unexplained variance in customer satisfaction might be due to social-party effects that our analysis is unable to capture.

Our analysis covers a broad spectrum of restaurant types – from quick-service and casual eateries to formal fine-dining establishments – all pooled together. We do not explicitly segment the sample by restaurant category, cuisine, or price tier in the regressions. Instead, we include all restaurants in one unified analysis, using fixed effects and large sample size to account for differences. This approach assumes that the relationship between features and satisfaction, while possibly varying in magnitude, is directionally similar across different restaurant segments. It provides an overall industry-wide perspective. We recognize that in reality, customer expectations and the marginal impact of certain features might differ by segment. For instance, live entertainment might be a welcome feature in a bar or family restaurant but could be viewed as a distraction in an upscale fine-dining context; similarly, kid-friendly amenities are irrelevant for a romantic boutique bistro but crucial for a family café. Because we do not have a variable labeling each restaurant as “fine-dining” vs “casual” (beyond what might be inferred from price or category keywords), we rely on the breadth of our sample to average out these differences. The city and state fixed effects soak up some location-specific aspects of restaurant style (urban areas might have more fine-dining, etc.), but they do not directly control for within-city variation in segment. Thus, our results should be viewed as average effects that apply in a general sense. In the discussion, we note that an interesting extension would be to analyze segments separately – for example, running the same regressions on subsets like high-end restaurants versus budget eateries – to see if the patterns hold consistently. For now, maintaining a unified sample increases statistical power and allows us to identify broad trends that are not overly driven by one category of restaurant.

One specific aspect of restaurant heterogeneity worth discussing is price level. Yelp provides a price range indicator for each business (commonly denoted by $, $$, $$$, etc.), which reflects the approximate cost per person. We did not include the price category as a control variable in our main regressions because data on price is limited. Our focus is on how tangible features (services, amenities, etc.) correlate with satisfaction, and including price could confound that relationship. Price and features are often correlated – e.g. upscale restaurants (higher price category) tend to offer more amenities and might also generally receive higher ratings due to service quality and ambiance that come with the higher cost. If we controlled for price, we might mask some of the effect of interest (since price itself can influence satisfaction and is intertwined with the level of service a restaurant provides). Moreover, the Yelp price tiers are broad and do not capture nuances of value-for-money or customer income, making them a blunt instrument. However, we acknowledge that price influences customer expectations and satisfaction. Customers typically tolerate fewer shortcomings at high-priced venues and might be more forgiving at cheaper eateries; perceived price fairness can affect satisfaction and loyalty. Because we omit price, our feature coefficients should be interpreted as unconditional correlations that partly reflect any price-related effects. We assume that by comparing restaurants within the same city (with city fixed effects), a portion of price variation is implicitly controlled – for example, New York City NYC has different price standards than a small town, but within NYC our comparison is between restaurants facing similar cost-of-living and customer income conditions. Still, there could be residual bias if, say, even within a city, higher-end restaurants both have more features and draw more praise. Our large and varied sample mitigates this to an extent (since it includes many low-, mid-, and high-price venues in each city). We emphasize this as a limitation: lacking direct price controls means we cannot separate the effect of a feature from the premium service context in which it often exists. Future research or robustness checks could incorporate the price tier or average meal cost to examine how controlling for price impacts the feature-satisfaction relationships. For our study, we maintain price as an implicit factor and focus on operational features, with the understanding that our results capture how these features matter in practice given the restaurants’ market positioning.

We employ an econometric regression approach to quantify the relationship between restaurant features and customer satisfaction outcomes. Given the nature of our dependent variables, we use two types of models: ordinary least squares (OLS) for the continuous outcomes and logistic regression for the binary outcome. Specifically, we run OLS regressions when the dependent variable is either the restaurant’s average VADER sentiment score or its mean star rating (both treated as continuous measures). When the dependent variable is the AFINN-based positivity (the fraction of positive reviews), we model it in a logistic framework by considering each restaurant’s positive vs negative review count, effectively analyzing the log-odds of a review being positive. In all specifications, we include fixed effects for city and state. This controls for unobserved regional characteristics such as local taste preferences, urban/rural differences, or regional trends in using the Yelp platform. The fixed effects ensure that identification comes from within-area variation: for example, when estimating the effect of offering outdoor seating, we are comparing restaurants in the same city/state – some with outdoor seating, some without – rather than comparing a sunny California locale to a colder climate where outdoor seating is less common. By design, this differences-out broad regional confounders. We also control for each restaurant’s review count (the total number of reviews in 2019) in the regressions. This variable acts as a proxy for the restaurant’s popularity or customer traffic. It’s conceivable that popular restaurants both attract more reviews and potentially receive different kinds of feedback (either harsher scrutiny or a dilution of extreme opinions as more people contribute). Including review count helps adjust for a potential selection effect: a restaurant with 500 reviews might naturally have a more “regression to the mean” average rating than a place with just 5 passionate reviews. Finally, we compute robust standard errors clustered at the city level for all models. Clustering by city means we allow for the error terms of restaurants in the same city to be correlated (accounting for any city-specific shocks or influences that hit multiple restaurants similarly). This yields more conservative standard errors and ensures that our inference is not overly optimistic in the presence of within-city correlations.

A central part of our methodology is converting unstructured review text into quantitative sentiment metrics. We use two complementary lexicon-based tools for sentiment analysis – VADER and AFINN – chosen for their accuracy and relevance in an economic context of consumer reviews. VADER (Valence Aware Dictionary and Sentiment Reasoner) is a rule-based sentiment analysis tool specifically attuned to social media and consumer review language. It provides a continuous sentiment score for each piece of text on a scale from −1 (extremely negative) to +1 (extremely positive). VADER is ideal for our purposes because it can handle nuances like sarcasm, slang, negations, and intensity (for example, it knows that “good” vs “GOOD!!” conveys different strengths of positivity). We applied VADER to every review to obtain an individual sentiment score, and then took the average per restaurant as described. AFINN is a lexicon developed by Nielsen (2011) that assigns a polarity score (positive or negative integer) to many English words. We use AFINN to derive a binary sentiment indicator for each review: by summing the scores of all sentiment-bearing words in a review, we get a net score, and we label the review positive if the net score is above zero and negative if it is below zero (reviews with a net score of zero, indicating neutral sentiment or perfectly balanced positive/negative words, are relatively rare and were excluded from this particular measure). After this classification, we calculate the proportion of positive reviews for each restaurant. The choice of these two tools was motivated by both methodological and economic considerations. Methodologically, VADER and AFINN have been shown to perform well on consumer review data and are widely used in economics and business research to gauge sentiment. They each have advantages: VADER’s continuous score captures gradations of sentiment (useful for detecting subtle improvements or declines in satisfaction), while AFINN’s binary classification offers a clear-cut rate of positive experiences (which can be intuitively linked to probabilities of customer satisfaction). Economically, these measures map onto the concept of utility or satisfaction derived from an experience – higher sentiment or a greater chance of a positive review suggests higher utility. We deliberately selected lexicons that yield interpretable metrics: an increase in average VADER score or in the positive review fraction can be directly read as an increase in customer satisfaction. Alternative lexicons like Bing (Liu) or National Research Council Canada Lexicon (NRC) could have been used, but they either provide only categorical positive/negative tags similar to AFINN or focus on specific emotions (joy, anger, trust, etc.) which, while interesting, are less straightforward to aggregate into an overall satisfaction measure. By using VADER and AFINN, we capture overall sentiment in a way that aligns with economic interpretations of consumer satisfaction (i.e. how positively consumers speak about their experience) and we leverage tools that have proven accuracy in the review context (Hutto and Gilbert, 2014, for VADER’s validation).

All textual data were preprocessed before sentiment analysis to improve accuracy. We filtered out non-English reviews so that language nuances would not confound the sentiment tools (the vast majority of reviews in our sample were in English). We then performed basic cleaning: lowercasing text, removing uniform resource locator links or markup, and correcting obvious misspellings when necessary. We did not remove standard stop words or punctuation wholesale, because many such elements carry sentiment information. In fact, VADER accounts for punctuation emphasis (e.g. exclamation points increase intensity) and treats negation words (“not”, “never”) specially by flipping sentiment polarity that follows. We wanted to preserve these cues. Similarly, we kept phrases intact rather than stemming or lemmatizing aggressively, since phrases like “not bad” or “would highly recommend” need to be read in context by the algorithm. VADER’s rule-based engine parses each review with these considerations, ensuring that a phrase like “terribly amazing” is interpreted in context – “terribly” in this case would be read as an intensifier that amplifies “amazing” (resulting in a strongly positive sentiment), rather than as a separate negative remark. Thanks to such rules, VADER can handle mixed sentiment constructs and sarcasm to a degree. On the other hand, the AFINN approach is a straightforward sum of word sentiments. In a phrase like “terribly amazing,” AFINN would assign a negative score to “terribly” and a positive score to “amazing.” When summed, the net score would likely still be positive (assuming “amazing” carries more weight), so the review would be classified as positive overall. In cases where positive and negative words cancel out exactly, the review is effectively neutral, and those instances we set aside as neither positive nor negative. By using both methods, we cross-validate our sentiment extraction – if a restaurant has a high average VADER score, it almost invariably also has a high positive-review ratio, and vice versa, which we indeed observe. The sentiment processing pipeline thus involves: (1) language filtering and minimal text cleaning; (2) applying VADER to get a nuanced average sentiment; (3) applying AFINN to categorize reviews and get a positive percentage; and (4) aligning these with the star ratings. This multi-pronged approach ensures that our sentiment measures are robust and capture the underlying customer satisfaction from different angles.

As a robustness check, we also include a third outcome measure: the average star rating, the traditional 1-to-5 scale commonly used on review platforms. We compute the mean star rating for each restaurant (denoted mean_stars_x) and treat it as a complementary measure of satisfaction. This check allows us to compare results from sentiment-based analyses to a more explicit and commonly understood customer satisfaction metric.

Our key predictors are restaurant attributes drawn from Yelp business profiles. We organize 70 binary attributes into seven broad categories reflecting distinct facets of the restaurant (see Table 8 for the full list). For each category, we construct two types of measures to capture the restaurant’s feature offerings. The first is an “Any Feature” dummy (Or logic) that equals (1) if the restaurant offers at least one attribute in that category, and (0) if it provides none. This indicator tests for threshold effects – whether having any feature in a category (versus having none) boosts customer satisfaction. The second measure is a “High Feature Count” dummy (Mean logic) indicating whether the restaurant offers an above-average number of attributes in that category. We calculate the share of possible features present in each category for each restaurant, and assign (1) if this share exceeds the sample mean, and (0) if it is below average. This captures whether extensive feature offerings in a category (beyond the typical restaurant’s offerings) yield additional satisfaction benefits. In the regression models, the High-feature dummy measures the marginal effect of being a feature-rich restaurant (relative to the baseline of a below-average offering) in that category.

We tailor the regression model to the nature of each dependent variable. For the continuous outcomes – the average VADER sentiment score and the mean star rating – we estimate a linear OLS regression with city and state fixed effects. The OLS coefficients on the attribute variables can be interpreted as the marginal impact of those features on the sentiment score or star rating, holding location and other factors constant:

where $X$ the seven attribute categories are represented, and $Reviewsi$ is the number of reviews. The error term $εi$ is assumed to be i.i.d. normal with mean zero and constant variance.

In contrast, we use a fixed-effects logistic regression for the binary sentiment outcome (positive vs negative review tone). The logistic model estimates how restaurant attributes affect the log-odds of a review positively, again including city and state fixed effects. This specification is appropriate for the dichotomous nature of the AFINN sentiment label and ensures predicted probabilities remain in the [0,1] range:

With the observed outcome rule $yi=1$ if $yi*>0$⁠, and 0 otherwise. This yields the probability model:

This specification allows us to account for unobserved heterogeneity at the city and state levels through fixed effects $γc[i]$ and $δs[i]$⁠, ensuring comparisons across similar geographic markets. Standard errors are clustered at the city level to correct for within-location correlation.

Our identification strategy relies on certain assumptions about the data-generating process, which we address here. First, we assume a linear and additive relationship between the feature variables and the outcomes (or a log-linear form in the logistic case). This means we model the effect of, say, adding outdoor seating as shifting the average sentiment or odds of a positive review by a constant amount, holding other factors fixed. Real-world relationships might deviate from strict linearity – there could be interaction effects between features (perhaps the boost from having live music is larger for restaurants that also have a full bar), or non-linear responses (the tenth added feature could matter less than the first). We partly capture non-linearity by including the “High” feature dummies to test for diminishing returns, as described. These dummies allow a piecewise-linear effect: one jump for going from 0 to some features, and another adjustment for going from a few features to many. This flexibility accounts for threshold effects that we did observe in the data. However, we did not include every possible interaction between feature categories to avoid overfitting and multicollinearity. We implicitly assume that feature effects are roughly independent or that any complementarities are second-order. We tested some interaction terms in preliminary analysis and did not find qualitative changes in core results, but a comprehensive interaction model was infeasible given the dozens of features. Thus, additivity is an approximation; we believe it is reasonable for identifying the primary drivers, especially with the grouping strategy reducing extreme collinearity (post-grouping, variance inflation factors dropped to acceptable levels, all well below 10).

Second, our regression estimates assume conditional independence – that after controlling for the included covariates (features, fixed effects, review count), there are no omitted variables systematically affecting both the presence of features and the review outcomes. In an ideal experiment, features would be randomly assigned to restaurants; in our observational data, we know that’s not true (restaurants deliberately choose their features, and those choices could correlate with other qualities). We have taken steps to mitigate omitted variable bias: the city and state fixed effects soak up unobserved differences across locations (e.g. a city with generally higher ratings or a culture of harsh critiques), and the large number of restaurants provides variation such that we can compare like with like as much as possible. Nonetheless, some unobserved heterogeneity remains. For instance, consider restaurant quality or management competence, which is hard to measure but surely influences customer satisfaction. It’s plausible that higher-quality restaurants are both more likely to invest in certain features and get better reviews. If so, our coefficients on those features could be upward-biased because we might be partly capturing the effect of underlying quality. We attempted to address this in two ways: (1) including the restaurant’s review count as a control – this can indirectly capture a restaurant’s popularity and longevity, which often correlate with quality reputation; and (2) leveraging the sheer scale and diversity of our data so that, hopefully, the idiosyncratic traits of any single restaurant carry less weight. While these measures help, we cannot conclusively claim causality. We therefore interpret the results as associative: they tell us which features tend to be present when sentiment and ratings are higher, after accounting for observable factors. We are transparent about this limitation. The consistency of results across different subsets of data and alternative model specifications (not reported in detail, but checked) gives us confidence that the associations are robust, but readers should keep in mind that unmeasured factors (like food quality, chef talent, or local competition) could be influencing both features and outcomes. We suspect that any such bias would need to be quite strong to overturn our main findings, given the pattern of threshold and diminishing returns effects we observe (it’s unlikely that unobserved quality would coincidentally create a non-linear jump then plateau pattern exactly aligned with feature counts). Still, caution is warranted in making causal interpretations.

Third, we assume the error terms are independent across restaurants, or at least uncorrelated after accounting for the fixed effects structure. There are reasons this might be violated: restaurants in the same area might all be impacted by a local event or trend (for example, a city-wide food festival week might temporarily boost all reviews), or there could be correlated shocks (like a regional power outage affecting many businesses on a given day). By clustering at the city level, we have explicitly allowed the errors for restaurants within each city to be correlated in an arbitrary way. This adjustment should capture most local spillovers or shared influences, and it aligns with our use of city fixed effects (both assume the city is the relevant cluster of similarity). We did consider clustering at broader levels (state) or narrower (zipcode), and results were materially unchanged – city proved a reasonable middle ground. The independence assumption is more likely to hold between cities (Chicago diners’ review patterns aren’t directly influencing Miami’s, for example). Another aspect of the error structure is heteroskedasticity (non-constant variance of errors). With a dataset of this size and variety, heteroskedasticity is almost certain – some restaurants (especially those with very few reviews near our cutoff of 5) will have more variability in their average sentiment than those with hundreds of reviews. We address this by using robust standard errors, which are valid even if error variance differs by observation. Our inferences thus do not rely on homoskedastic errors.

Finally, the large sample size (tens of thousands of observations) strengthens the reliability of our estimates. With such a large sample size, the Central Limit Theorem assures that even if error terms are not perfectly normally distributed, our OLS estimates will be approximately normally distributed and our hypothesis tests will be reliable. The rich dataset also allows us to include many controls and fixed effects without running out of degrees of freedom, and to detect even modest effect sizes with statistical significance. The results remained stable under alternative modeling choices (e.g. using a logit vs OLS for the star rating banded into high/low, or using a Poisson pseudo-likelihood for count of positive reviews – methods not reported in tables for brevity). These robustness checks increase our confidence that the model’s core assumptions are not flagrantly violated. In sum, while the fixed-effects OLS/logistic framework has its limitations, we have taken care to meet its assumptions where possible and to be open about where assumptions may be stretched. The combination of theoretical justification (utility theory and prior literature), data checks, and robustness tests suggests that our methodology is appropriate for uncovering the determinants of online customer reviews in this context. The next section presents our findings under this framework, noting how the evidence aligns with the hypothesized threshold effects and diminishing returns in restaurant features.

Tables 2 and 3 present the fixed-effects regression results for review text sentiment as measured by the VADER lexicon (where sentiment is a continuous score from −1 to +1). Table 2 (High logic) shows that core service features are the primary driver of positive sentiment. Restaurants offering an above-average number of service-related features – for instance, having delivery, table service, catering, or reservations – exhibit significantly higher average sentiment scores than those lacking in service breadth. The coefficient on the service-high dummy is about +0.02 (p < 0.10), meaning that going from a basic to an above-average service offering raises the sentiment score by ∼0.02 points. While this effect may appear small in absolute terms, it is meaningful given the sentiment scale and the large sample, and it signals that customers reward restaurants for reducing friction and increasing convenience (for example, the ease of getting food delivered or securing a table). This finding aligns with Lancaster’s characteristics model of utility – consumers derive real utility from these key service attributes – and echoes the SERVQUAL literature emphasizing reliable, responsive service as a core satisfaction dimension (Parasuraman et al., 1988).

By contrast, no other category being “above-average” produces a significant sentiment boost. Having an above-average number of amenities, ambience features, payment options, menu items, or entertainment options does not appreciably increase the sentiment score. The coefficients for those high-feature dummies are near zero and statistically insignificant. This suggests diminishing marginal utility for features beyond the basics. Once a restaurant meets a baseline level of expected offerings (the core services and a decent environment), additional frills do not translate into additional emotional enthusiasm from customers. Essentially, consumers act as satisficers: beyond the threshold of “good enough,” extra features yield negligible returns in sentiment. This pattern is consistent with the diminishing returns observed by Ryu and Han (2011) in restaurant upgrades, and it reflects the idea of hygiene factors in service quality (Kano, 1984). Certain attributes (e.g. basic service, cleanliness, a minimum comfortable ambience) are expected – their absence will hurt satisfaction, but exceeding the expectation does not proportionally help once the need is met. One minor point of interest is the menu breadth dummy, which is positive but not significant in Table 2. A wide menu by itself doesn’t boost sentiment, and this could hint at a subtle trade-off noted in prior studies: offering too many options might overwhelm or dilute perceived specialty (Iyengar and Lepper, 2000; Johns et al., 2013). In sum, under the High logic we find that only major service upgrades reliably lift sentiment, whereas piling on extra features yields virtually no further benefit and can even risk mild drawbacks (as seen with very large menus or intense focus on alcohol, discussed later).

Table 3 (Or logic) reinforces these conclusions by testing threshold effects. Here we code whether a restaurant has any feature in a given category (versus none). We find strong threshold effects for core services: simply having at least one core service feature (e.g. offering either delivery or takeout or reservations, when previously offering none) is associated with an increase of about +0.024 in the sentiment score (p < 0.05). In practical terms, going from zero to one basic service yields a noticeable uptick in customer sentiment, underscoring the importance of meeting fundamental consumer expectations. Ambience shows a smaller threshold effect: having any dedicated ambience feature (say, a patio or special décor) adds roughly +0.01 to sentiment on average, significant at the 10% level. This modest gain suggests that a pleasant atmosphere does contribute to a better experience, though its impact is weaker than that of core services. These findings resonate with the DINESCAPE framework (Bitner, 1992; Ryu and Jang, 2008), where the physical environment enhances experience, but here we see it takes at least some effort in ambience to make a perceptible difference. Meanwhile, other categories show no significant threshold gains: simply offering one amenity (versus none), one entertainment option, or one special payment method yields no significant change in sentiment. Such features operate as background satisfiers – nice-to-have options that diners may take for granted. Providing, for example, a single extra amenity like free Wi-Fi doesn’t automatically elevate the customer’s mood in reviews. In fact, the only notable non-core effect under Or logic is a negative coefficient on the Payment feature dummy: having any unconventional payment option (beyond standard credit/cash) slightly reduces sentiment (around −0.01 to −0.02, p < 0.10, as implied by the text). Although marginal, this negative effect is consistent across our analyses and suggests that overly complex or novel payment systems might irritate some customers or signal a focus away from the dining experience. It reinforces the idea that deviating from “tried-and-true” basics in some peripheral ways (like an overload of payment technology or promotions of alcohol) can backfire.

Taken together, the VADER sentiment results demonstrate two key points: threshold benefits (especially for core services) and diminishing returns beyond the threshold. Moving from “zero to one” in critical service areas yields a meaningful improvement in customer sentiment, whereas moving from “some to many” features yields virtually no additional joy. This empirically validates the theoretical expectation that consumers gain high utility from the first unit of a crucial attribute but very little from subsequent units. It also provides a textual evidence base for the notion that restaurants should nail the basics before adding extras. In essence, customers writing reviews clearly reward restaurants that remove major pain points (e.g. lack of delivery or slow service) and provide a comfortable environment, but they do not disproportionately praise those that go from a decent amount of extras to a great many extras.

We next examine the determinants of a review being positive in tone using the AFINN lexicon-based measure. Instead of a continuous sentiment score, this approach classifies each review as positive or negative based on the net valence of words, and we model the log-odds of a positive review using logistic regressions (Tables 4 and 5). Despite the different scale, the AFINN results confirm the earlier patterns.

Under the High logic (Table 4), most feature categories again show little to no effect on the likelihood of a positive review. The one notable exception is Amenities: having an above-average number of amenity-related features (comfort and convenience extras like free Wi-Fi, parking, kid-friendly options) correlates with a significant increase in positivity odds. Specifically, the amenities-high dummy carries a log-odds of roughly +0.23, which translates to about a 20–25% higher odds of a review being positive. This suggests that when a bundle of comfort features is provided together, it can collectively enhance customers’ satisfaction enough to push some experiences from negative toward positive. Notably, this cumulative amenity effect did not appear in the VADER continuous sentiment, implying that while extra amenities might not elevate the average tone of all reviews, they can make the difference in tipping a borderline experience into “positive” territory. In other words, a robust package of amenities acts as a hygiene factor – largely unnoticed until absent, but when present in sufficient volume it ensures more customers leave happy rather than disgruntled. This finding aligns with Kano’s model (1984) and the idea in service research that a threshold of comforts must be met to avoid dissatisfaction. By contrast, focusing too much on alcohol-related offerings shows a significant negative effect: the alcohol-high dummy is around −0.26 in log-odds (p < 0.01), meaning restaurants that heavily emphasize alcohol (e.g. extensive bar menus or a bar-like atmosphere) tend to receive fewer positive-toned reviews. This supports a “disamenity” interpretation– an ambience dominated by alcohol may alienate certain diners or signal a mismatch for families, echoing concerns by Ryu and Jang (2008) that a loud bar setting erodes utility for diners seeking a calm meal. No other “High” category significantly shifts positivity odds, again indicating that above-average service or menu breadth alone doesn’t guarantee more positive reviews. This reinforces the diminishing returns idea: once a baseline is exceeded, more service or menu variety doesn’t further convert negative experiences to positive, likely because customers already expect a good level of service and adequate menu choices as standard (any additional beyond that is not a delight, just “okay, they have a lot of options”).

Under the Or logic (Table 5), meeting core service thresholds has a pronounced impact on review positivity. Simply offering at least one basic service (versus none) increases the odds of a positive review by about 26% (log-odds ≈ +0.26, p < 0.01). This is a sizeable effect in the context of binary outcomes – it implies many customers who would have written a neutral or negative review might turn positive if the restaurant provides a key service like delivery or takeout. This threshold jump mirrors our earlier sentiment result and underscores Lancaster’s utility model: the first unit of a needed service yields high utility. We also see a smaller but suggestive boost from Ambience (threshold): having any notable ambience feature slightly raises positive-review odds (on the order of a few percentage points, p < 0.10). Even one appealing atmospheric element – say a nicely decorated interior or pleasant music – can nudge some customers to feel positive about their experience, although the effect is modest. On the flip side, adding any special Payment option lowers the odds of a positive review (log-odds ≈ −0.19, p < 0.10). While this effect is only marginally significant, it consistently appears negative across models. It suggests that introducing non-standard payment methods (for example, cryptocurrency acceptance or proprietary app payments) might subtly detract from customer satisfaction. One interpretation is that it could introduce complexity or signal a tech-forward focus that isn’t matched by flawless execution, thereby frustrating some guests. This finding echoes Luca (2016) in cautioning that more options are not always better – extra features can distract or even negatively signal priorities. Apart from these, the threshold dummies for amenities, entertainment, and menu breadth are insignificant, reinforcing that just having one such feature isn’t enough to sway customer opinions. A lone extra like a TV for entertainment or a single vegan option on a large menu doesn’t meaningfully change the fraction of happy customers; these things might only matter if numerous and excellent, as the high-amenities result indicated.

We acknowledge that the adjusted R² values in our regressions are relatively low, often in the single digits. However, this is expected and well-documented in studies involving large cross-sectional datasets of consumer reviews, where much of the variation is inherently idiosyncratic or driven by unobserved factors. What strengthens our conclusions is not high explanatory power, but the consistent and statistically significant effects of key predictors across multiple models. The fact that core service and ambience features repeatedly emerge as meaningful correlates of satisfaction—even in the presence of considerable noise—underscores the reliability of our findings. In this context, the low R² values reflect the complex and subjective nature of review data and are consistent with benchmark results in the existing literature.

In summary, both sentiment analyses tell a consistent story. Core services and, to a lesser extent, ambience are what move the needle on customer satisfaction, especially by ensuring a restaurant clears a basic quality threshold. Peripheral extras (entertainment gadgets, expansive menus, exotic payment methods) have negligible impact once those basics are met. We see evidence of threshold effects (“going from nothing to something” yields a boost) and diminishing returns (“going from something to a lot” yields little or even backfires). This robustness across two different sentiment metrics strengthens our confidence in the results. The convergence between VADER’s continuous sentiment scores and AFINN’s positive/negative classifications suggests that these are genuine patterns in the underlying consumer experience, not artifacts of any one measurement approach.

For restaurant owners and managers, the implications are direct and actionable. The best way to improve customer satisfaction and build a strong online reputation is to excel in the fundamentals. Our results consistently indicate that investing in reliable, responsive service and maintaining a pleasant ambience yields the highest returns in terms of positive reviews and ratings. These are the areas where improvements translate into appreciable gains in customer sentiment. Managers might thus consider allocating more resources to staff training, speed of service, cleanliness, and comfort – the basic elements of a good dining experience. Importantly, doing so is not just theory but quantitatively supported: for example, introducing a core service like delivery can lift average ratings by a significant fraction of a star, which in turn can increase customer traffic and revenue. On the other hand, simply adding more low-impact features will not help once the basics are covered. Our findings warn against the instinct to compete by offering every conceivable extra (be it an overly expansive menu, exotic payment options, or niche amenities). Such additions have little effect on satisfaction after a point, and in some cases they even carry subtle downsides (diminishing returns or confusion). In economic terms, resources spent on those extras have an opportunity cost – they could have been spent on core quality improvements that customers actually notice. Strategically, this means a restaurant that “does a few things very well” is likely to garner better reviews than one that “does many things okay.” This echoes the idea of focusing on core competencies: a burger joint that makes outstanding burgers and provides quick, friendly service will likely outshine a diner that offers burgers, sushi, pizzas, and more but excels at none. It also resonates with the concept of signaling in consumer markets. By doubling down on fundamental quality, restaurants send a signal of competence and reliability to customers. For instance, a restaurant known for consistently great service signals that it values the guest experience, thereby attracting positive word-of-mouth and reviews. In contrast, a restaurant advertising a long list of novelty offerings might signal a lack of clear focus, potentially eroding consumer trust. In summary, when it comes to driving satisfaction, quality beats quantity of features. Managers should thus innovate and differentiate primarily in areas of core service delivery and customer experience, rather than an arms race of menu items or gimmicks.

Our findings also carry implications for digital platforms and policymakers interested in the design of online feedback systems and consumer welfare. Review platforms like Yelp may consider evolving their design to better emphasize the information that truly matters for consumer decision-making. Currently, business profile pages often display a long checklist of attributes (“Has Wi-Fi,” “Offers Parking,” “Accepts Credit Cards,” etc.) with equal weight. However, our results suggest that core indicators (e.g. does the restaurant offer basic services like delivery or reservations? does it have a comfortable ambience?) have an outsized influence on customer satisfaction, whereas many other details do not. Platforms could leverage this insight by highlighting core service attributes more prominently – for example, using badges or icons for “Delivery Available” or “Outdoor Seating” at the top of the page, while relegating less impactful attributes to a secondary list. They might also refine search and recommendation algorithms to reward restaurants that meet core quality thresholds. For instance, a platform could incorporate a “core service score” that boosts businesses which provide the essential features most valued by customers. This would nudge the market toward improving fundamentals, ultimately enhancing consumer welfare. Additionally, digital feedback systems could integrate sentiment analysis to identify what aspects of the experience drive satisfaction, possibly giving restaurants qualitative feedback on their service and ambience from review text. On the policy side, regulators concerned with consumer protection and fair competition should note that more is not always better in terms of information disclosure. Encouraging or requiring businesses to simply list countless features might overwhelm consumers or create a false impression that the quantity of amenities equals quality. Instead, regulators (and consumer information platforms) might work to ensure that the truly critical aspects – health and safety, basic service availability, accessibility – are clearly communicated to consumers. By doing so, they help level the playing field: a small family-run restaurant that excels at basic service and food quality can compete on more equal footing with a large chain that might tick more boxes on a checklist but doesn’t deliver on core experience. In broader societal terms, recognizing the primacy of basics could lead to a fairer marketplace. If both consumers and platforms start to emphasize fundamentals over frills, businesses won’t feel compelled to engage in feature bloat; instead, they can focus on improving what really matters to everyone. This not only improves average customer satisfaction but also lowers unnecessary costs (which could translate into better prices or wages) and reduces decision fatigue for consumers. Finally, from a public policy perspective, our findings suggest that any initiatives to support local businesses (through training, grants, etc.) would do well to target enhancements in service quality and ambience – the areas with proven payoff – rather than encouraging superficial upgrades.

While our study provides robust evidence on the determinants of online reviews, it has certain limitations that open avenues for future work. First, we examined one platform (Yelp) in a specific timeframe (2019). This cross-sectional approach helped control for temporal shocks and platform-specific effects, but it remains to be seen whether the patterns hold in other online review environments or over longer periods. Future research could use panel data to perform event studies – for example, tracking whether a restaurant’s reviews improve after it adds a new service like delivery or undergoes a renovation. Such analyses would strengthen causal inference by observing within-restaurant changes over time. Second, although we include fixed effects for location (city and state) and controlled for restaurant popularity (review counts), there may be observed restaurant characteristics that influence reviews. Factors like brand reputation, chef quality, or local competition weren’t directly observed in our feature data. These could be correlated with both the presence of certain features and customer satisfaction. For instance, upscale restaurants might invest in ambience and also generally earn higher ratings due to quality food and service – we partly account for this with price category controls, but not perfectly. Future studies might incorporate more granular controls or utilize experimental designs (e.g. A/B testing of feature changes) to isolate the effects of attributes. Third, our measure of “food quality” was proxied imperfectly by menu breadth (number of cuisines or items), which did not show significance. It could be that true food quality is relatively homogeneous in our sample or masked by selection bias (diners choose places that match their taste), or simply that it’s hard to quantify without explicit data. Further research could combine textual analysis techniques (to directly evaluate mentions of food quality in reviews) or include health inspection scores, etc., to better capture the food factor. Despite these limitations, our study provides a foundation for understanding how digital feedback reflects real-world service attributes. We encourage future researchers to build on this work by exploring other sectors (e.g. hotels or retail) for similar threshold dynamics, by testing the generalizability across cultures or regions, and by examining how businesses can optimally invest in improvements. In conclusion, our findings highlight a vital lesson: in the age of online reviews, meeting customers’ basic expectations brilliantly beats exceeding them in trivial ways. Restaurants that focus on the core drivers of satisfaction – great service, a pleasant atmosphere, and of course a good meal – will see the dividends in their reviews. Platforms and policymakers, in turn, should create an environment that rewards quality over quantity of features, ensuring that both businesses and customers thrive in the online review ecosystem.

As a robustness check, we analyzed whether the above patterns hold when using the traditional star rating outcome instead of text-based sentiment. Tables 6 and 7 replicate our regression analysis with the restaurant’s average Yelp star rating as the dependent variable (a 1–5 scale). Despite the different nature of this outcome, the results closely mirror the sentiment findings, indicating that the drivers of textual sentiment are indeed the drivers of overall satisfaction as captured by star ratings.

Under the High logic model for star ratings (Table 6), most feature categories show no significant effect on a restaurant’s mean rating, with one notable exception: Ambience. Restaurants that offer an above-average number of ambience features (truly exceptional atmosphere or décor beyond the norm) have significantly higher average ratings. Specifically, going well above the typical ambience (investing in standout interior design, unique atmosphere, etc.) is associated with roughly a +0.27 star increase in the mean rating (p < 0.10). A quarter-star boost is quite meaningful in aggregate terms: on platforms like Yelp, even a fraction of a star can be the difference that tips a business from, say, 4.2 to 4.5 once rounded, affecting consumer perceptions. This result suggests that exceptional ambience can delight customers enough to be reflected in their star evaluations. In contrast, no other category being above-average provides a statistically significant lift in stars. Even “above-average” service, which did improve sentiment, does not translate into a higher star rating once we account for location and other controls. One interpretation is that consumers expect a baseline of good service; exceeding that baseline may improve their mood (sentiment) but not their formal rating, because they consider it the norm or a basic requirement. Outstanding service might thus be “baked in” to avoid negative reviews (a hygiene factor), rather than something that earns extra stars for being above par. By contrast, ambience seems to function more as an exciter or differentiator: diners consciously reward a restaurant with higher stars when its atmosphere is notably superb, whereas reliable service is taken as a given (only its absence is punished). This interpretation is bolstered by industry observations that higher-end restaurants often distinguish themselves with ambiance and service touches to earn those extra half-stars of approval. It appears that once basics like service are satisfactory for everyone, ambiance becomes a key area where exceeding expectations can still garner incremental praise in star ratings.

Under the Or logic model (Table 7), we find strong threshold effects in star ratings very much akin to those in sentiment. The clearest is for Service: offering any core service feature (versus none at all) is associated with approximately a +0.10 star increase in the average rating (p < 0.01). A tenth of a star may sound small, but in competitive restaurant markets, an advantage of 0.1 in average rating can sway consumer choice, as prior research on online ratings suggests. Over thousands of customers, that little bump reflects a consistent improvement in satisfaction for having met a basic need. This finding underscores that basic service offerings (like delivery, takeout, or reservations) give a restaurant a real competitive edge in aggregate ratings. In addition, having any Ambience feature yields a slight positive bump of about +0.03 stars (p < 0.10). This implies even a single appealing ambience element (an outdoor patio, tasteful decor, etc.) can marginally improve overall ratings. Though the effect is modest, it reinforces that customers do notice and appreciate an enhanced atmosphere, even if it’s just one notable aspect. Conversely – and consistent with the sentiment results – having any special payment option shows a small negative association with ratings (roughly −0.07 stars, p < 0.10). In other words, restaurants that try to be too fancy or cutting-edge with payment technology might see their average rating slightly dip. While this negative effect is not large in absolute terms, it is statistically consistent with the idea that certain “over-engineered” conveniences add friction or unintended signals (for example, a restaurant that advertises every conceivable payment method might be perceived as gimmicky or might introduce checkout hassles, detracting from the overall experience). Meanwhile, threshold indicators for Amenities, Entertainment, and Menu breadth remain insignificant in the star models. Simply offering one amenity like free Wi-Fi, one entertainment option, or a marginal expansion of the menu does not move the average rating needle. This confirms that customers don’t reward these individual extras with higher stars – at least not unless they accumulate to an exceptional level (as seen with the high-amenities log-odds earlier, which affected sentiment more than stars).

Overall, the star rating robustness check paints a consistent picture: focusing on core services and a basic pleasant atmosphere yields tangible benefits in both sentiment and stars, whereas layering on peripheral features provides little to no benefit (and can even be slightly detrimental). The fact that Tables 6 and 7 produce very similar patterns to the sentiment-based Tables 2 and 5 lends additional credence to our findings. It indicates that our conclusions are not an artifact of how we measured sentiment – they reflect genuine drivers of customer satisfaction that manifest in both what customers say and how they rate their experience. This convergence across metrics increases our confidence in the robustness of the results. From a managerial standpoint, it means strategies to improve Yelp performance should concentrate on getting the basics right and hitting key thresholds (ensuring at least some service options are available, maintaining a decent ambience) rather than overloading on minor features. The star-rating evidence supports the view that once fundamental expectations are met, extra features yield diminishing returns in customer evaluations. In fact, adding too much may even backfire slightly – an outcome in line with the classic notion that excessive choice or complexity can undermine the consumer experience (Iyengar and Lepper, 2000). By using both textual sentiment and star ratings, we demonstrate a robust, coherent result: core quality attributes drive positive customer evaluations, and most other attributes either don’t matter much or matter only up to a threshold. This reinforces recent research emphasizing service and atmosphere as critical to restaurant success, and it provides a clear direction for where improvements pay off. (see Table 8).

This research set out to determine which tangible restaurant attributes most effectively lift online customer satisfaction. Using a unique dataset of 580,000 Yelp reviews from 36,000 restaurants, we found a clear hierarchy of what matters to diners. Adding one core service feature – such as offering delivery, takeout, reservations, or table service – significantly improves review sentiment and increases average ratings by around 0.1 stars on the platform. Even a single addition in this domain (for example, starting to offer delivery when previously none was available) yields a noticeable uptick in customer satisfaction metrics. Modest ambience upgrades (e.g. improving décor, adding comfortable seating or pleasant lighting) also provide positive returns, though smaller in magnitude. In contrast, piling on amenities or other extras beyond the basics shows little benefit. Once basic expectations are met, adding more peripheral features – expanding an already large menu, offering every conceivable payment method, or tacking on niche amenities – does not improve customers’ expressed satisfaction, and in a few cases it even correlates with slightly lower ratings or sentiment. These findings empirically support economic theories of consumer utility with threshold effects and diminishing marginal returns. Diners derive the most utility from high-salience attributes that address fundamental needs (food served conveniently, an enjoyable atmosphere), and they experience sharply diminishing returns on additional frills once those needs are satisfied. In essence, going from “none to some” has far greater impact than going from “some to many.”

By connecting structured restaurant features to both text-based sentiment and star ratings, our study contributes new evidence on consumer behavior in experience goods markets. We introduced a scalable method that combines lexicon-based text analysis with fixed-effects regressions, allowing us to quantify nuanced consumer reactions while controlling for confounders. Notably, the agreement between the sentiment results and the star rating results strengthens confidence in our conclusions. Text analysis of reviews, often richer in detail, ended up revealing the same core drivers of satisfaction as the numerical star ratings. This cross-validation suggests that our findings are not an artifact of any one measure; rather, they reflect genuine underlying drivers. To our knowledge, this is the first national-scale evidence linking verified feature data (actual attributes listed on business profiles) to both review text sentiment and aggregate ratings. We also provide the first clear evidence of threshold effects and diminishing returns in online restaurant reviews at scale, showing that consumer feedback patterns conform to longstanding theoretical expectations (Lancaster, 1966; Kano, 1984) when it comes to core vs peripheral product attributes.