More Than Words: The Influence of Affective Content and Linguistic Style Matches in Online Reviews on Conversion Rates
Summary (6 min read)
Introduction
- The influence of affective content and linguistic style matches in online reviews on conversion rates.
- The inherent inseparability of content and style in reviews suggests that insights into linguistic style may extend beyond verbatim content to enhance understanding of their overall impact on customer decision making and thus retail performance.
- In light of these gaps and concerns, the authors aim to provide theoretical and managerial guidance on the influence of textual properties of consumer reviews on online retailers’ conversion rates in three ways.
Conceptual Foundations
- Feldman and Lynch (1988) posit that the relative weight of heuristic inferences, as decision inputs, depends on two context-dependent facets: their relative accessibility and their diagnosticity compared with alternative inputs.
- Existing research has accordingly focused on the diagnosticity of readily extractable, quantifiable customer review information cues, such as quality ratings (Chevalier and Mayzlin 2006), volume (Duan, Gu, and Whinston 2008), and reviewer identity information (e.g., name, location; Forman, Ghose, and Wiesenfeld 2008), as well as on product-related aspects such as product popularity (Zhu and Zhang 2010) and price (Yong 2006).
- There are only approximately 500 function words in the English language, but this deceptively small category comprises roughly 55% of people’s daily word usage and provides insight into conversants’ personalities (Bird, Franklin, and Howard 2002).
- Recent research drawing on communication accommodation theory (Giles and Smith 1979) has also suggested that common ground may be an automatic outcome of LSM, irrespective of content (Ireland and Pennebaker 2010).
Affective Content
- Affect drives evaluation and decision making (Lench, Flores, and Bench 2011).
- Previous theory and research thus suggest that affective cues elicit automatic affective responses, which require few processing resources, emerge rapidly, and guide attitudes and actions (Baumeister et al. 2007; Cohen et al. 2008).
- The authors test this relationship to corroborate previous research and to demonstrate the ability of text analytic techniques to detect theoretically meaningful, well-established relationships.
- Accordingly, the authors posit the following: H1a: There is a quadratic relationship between changes in aggregate positive affective content in a product’s reviews and changes in conversion rate.
- At the extremes of positive affect change, each additional increase should have a smaller impact on conversion rate change.
LSM
- Third, in line with Ireland and Pennebaker (2010), the authors derived the overall LSM score of a particular review by taking the average LSM score across all function words.
- If a review used “despite” four times in a text of 100 words, it would yield an intensity of .04.
- Therefore, the authors obtained a measure of overall LSM for reviews published about a particular product in a given week.
- The authors conducted a pilot study to confirm the validity of the LSM measure, in terms of actually eliciting social identification by the target audience and strengthening the impact of the reviews on purchase intentions.
Setting
- The authors gathered data using automated JavaScripts to access and parse HTML and XML pages describing books available for sale on Amazon.com, the leading electronic retailer.
- The authors chose this research context because of the unique traceability features of customer review information and their influence on conversion behavior on retail sites.
- All information is publicly accessible and updated frequently on the retailer’s webpage.
- Therefore, in addition to the information conveyed through the customer reviews, the authors could collect and control for other product- and review-related information, such as price, review volume, review helpfulness, and advertising, all of which may affect consumers’ purchase decisions.
- Similar data about the conversion rates of websites could be obtained from software packages designed to track or retrace online visitor behavior.
Sample
- The authors initial sample included 641 unique books across all subgenres, released between April 15 and May 5, 2010, which received at least one customer review during the observation period.
- The authors selected only books launched in this time period to ensure that the sample books were in approximately the same stage in their product life cycle.
- To explore how changes in customer reviews influence the conversion behavior dynamics of visiting customers, the authors retrieved conversion rates and customer reviews, along with general product and price information, at weekly intervals for 17 consecutive weeks.
- 92 / Journal of Marketing, January 2013 the authors eliminated 36 books that were unavailable for purchase (out of stock) for some period of time during the data collection.
- In a second stage, two independent coders assigned the books to subgenres using Amazon.
Measurement Development
- The authors obtained conversion behavior data from the conversion rate information provided by Amazon.com, namely, the “What Do Customers Ultimately Buy After Viewing This Item” information, which listed the percentage of customers who bought the product featured on the retail page (Bray and Martin 2011).
- To assess the affective content and LSM of the review texts, the authors then conducted a content analysis of the reviews’ qualitative text comments.
- Content analysis is an increasingly popular method to study user-generated posts online (Singh, Hillmer, and Ze 2011).
- To transform text comments into quantitative data, content analysis uses automated, systematic procedures that ensure the objectivity, reproducibility, and reliability of the data analysis (Chung, Pennebaker, and Fiedler 2007).
- Preparing the data for automated text analysis entails archiving the review texts and converting them into text files, a process that produced several review text files for each product (minimum = 1, maximum = 503).
Fiction
- Originally developed to analyze emotional writing, LIWC dictionaries offer strong, reliable convergence between the dimensions they extract and content ratings performed by human coders (Pennebaker et al. 2007).
- Using word counts for a given text, LIWC calculates the proportion of words that match predefined dictionaries.
- Research into collective settings has examined group similarities using direct consensus models, which take the group average as the preferred mode of aggregation (Bliese 2000).
- The authors results empirically justify the data aggregation of individual linguistic styles in reviews to derive a common “genre” linguistic style; for all the function word variables, the ICC 1 values are significant (F-values, p < .05), ranging from .62 to .94.
Control Measures
- The authors constructed their control variables from observed Amazon.
- In addition to the traditional star rating measure, which offers a five-point product quality rating per review, the authors controlled for the effects of increased review quantity posted in a particular week for each book.
- Because price dynamics influence online purchase decision making (Xinxin and Hitt 2010), the authors also considered the impact of changes in discounts (percentage margin of the original price in week t).
- Using the same measure, Mudambi and Schuff (2010) show that the percentage of people who find a review helpful is related to the diagnosticity of the retail site.
- Finally, the authors collected data about additional advertising of the books, but because there were virtually no such occurrences and the effect was highly insignificant, they did not include this variable in the final model.
Analyses and Results
- To capture the influence of their explanatory variables on changes in product site conversion rates, the authors specified a 94 / Journal of Marketing, January 2013 dynamic panel data model.
- The authors assessed the impact of the preceding changes in the explanatory variables on subsequent changes in the product site’s conversion rates, to reduce potential problems associated with autocorrelation and remove the impact of time-invariant unobservable factors.
- By studying the within-product changes rather than absolute levels, the authors could eliminate observed and unobserved differences between the books, which might influence conversion rates.
- This feature does not change over time, so the authors removed (by statistically controlling for) it with the general method of moments first-difference transformation.
- The initial condition was zero for all products; the launch date, as a first observation, produced a conversion rate of zero.
Overall Model
- The authors also assume that E [uit |t] = 0, where the expectation of change in the error term, given the information More Than Words / 95 TA bL e 3.
- De sc rip tiv e set at time t, is 0, instrumented using the lagged value from the previous week (t – 1); uit is the change in the random error term (the authors excluded fixed effect errors from the model by first differencing); and t denotes the week.
- First, on the basis of past conversion behavior, Amazon.
- Past changes in conversion behavior may influence the amount of helpfulness votes that reviews on that site accumulate in a subsequent week.
- Collectively, the traditional fixed effects estimators therefore must be biased.
Hypothesis Testing
- Table 4 outlines the results of the models.
- Because the authors used first differencing and the lagged values for conversion rate, the sample size for the models fell to 4763 observations (591 books).
- There was insufficient evidence to reject the assumption of no autocorrelation in the differences, so their generalized method estimators likely yielded unbiased and constituent estimates (Arellano and Bond 1991).
- The exogenous variables indicated frequency (percentages of affect-laden positive and negative content words in the review text; match percentage in function words), so the authors mean-centered the variables and calculated the interaction term by multiplying mean-centered variable scores.
Variables Model 1 Model 2 Model 3 Model 4 Model 5
- The authors compared the models by computing the chi-square difference test, which confirmed that the main, interaction, and squared effects added explanatory power to the original Model 1 (p < .01).
- In Model 5, the authors additionally split up affective content into positive and negative affective content.
- The authors use the estimates reported from Model 4, including all hypothesized effects, to discuss their results.
Results
- The negative coefficient of the squared term for positive affective content indicates that the effect tapered off, so in line with H1a, overtly positive changes in the affective content in customer reviews had a smaller positive impact on conversion rate, compared with moderate changes in positive affective tone.
- The coefficients of positive and negative changes in affective content differed significantly (p < .001), and negative changes had stronger impacts on the conversion rate.
- The authors chose values for LSM to be one standard deviation below the mean (low), at the mean , and one standard deviation above the mean (high).
- Changes in review quantity also had significant effects on changes in conversion rates (review quantity = .002, p < .05), in line with prior studies that have indicated that product sales can be explained by review volume (Godes and Mayzlin 2004).
The Combined Impact of Changes to Affective Content and LSM on Conversion Rate
- Changes in helpfulness were weakly significant, positive predictors of subsequent conversion rate changes (helpfulness = .006, p < .10).
- The authors tested the effect of changes in star ratings on subsequent conversion rate changes, but in line with Yong (2006), they found no significant relationship.
- The descriptive information in Table 3 reveals that the average star rating for all products was mildly positive (4.15) and did not change much over time, in line with recent research that suggests star ratings converge to an average within a few weeks (Moe and Trusov 2011).
- In contrast, review texts are nuanced and still provide new and relevant information over time that may color overall product evaluations more than the overall star rating dynamics.
- In contrast, variance in the weekly changes of affective content was not significantly related to subsequent conversion rate changes (affective content variation = .002, p < .411).
Extending Extant Research
- This study contributes to contemporary research on customer reviews by outlining a method to dissect customer review texts to reveal their semantic content and style properties, as well as demonstrating the dynamic influence of text properties on conversion rates in online retail sites.
- Various concerns persist about the validity of reviews (Mudambi and Schuff 2010), and the authors confirm that in the case of sharp increases in positive affective content, the conversion rate increases are smaller than if the positive affective content increase were more moderate.
- Yet the authors fail to find a similar attenuating effect for extremely negative changes.
- This finding, though unexpected, is consistent with previous research that indicates that negative affective cues can be more powerful than positive ones for driving judgment and behavior.
- Online review settings remove the faceto-face contacts that traditionally have informed word-ofmouth recommendations, but their research reveals that the contents of reviews have significant effects when their linguistic style elicits source similarity perceptions.
Corroborating Extant Research
- Customer review phenomena have stimulated exceptional research studies aimed at uncovering relationships between the information diagnostics provided in such settings and retail performance, and yet the field still lacks a good synthesis and reconciliation of evidently divergent findings (Zhu and Zhang 2010).
- Their results highlight that price dynamics remain a decisive purchase criterion, irrespective of any other information.
- Third, in line with Mudambi and Schuff (2010), who argue that more helpful reviews increase the diagnosticity of an online retail page, the authors find that increases in helpfulness perceptions of featured customer reviews enhance subsequent conversion rates by enhancing site diagnosticity.
- By collecting longitudinal weekly data and converting their model into first differences, the authors statistically removed fixed product effects, such as inherent quality, which could account for previous mixed findings (Zhu and Zhang 2010).
- Similar to Chen, Wu, and Jungsun (2004), the authors find no significant impact of changes in star ratings on the retail site’s conversion rate for that product.
Limitations and Further Research
- The authors results are consistent with the proposition that customers read and rely on review text information in their purchase decision making (Chevalier and Mayzlin 2006).
- Online retailers, such as Amazon.com, increasingly prompt customers with complementary or alternative product choices to cross- and up-sell, so ongoing research should investigate how these products might affect the conversion success of a focal product.
- Fifth, following CAT (Giles 2009), greater LSMs reflect greater identification with a conversant (Ireland and Pennebaker 2010).
- This study’s insights into people’s emotional states and their interaction with the development of a shared language, retrievable through function word usage, suggest that text-based sentiment analysis of customer reviews should incorporate these words.
- These purchasers could improve their return on the significant investments in customer reviews by selecting those reviews that are most impactful in terms of conversion rates, as determined by the elicited affective content and linguistic style convergence.
Method
- For the pilot study, the authors solicited the cooperation of 230 business students who were in at least their third year of study (119 women [51.7%], mean age = 21 years [SD = 2.32]), to ensure that they had substantial exposure to business jargon over the course of their studies.
- The self-reported mean number of business books read was 15 (SD = 5).
Procedure
- Participants were invited by e-mail to an online experiment.
- After instructing them to picture themselves in a situation in which they needed to buy a new business book, they were randomly assigned to one of four experimental groups and sequentially shown two typical Amazon.com pages featuring a business book with two customer reviews respectively, taken from their review sample.
- The dependent variables in their pilot study consisted of purchase intention and identification with the reviewer.
- The authors used four multiple-item, seven-point Likert-type scales (coefficient = .952) to measure participants’ perceived identification with the reviewers (Prendergast, Ko, and Yuen 2010).
Results and Discussion
- The goal of the study was to determine the degree to which LSM (1) influenced participants’ identification with the reviewers and (2) interacts with affective content to increase (or attenuate in the case of low LSM) the impact of the reviews on purchase intention.
- The main effect of affective content and the interaction effect were not significant.
- Chen, Pei-Yu, Shin-Yi Wu, and Yoon Jungsun (2004), “The Impact of Online Recommendations and Consumer Feedback on Sales,” in Proceedings of the International Conference on Information Systems, ICIS 2004.
- Clark, Herbert H. and Susan E. Brennan (1991), “Grounding in Communication,” in Perspectives on Socially Shared Cognition, L.B. Resnick, J.M. Levine, and S.D. Teasley, eds.
Did you find this useful? Give us your feedback
Citations
800 citations
Cites background from "More Than Words: The Influence of A..."
...For instance, Ludwig et al. (2013) find that linguistic styles can positively affect source perceptions and lead to higher conversion rates....
[...]
359 citations
329 citations
Cites background from "More Than Words: The Influence of A..."
...Online rating also facilitates the purchasing experience of other customers who wish to compare alternatives before ordering, since the ratings visually present the feedback of previous customers (King et al., 2014; Ludwig et al., 2013; Qiu et al., 2012)....
[...]
...com demonstrated a direct relationship between online features, such as linguistic style and star rating, and conversion rates (Ludwig et al., 2013)....
[...]
321 citations
Cites background from "More Than Words: The Influence of A..."
...For example, Ludwig et al. (2013) include price discount in the model when studying the relationship between product reviews and conversion rate to control for this factor....
[...]
...Berger and Packard (2018), for example, compare lyrics from different genres, and Ludwig et al. (2013) include reviews of both fiction and nonfiction books....
[...]
...For example, a temporal sequence of documents or a portfolio of documents across a group or community of communicators can be examined for interdependencies (Ludwig et al. 2013, 2014)....
[...]
...In some cases, the researcher is interested in measuring the similarity between documents (e.g., Ludwig et al. 2013)....
[...]
300 citations
Cites background or methods from "More Than Words: The Influence of A..."
...Customer ratings supported by lengthy textual reviews containing rich information are favored by customers, and thus hotels should identify and promote the most influential reviews and provide instructions to motivate customers to write powerful reviews (Ludwig et al., 2013)....
[...]
...The relationship between the technical attributes of online textual reviews and customers’ ratings also influences future customers’ demands because customers tend to read both textual reviews and ratings to justify their consistency (Chevalier and Mayzlin, 2006; Ludwig et al., 2013)....
[...]
...The authors employed a text mining approach to find the influence of the linguistic style of online reviews on customers’ conversion rates among product websites (Ludwig et al., 2013)....
[...]
...Customers’ linguistic style is influenced by their motivation for writing online reviews (Ludwig et al., 2013)....
[...]
..., 2018), customer conversion rates (Ludwig et al., 2013), and hotel performance (Blal and Sturman, 2014)....
[...]
References
26,580 citations
"More Than Words: The Influence of A..." refers methods in this paper
...Therefore, we used a general method of moments estimator (Arellano and Bond (1991)....
[...]
...There was insufficient evidence to reject the assumption of no autocorrelation in the differences, so our generalized method estimators likely yielded unbiased and constituent estimates (Arellano and Bond 1991)....
[...]
10,792 citations
"More Than Words: The Influence of A..." refers background in this paper
...The pooled augmented Dickey-Fuller test verified that our series in conversion rates was stationary (p < .01); the conversion rate observations were independent of time (Levin, Lin, and Chu 2002)....
[...]
7,489 citations
4,585 citations
"More Than Words: The Influence of A..." refers background in this paper
...…diagnostic cues and the corresponding automatic responses may be best understood according to a positive–negative continuum (Baumeister et al. 2007; Russell 2003); thus, research examining the content of online text has applied this approach to understand links between the mood of online messages…...
[...]
...…thus recognize the effectiveness of customer reviews for converting customer visits into sales; Roku, the market leader in innovative applications for digital media, attributes a 20% lift in its online conversion rates to the appearance of approximately 17,000 (both positive and negative)…...
[...]
...Affect in and of itself is not a linguistic property but refers to an “internal feeling state” (Cohen et al. 2008, p. 297) that is “consciously accessible as the sim- More Than Words / 89 plest raw (nonreflective) feelings evident in moods and emotions” (Russell 2003, p. 148)....
[...]
Related Papers (5)
Frequently Asked Questions (15)
Q2. How did the authors calculate the interaction term?
The exogenous variables indicated frequency (percentages of affect-laden positive and negative content words in the review text; match percentage in function words), so the authors mean-centered the variables and calculated the interaction term by multiplying mean-centered variable scores.
Q3. Why did the sample size for the models fall to 4763 observations?
Because the authors used first differencing and the lagged values for conversion rate, the sample size for the models fell to 4763 observations (591 books).
Q4. What is the effect of changes in helpfulness on conversion rates?
Changes in helpfulness were weakly significant, positive predictors of subsequent conversion rate changes (helpfulness = .006, p < .10).
Q5. What is the role of linguistic styles in the evaluation of a customer review?
In an online customer review context, readers often have little but the review text to use to form their perceptions of the review’s diagnosticity, so linguistic styles may serve as identity-descriptive information that, as a heuristic cue, shapes consumers’ evaluations of the review and thus of the product.
Q6. What is the effect of affective content on the evaluation of a product?
Evidence from previous studies that use experimental manipulations to prime affective states suggests that exposure to affective cues influences evaluations and/or judgments of attitude objects, such as brands and products: Positive (negative) affective cues lead to more positive (negative) evaluations and judgments (e.g., Lau-Gesk and Meyers-Levy 2009).
Q7. What are the main aspects of the diagnosticity of affective content words?
Two facets of diagnosticity documented in prior literature seem relevant to (conversion) behavior: (1) perceived representativeness, which is related to the extent to which consumers believe that affective content reflects the target and whether the representation of the sender indicates qualifications to express his or her opinions, and (2) perceived validity, or whether affective cues appear consistent with other cues and across multiple sources (Gasper and Clore 1998).
Q8. Why did the authors conduct separate calculations for title and text intensities?
Because review titles are particularly prominent, the authors mined and conducted separate calculations for title and text intensities similar to Cao, Duan, and Gan (2011).
Q9. What is the effect of the anonymous nature of online review settings?
The anonymous nature of online review settings makes it difficult to establish sender qualifications, but extreme deviations in affective cues lower the value of feelings as information and elicit counterproductive effects by reducing diagnosticity (Andrade 2005).
Q10. What are the main concerns about the validity of reviews?
Various concerns persist about the validity of reviews (Mudambi and Schuff 2010), and the authors confirm that in the case of sharp increases in positive affective content, the conversion rate increases are smaller than if the positive affective content increase were more moderate.
Q11. What does the research show about the effect of online review settings?
Online review settings remove the faceto-face contacts that traditionally have informed word-ofmouth recommendations, but their research reveals that the contents of reviews have significant effects when their linguistic style elicits source similarity perceptions.
Q12. What is the linear relationship between affect and the linguistic style of the review?
The linear relationship tests the notion from the affect transfer and priming literature that predominantly negative (positive) reviews over time increase the negative (positive) affect conveyed through reviews, leading to reduced (increased) product conversion rates.
Q13. What is the effect of change in affective content on conversion rate?
The figure illustrates how a change in the reviews’ content toward more positive affect leads to higher predicted changes in conversion rate and yet tapers off at extreme degrees of change.
Q14. What is the effect of perceived rapport on consumer judgments and behaviors?
Such perceived rapport provides readily accessible diagnostic information, which directly influences consumer judgments and behaviors if they process information heuristically, as is the case for online information searches (Chaiken and Maheswaran 1994; Jones, Ravid, and Rafaeli 2004).
Q15. What is the importance of function words in determining conversational outcomes?
Humans are highly attentive to the conveyance of messages (Giles and Smith 1979), and prior work in several scientific disciplines has demonstrated the importance of function words for determining conversational outcomes (Huffaker, Swaab, and Diermeier 2011).