Rethinking online product review usefulness

Размещено на http://www.allbest.ru/

2

FEDERAL STATE EDUCATIONAL INSTITUTION

OF HIGHER EDUCATION

NATIONAL RESEARCH UNIVERSITY

HIGHER SCHOOL OF ECONOMICS

Saint Petersburg School of Economics and Management

Department of Management

Bachelor's thesis

Rethinking online product review usefulness

Panov Marsel Gennadievich, Yulia Fedorenko Andreevna

Saint Petersburg 2020

Abstract

This thesis deals with the issue of obscuration of many online product reviews by the more attractive “top” section and proposes a solution in form of creating a better sorting algorithm that will base its rankings on qualities of each individual review, instead of vote aggregates that are generated by the customers, who have to manually view and rate every review. Our model performed well in an experiment and successfully reranked 68% of the reviews, proving that an accurate prediction of review helpfulness based solely on its qualities is possible.

We focus on Russian online market community and analyze more than 100,000 reviews to understand how quickly people get enough information from reviews to stop reading further, reviews of what product types don't benefit or are even harmed by the current sorting algorithm and what review features can be used to successfully predict its actual helpfulness.

Our research contributes to improving customer's purchase decisions, by rearranging faulty rankings of the current sorting algorithm and presenting customers with a better set of reviews that are able to give more insights into the actual product performance.

Our academic contribution consists of first-time testing of how many previously successful indicators of review visibility perform for a related concept of helpfulness and shedding light on the effects of some new indicators of credibility, depth, informativeness and the underlying product's technical complexity. We categorize several specific content constructs that can assess what kind of claims a review includes.

Keywords: Reviews, Review usefulness, Consumer sentiment, Predictive modeling, Linear regression, Sorting algorithms

Contents

Introduction

1. Other keywords will be defined later in the text

1.1 Literature Review and theoretic background

1.2 Source credibility and authority influence

1.3 Level of review oversaturation

2. Data Collection

2.1 Sources

Conclusion

Appendix

Introduction

Background

Online retail industry grows by the day and navigating through the abundance of offerings from different providers is no easy task for the potential consumer. Obviously, each product and service has some sort of description about its features and a price tag, along with terms of service and channels of distribution. However, this is often not enough, as even after filtering through all too expensive, lacking or inconvenient offerings the customer is still left with a plethora of options with seemingly no way of knowing how to make a rational choice between them. Customers solve this issue by analyzing product reviews - electronic network of word-of-mouth experiences contributed by peers, who had decided to purchase the product in the past and then described the level of their satisfaction, how it works in real life and what issues one might expect to arise with its use.

Reviews are undoubtedly beneficial for the potential customers lurking the online markets: they narrow down their choice options, decrease the time needed to decide on the purchase of each item, mitigate the stress of deciding between the two or more similar items and decreasing the probability of future buyer's remorse (the customers familiar with what to expect are more likely to forgo bad options) - an all-round user experience improvement. If a viewer purchases a product after all, he might leave a review himself and contribute to the experience pool. This can be due to customer noticing an important aspect of the product that no one has yet mentioned and thinking it would provide useful insights for future customers, his desire to warn others about the deceitful nature of majority of the opinions on the product or simple gratefulness and satisfaction.

Companies benefit from these reviews by collecting valuable feedback on the product and, more importantly, basically having another channel of marketing, which is free, self-sustaining and is in higher regard by the customer community than many company promotional materials (Sbaffi & Rowley, 2017). The reviews are often used by the companies in their marketing campaigns. Especially by those, where the experience is the key element: namely entertainment and recreational services (e.g. nightclubs and videogame developers). Helpful reviews also increase sales of referenced products (Chevalier & Mayzlin, 2006; Clemons et al., 2006; Chen et al., 2008).

So, now we can see that reviews are a very important part of marketing and obtaining customer feedback for the companies, communication of online market communities and customers' decision-making process. Reviews themselves usually consist of 2 components: a quantitative scale (commonly a 1 to 5 stars rating) and a qualitative description in form of a comment. They also have user-submitted ratings on the review itself that are aimed to determine the usefulness of the review. If the user finds that the review positively added to his evaluation and comparison processes then he is expected to upvote a review and, if it didn't, he would either ignore or downvote it. The latter usually occurs when the customer knows in advance that the information in the review is misleading, incompetent or outright false. Although, a more common way of rating formation is agreement/disagreement expression by a fellow product user, who may vote on other reviews along with or instead of writing a review himself (people are often lazy, especially if they are satisfied with the product and/or have nothing new to add). Void of substance reviews tend to receive no helpfulness rating, because people simply don't see them as deserving attention altogether (Cao, Duan & Gan, 2011).

This review user rating system is a matter of necessity, as the other side of the coin is getting lost in the sea of controversy that often engulfs the user in review section of many offerings, especially most popular ones where the amount of reviews can approach thousands. This creates a challenge of both determining the sincerity, trustworthiness and informative value of each review and creating an aggregate view of what is reasonable to expect from the product. The trustworthiness issue arises from the company third-party and/or freelancer affiliates that create fraudulent reviews, not backed up by the actual product experience, but initiated with a monetary compensation and written in accordance with the client's (company) desires (Mayzlin, 2006). Because fraudulent reviews don't represent real experience, they misguide the customer and decrease the probability of product satisfaction. After a while even gullible customers, who had believed fake/bad reviews and bought a product that doesn't match their expectations, become aware of the credibility issues and start being suspicious of each reviewer's intent (Park & Lee, 2009), now focusing on review ratings. The existence and even abundance of “junk” reviews You can see an example in the “Samples” section below that offer very little substance (informative value) for the customer exacerbates the issue of even determining what reviews he should spend time assessing. Overly praiseful reviews suffer from the same issue, as they often lack critical analysis, focusing solely on positive aspects of the experience. When people assess aforementioned low-quality reviews, they, in theory, downvote them.

Problem Statement

Customers often rely on their peers' collective sentiment (Sbaffi & Rowley, 2017) and so if reviews are abundant, customers sort them by their helpfulness and look only at the most helpful (those with the highest number of upvotes). Doing so, they expect to see only the best, most informative, least time-consuming to read and, desirably, most unbiased and credible reviews. Sadly, fair voting on every review happens rarely, if ever: sorting by helpfulness is a prevalent strategy and so top reviews rise higher, gaining more and more votes, while unpopular ones fade to obscurity and are very rarely noticed during customer review analysis. This leads to a section of reviews being completely disregarded. But those reviews might contain unique insights about the product exploitation or even attempt to reverse the shaped public sentiment about the product. They are basically treated the same say as heavily downvoted reviews.

To combat the neglection of obscured reviews, we will take a shot at identifying beneficial and harmful qualities of reviews, based on helpfulness scores of the top section. We will build a predictive model and then test their usefulness in a real-life controlled experiment. Besides hopefully providing a way to sort reviews more fairly and efficiently, we will conclude by suggesting changes to the structure and contents of recommendations, that show up to a user when he writes a review, to serve as a guideline.

Scope of the thesis

In this research we want to contribute to the understanding of what constitutes a useful review. We will examine Russian online retailers and scrape the reviews on samples of different products to identify the qualities of an ideal review. We expect general customer sentiment to have already been defined and not to be a subject to significant change, so our findings will hopefully point out beneficial features of a review that could be used to develop a more competent sorting algorithm that doesn't rely on ratings alone and wouldn't require human evaluation for each individual review. In short, we hope to deal with review neglection and create an automatic system for their fair ranking, based on some beneficial qualities of the reviews.

Note, that we will not focus on the effects, reviews have on the product they are written for (e.g., its sales or popularity), but on the value that customers, who look at this review receive and factors that contribute to said value.

Professional significance

Current functionality of many online marketplaces (like Amazon.com, Steampowered.com or Ozon.ru) relies on users to submit their ratings for the reviews and then uses the aggregate scores of upvotes as one of the sorting mechanisms. This approach is very cost-effective as we authorize the users to be their own judges, however it is far from perfect: if the reviews are plentiful, customers will not spend time evaluating all of them focusing only on those several, that are already at the top of the sorted lists and may even skip those of lesser volume as they go. This is undesirable, because we want a fair assessment on all reviews - not just the best of oldest; and the neglection of smaller reviews might lead to overlooking endemic valuable insights. For example, products that had suffered from manufacturing or design flaws in the past and gained a generally negative scope of reviews because of it, can't benefit from the upsurge of positive reviews on fixed/upgraded products - people will continue to only look at the old, very negative reviews and thus won't evaluate an improved product fairly. This is a consequence of the modern ranking system's design. If we won't change the ranking system based on conservative public sentiment, companies will have a hard time attracting customers, when their reputation has already been tarnished.

Another current problem is that good authors, when they write a review should be rewarded with recognition of their efforts in the form of a fair assessment of its value (helpfulness). However, it doesn't always happen and a lot of them decide to forgo writing reviews in the future. This decreases the average potential quality of the review base (good authors, who tend to spend more time and effort on their reviews desert, while others remain). Also, this lack of review writing incentive may be the reason for low realized number of positive and neutral reviews (compared to the number of positive and neutral experiences with the product from all users: not only those, who wrote a review), as there exist clear incentives to write a negative review: (1) sense of enforcing justice in a form of condemnation, in hope to dissuade others from purchasing a product and (2) reliving one's anger and frustration by speaking out. If online marketplaces redesign their current ranking system, it might change the distribution of review extremity for the better and improve public opinion about non-negative reviews, as well as increasing average review quality.

Below we summarize all expected implications of our findings and their theoretical

Table 1. Potential benefactors of this research.

Key Terms

“Products” refer to items available for sale that are listed on online marketplaces. The term “underlying products” is used to refer to those products in relation to their respective reviews (e.g., “The underlying product of these reviews is technically complex”). “Users” are people, who purchased and successfully utilized the product. We refer to both product's usage and the level of satisfaction its user had with it as “product experience”, “experience with the product” or “user experience”. “Reviewers” are people, who leave reviews on purchased products. In order to write a review and become a “reviewer”, one must first use the product and become its “user”, so we will often use those two terms interchangeably.

“Customers” in our study are people who use online marketplaces to screen and purchase products. When they turn to reviews for guidance as a part of their purchase decision-making process they become “viewers”. Thus, all viewers are potential customers and in situations where we talk about the reviews, we use these two terms interchangeably.

1. Other keywords will be defined later in the text

1.1 Literature Review and theoretic background

In this section we discuss previous findings in scientific literature regarding the reviewing process, factors affecting perceived review value and other related issues. We also offer some new explanations as to the causes of discussed issues and speculate about new relationships, providing both logical justification for our choices and predicting their correlation with review helpfulness.

Review helpfulness vs review visibility

Schuff (2010) defines online customer reviews as “peer-generated product evaluations posted on company or third-party web sites” and then subsets a group of helpful reviews as “a peer-generated product evaluation that facilitates the consumer 's purchase decision process”. Reviews receive votes from their peers based on the value they provide in a decision-making process, when selecting a product out of a bunch. Usually users are asked something along the lines of “Do you find this review helpful?” or “Was this review helpful for you?” and presented with the options to either up- or downvote a review. The system then aggregates scores of “likes” and “dislikes”, which can be used to sort reviews by their helpfulness and filter out those that didn't receive much attention or have too many negative-to-positive votes ratio. Voting isn't obligatory and the factors that influence people's decision to vote is the primary subject of many papers (Zhang et al., 2014; Huang, 2015; Liu & Park, 2015; Salehan & Kim; 2016; Law et al., 2016).

The papers that we will discuss in this section, up to “New explanatory factors” haven't dealt with reviews' helpfulness, but with the severity of review obscuration We use terms “review neglection” and “review obscuration” interchangeably, because they have identical outcomes (low noticeability and, therefore, scores).. Both topics use the same primary data as a basis for their analysis - the aggregate feedback scores (or ratings) made by the customer community. So, for example, if a review had 20 upvotes and 5 downvotes, its aggregate score would be 25, as feedback type doesn't matter and any vote indicates review's prominence. We, dealing with review's helpfulness, would say that it is 80% - the proportion of number of positive to total number of reviews.

While the purposes and methods of our and other authors' researches differ significantly, we claim that the factors that improve the noticeability of reviews can oftentimes be used for assessing their helpfulness as well. First, we provide background for their findings on individual factors' effect on review noticeability, which refers to the overall number of votes on a review (25 in the example above) and then support our position for why we think it might affect its helpfulness. Our choice to use factors that affect visibility to assess helpfulness is also partially motivated by the fact that our main goal is to redesign sorting algorithms to improve the visibility of more helpful More helpful, than the current score leaders, but obscured reviews specifically. So, if some factor indicates obscuration, it must be examined before all others.

We begin with examining the extent of review neglection.

Review ratings distribution

Many researches were concerned with disproportionate distribution of helpfulness ratings on the reviews: Kim et al. (2006) examined more than 20,000 Amazon reviews on MP3 players in 2006 and found that 38% of them had less than 3 helpfulness ratings; Cao et al. (2011) scraped for his research 3,460 reviews from Amozon.com on various tech products out of which 51.59% (1,785) had 0 votes, 20.23% (700) had only 1 vote and only a handful received enough votes to be properly evaluated based on the votes alone. Those right skewed distributions appear in nearly every paper we examined no matter the date, sampling methodology or underlying product, revealing a consistent pattern.

We expect this problem to be as acute in Russian online marketplace communities.

Factors of helpfulness:

Recency of the review

One possible explanation for review neglection is that people prefer to read the most helpful information, in order to conserve their time and maximize the benefits of the reviewing process. If there is no shortage of commentary (reviews are abundant) it only makes sense to focus on the best selection exclusively. Top-rated reviews gain more and more support (more votes, higher scores), while newer ones fade to obscurity (Cao et al., 2011). Top-review section becomes mostly stagnant and very old reviews get disproportionately more visible in comparison to reviews of any other period, given they had reached the top at some point.

Common sense, on the other hand, dictates that the fresher a review is - the more helpful it gets, because newer reviews can account for changes in product's design or performance and find some issues that could have only be discovered after a substantial amount of time had passed after the product had first been used. Theoretically -more helpful reviews are obscured as a result. We will factor time elapsed since posting in our analysis and try to get rid of this injustice.

Content of the review

In order to evaluate review's contents, viewers first have to notice, read and evaluate it, so it is not appropriate to speak of review contents as a review neglection factor. At least, not in a sense we have used to up until now. Remember, that voting isn't obligatory and customers might still decide not to vote after reading or even abandon the review midway. The outcome for the read, but disregarded reviews is the same as for unnoticed ones - lack of votes. Nonetheless, in order to cause less confusion and because calling these studies papers on review helpfulness would be a stretch, we will continue to treat content issues as visibility problems.

Kim et al. (2006) tested the dependency of aggregate scores of reviews on MP3 players and video cameras on structural (word length, sentences number and their average length), lexical (what words were used and how often), syntactic (proportion of nouns, adjectives, etc.), semantic (positive and negative sentiment words), depth (word count) and review extremity (star rating). Lexicon used, accentuations, depth and extremity all proved to have a significant impact. Later works confirmed the importance of word choice (Cao et al., 2011) but purposefully left out the mention of helpful and unhelpful semantic groups and abandoned their exploration entirely, so we can only conclude that the cumulative effect of word choice is positive and is greater than for any other factor in the model (with the highest B-coefficients). At the moment it is unclear what phrases and word combinations persuade viewers to vote and in what way they are swayed (to upvote or to downvote?). Since we are dealing with both positive and negative feedback, these questions can be easily redesigned to address review helpfulness. We will offer our solution for distinguishing beneficial and harmful contents of the review and their indicators in subsection “Review informativeness” below.

Cao, Duan & Gan (2011) examined the effects of written style, linguistic choices and format characteristics of a review, using ordinary logistic regressions with text mining and factor analyses. They found that the content is much more important than format (highest and lowest explanatory power for helpfulness levels respectively) for review visibility. The most explanatory power had the model that included all variable types and factored groups of linguistic content. The site they used to download reviews from, instructed their users to structure a review with both pros and cons of the product and a conclusion - all in separate segments. The presence of conclusion (1) and the number of words in the “cons” section (2) both had a positive impact on review visibility. (1) Can be explained by people's desire to skip reading an entire review and focus rather on its final verdicts, (2) Can be explained by the risk-averse nature of most customers, who are inclined to be more interested in negative aspects of the product (Vohs, 2001; Rozin 2001).

These structural qualities of a review are expected to affect its usefulness in the same direction (positively) and for the same reasons they do review visibility: due to their contribution to time conservation and risk mitigation for most customers. We will account for the effect, the relationships between different content segments have on review helpfulness. Specifically, with our data See “Data collection and sample choice” subsection of “Research Design” section below for the description of our observations' features it would be convenient to use the number of words in “comment” section to the number of words in both “pros” and “cons” sections to measure the ratio of generalizations and conclusions to separate arguments in a review.

Linguistic analysis is later run by (Krishnamoorthy, 2015) on another sample of 13,000 reviews, although following a different grouping strategy based on the linguistic category model (Semin & Fiedler, 1991) and coupled with some extra variables (including the control experience/search group classification). Krishnamoorthy's model had explained 77% of votes on the reviews, which once again proves the importance of semantics. The model itself is very useful as it classified the words based on their role in speech (part of speech) and the situations it allowed to describe (e.g. state verbs like “love”, “dislike”, “recommend” were separated from action words “bother”, “utilize”, “take”, which were further broken down). She found that the inclusion of interpretive (positive/negative connotation) and descriptive (no quality connotation) action verbs increased the helpfulness of subjectively perceived experience goods (read below). We will use her segmentation as a guidance for distinguishing between our own content segments, however we will tweak our grouping and base it on keywords necessary to convey very specific ideas Read more on that in “Review informativeness” section below.

1.2 Source credibility and authority influence

Source credibility is in part related to the contents of the review - reviewer popularity (if there exist ratings for reviewers themselves) and ratings of his comments are also dependent on (1) proper spelling, as grammatical errors greatly reduce visibility (here - viewer's desire to read after mistakes are spotted and, therefore, vote) (Forman et al., 2008) and (2) competent, professional approach, so that the reviews that sported slang vocabulary and abbreviations (like “dunno” instead of a “don't know”) had likewise much lower ratings (Kim et al., 2006).

In (Cheng Y. & Hui-Yi Ho H., 2014) the source credibility, measured by the number of reviewers' followers for the Taiwanese restaurant critics affected restaurant selection to a much greater extent than the review comprehensiveness. However, he only included the extensiveness (word count) and number of photos (extra visual support) in the reviews to account for the latter, so there might still exist undiscovered relationships and the credibility effect might not be as strong as they had discovered. Nevertheless, the findings might suggest that the audiences of some markets are more susceptible to the opinion of the critic authority, as other papers that studied either American review databases like CNET.com or giant e-commerce services like Amazon.com didn't identify this connection, instead proving that Western audiences vote on each comment based on the content of that comment alone: for example, top user overall reviewing experience - the number of reviews left by an author prior to the research and the total number of votes he had received before on all his reviews couldn't predict the visibility of his future reviews (Huang, 2015).

This will be the ground for our doubt over the sentiment uniformity across markets and one of the reasons to study how the models, that proved successful in the past perform in the Russian environment. Review's credibility has unambiguous effect on review visibility and is expected to have similar effects on its helpfulness. Therefore, we decided to heavily focus on it in our models, providing several new independent variables and interaction effects: purchase verification, ownership evidence, reviewer identity disclosure, presence of an expertise claim and its vindication, etc.

Review extensiveness

Another very straightforward, yet no less important aspect of the review is its extensiveness, that many researchers (Kim et al., 2006; Schuff, 2010; Zhang et al., 2014; Liu & Park, 2015; Law et al., 2016) used some sort of a word count to measure. Generally, the lengthier the review the more content it can theoretically hold. The findings of (Schuff, 2010) support the claim that customers consistently view more lengthy experience goods (see distinction below) reviews as less desirable than same length search goods reviews, but the standalone effect of perceived review extensiveness on review visibility is positive for both product types.

Conversely, Huang (2015) showed that Schuff's approach to metadata, equating quantity and quality, had faults and that wordcount variable could not go unconstrained. The study had questioned if there existed a certain threshold beyond which additional words in the review should diminish their value (and, therefore, theoretically, effect on helpfulness), as not only its comprehensiveness might have an objective limit, but also the content is not safe from users' rants, deliberate sabotage of producer's reputation, paragraphs of general information on the product already provided by the product description on its respective page or irrelevant complaints/exceptional cases. Quoting Huang: “A good review may be filled with details that could make it lengthy, but a lengthy review is not necessarily a good review”. For his study Huang separated reviews into 2 groups, based on their authors reputation: top-reviewers and “commoners”. The former were expected to have a higher degree of competence, which facilitated in their high ratings. The word counts only proved significant for the “commoners” group and positively affected visibility if it was below (!) the average for the group, although being 42% shorter on average than in the top-reviews.

Forth finding of (Cao et al., 2011) also supports the hypotheses of longer reviews receiving less attention, this time by showing that people might not even look at the review itself and assessing what its length should be, based on the length of the title (the site they used to download reviews from, instructed their users to structure a review with a title). The longer the title - the more it is likely to be a lengthy review, which people tend to bypass for time conservation.

Commentary isn't the only way to convey your experience with the product. When reviews are supported by photos, they become more visible and useful, as photos catch viewers' attention and provide additional information about the product (Cheng Y. & Ho H., 2014). Restaurant reviews with higher number of visual media receive more votes on average (upvotes, specifically) than those with fewer or no photos. The effect is significantly greater for experience goods (that restaurants and other services are a part of). Photos also provide additional evidence of really owning a product, increasing the trustworthiness of the review.

Concluding all previous findings, we can say with great certainty that the length of a review is among its more prominent features: it is easy for customers to decide if they want to read a certain review just by approximating how long it will take to digest and assess its contents and how it is structured. Extensiveness of a review should definitely be included in our analysis. However, we will not only regard the extensiveness of an entire review, but also its parts. Namely, it is interesting to see, how people react to simple, non-argumentative mentions of product's pros and cons in their respective sections. These segments are very frequently structured in the form of a bullet point list with some generalizing title for each point at the beginning, for example:

Product pros.

Screen. Very bright and the colors pop. Very pleasant, vivid image.

Viewing angle. You can look at the screen from a significant angle without changes in image's color, saturation or contrast.

Size

Connection”

In our example, only the first two pros are given sufficient arguments necessary to understand what exactly an author had in mind. An unfamiliar viewer can get lost with non-argumentative mentions of product's pros and cons, especially if they are ambiguous and can't be assessed with product's frontpage description (like “Connection”). On the other side, these reviews can validate the experiences, described in other reviews, as you already know what hides behind a keyword for a feature. Therefore, at the moment we can't predict how the public in general views non-argumentative bullet point mentions of product's features: with contempt or with gratitude.

Review extremity

The extremity of the review (i.e., its star-rating) is not only used to create an aggregate, comparable score for the product and facilitate a sense of ratings distribution, but also to filter the results if a customer wants to see only the reviews of a specific rating. Usually the latter preference is determined by the product type, customer's tastes and/or his initial attitude toward the product (Kim S.M. et al., 2006; Schuff, 2010). Since people are naturally predisposed to react more vividly to negative signals (Vohs, 2001), it is no surprise that many papers found a strong negative correlation between star ratings and review visibility (Kim et al., 2006; Schuff, 2010; Huang, 2015; Eisend et. al., 2015, etc.). Papers that dealt with a more structured reviews (Cao et al., 2011) found that the longer the “cons” section was, the higher were the ratings. Note, that the study measured review extremity unconventionally: as the difference between the review's rating and the aggregate average, so their findings speculate more about the deviations of ratings and not the ratings themselves.

We will factor review extremity effect in our helpfulness model. We don't include different sections' lengths, because we believe that the length of a section (pros/cons/etc.) is strongly correlated with review's star rating (that way we avoid multicollinearity).

Closely connected to the concept of extremity is the position of the reviewer, that can either be “two-sided” (when he provides both positive and negative commentary on their experience with the product) or “one-sided” (when he keeps to one side of the argument). Previous works (Schlosser, 2005; Eisend, 2006; Pavlou at al., 2006; Forman et al., 2008; Schuff, 2010) have shown that product type and position of the reviewer play an important role in determining review's benefit for the customer. Movies, for example, are hypothesized to be viewed by the public as very opinion-based, so people recognize that moderate (around 3-stars) two-sided reviews are more “rational” and level-headed, and thus they receive more credit that turns into higher visibility than extreme (very negative or very positive) two-sided reviews. For many other products (e.g. books (Forman et al., 2008; Schuff, 2010)), on the other hand, people favor neither moderate nor extreme two-sided reviews. We suggest, that customer's initial attitude towards the product is very important: if his initial sentiment is negative, he is more likely to vividly react to positive reviews and vice versa, because their comments shift his perspective and contribute something new to his understanding of the product. Since two-sided reviews include both praise and criticism, they are immune to being filtered out by the customers based solely on their initial attitude. Two-sided reviews offer both points of view: one, recognizable to the customer and two, competing claims that are yet more likely to be viewed from his perspective. As a result, simply more people wish to view and rate two-sided reviews than one-sided reviews. This more frequent rating behavior is a sign of two-sided argumentation importance for review helpfulness. Of course, we will view it in conjunction with (and controlling for) review's star-rating in our analysis.

As for now, there still doesn't exist a consensus on the effect that the star-rating has on review credibility and helpfulness. We think the inconsistency is connected to failure of group distinction. We see the solution for this and many other aforementioned problems in more control variables, namely those that are related to the type of the underlying product. We explore them in the next section.

Product type segmentation

Some voting behavior is only observed when we deal with goods of specific types, which can serve as proxy-indicators of goods' target audiences' characteristics and preferences. We surmise the divergence in findings is likely due to the types of products selected.

Search vs Experience goods

Not all products present an equal opportunity to evaluate their qualities with customer reviews. Products can be divided in search and experience goods (Nelson, 1970, 1974): search goods are those that can be evaluated before committing to the purchase (mainly, by analyzing the reviews) and experience goods require additional direct individual sampling/testing to be evaluated justly, due to their highly subjective value offerings. Technological and medical products are an example of search goods, while wine and music are experience goods (Nelson 1970, Huang et al., 2009). Note, that products exist on a spectrum between pure search and pure experience goods.

This subjectivity of experience goods creates a barrier for using peer product reviews to assess the offering, thus requiring a different sort of content and within them (Schuff, 2010).

We already mentioned one of differences between search and experience goods in that review extensiveness benefits more to search goods than to experience goods (Schuff, 2010), but the former are more valuable if they contain descriptive adjectives and emotional semantics (Krishnamoorthy, 2015). In general, experience goods were proven to require people viewing and rating their reviews more frequently (Eisend et al., 2015).

The drawback of this grouping method is the qualitative nature of the segmentation criteria, resulting in high levels of interpretability and ambiguity of the product category. How much subjective value defines a product as an experience good? How can we measure it? Couple that with the internet blurring the lines between the two groups and you get a very inconsistent classification. Academics deal with this issue by studying only those pure search and pure experience goods mentioned in the earliest works (Nelson, 1970).

At the moment we can only forecast that this typization most likely affects the relationships between review extensiveness, content and linguistic choices and its helpfulness. Other interactions remain hidden, but may pop up later in our analysis.

Price category

Experience/search goods typization is a staple of almost any modern regression on review helpfulness and the shortage of unambiguous goods leave them little wiggle room for further classification: most mentioned papers focused primarily on no more than 5-6 products, ranging from mp3-plyers to videogames and restaurants. However, it is reasonable to guess that customers with varying financial resources and spending habits will exhibit varying reviewing/rating behavior. The pricier an item gets the more risk its buyer accepts, as more resources have to be spent on acquisition. This begets in customers more meticulous attention to details, that are often found in reviews, persuading buyers to view and rank.

Cao et al. (2011) justify our guess with their study on software reviews. Free software receives less votes than paid software, because users don't have any direct financial risks, when trying out free software, so the “stakes” are lower and therefore they don't (1) see reviews as valuable when making a decision to try, (2) look them up, (3) vote on them (given, not all people abandon the reviewing stage, but enough to warrant the price distinction).

Sadly, Cao and others only distinguished between free and non-free products. It is still a question if the relationship between the financial requirements for the purchase of the product and review's visibility will persist on a scale between most cheap and most expensive products.

In the context of review helpfulness, it is reasonable to assume price category We will have 3 classes of products: “cheap”, “midrange”, “expensive” to affect more objective review qualities: credibility and informativeness. Credible reviews, in theory, should receive extra trust from customers, so it would be easier for them to be sure of making riskier (pricier) purchase decisions. More information in a review makes it just that more valuable with the rising of the stakes (a chance to select a suboptimal product).

Research gaps

Research gaps summary

As we already mentioned, scientific community has so far directed their efforts at identifying review qualities that influence reviews' visibility, whether speaking about their influence on customers' decision to vote after reading a review or even customers noticing that review in the first place. These studies and their findings are useful in understanding why the top reviews lead, but they instantly assume that reviews with lower scores are intrinsically worse than top-section leaders and deserve the votes they receive. This is simply not the case: obscured low-score reviews might be just that - obscured, by more prominent and popular top-section reviews. Intrinsic quality of a review doesn't depend on and shouldn't be evaluated with customer community participation in the voting process, but on the actual features and content of a review alone. Previous studies are more about customer voting inclinations than about benefits reviews provide.

Papers, that distinguished between aggregates of up- and downvotes face another problem: their findings are incomparable. Since aggregate scores are absolute values they are affected by the underlying product's popularity. For a very popular product, the highest-ranking review (review 1) could have, say, 100 upvotes and 25 downvotes and for a not so popular product, the highest-ranking review (review 2) could have just 20 upvotes with no downvotes. Previous studies would consider the second review less visible and therefore call its features detrimental or less effective. We, on the contrary, would generalize the scores and remove the effect of product's popularity. This way, we measure each review's helpfulness ratio - what proportion of customers found the review useful. We would conclude that the second review is more useful with 100% of customers benefiting from it, as opposed to 80% for the first review. We could then advise to make more informative reviews more visible, by adjusting the sorting algorithm. Previous researches lack such applicability and can't even compare their results in many cases, as none factored product's popularity in their models.

Despite that, we will still use some of visibility-affecting review qualities to assess helpfulness. The only problem is that all reviewed studies targeted Western and South-Asian online markets and we can't be sure that similar relationships that we use to justify our explanatory and control variables will be found in Russian online communities. We will have to answer this question not only to empirically validate our theoretical background, but also to showcase the extent of review neglection in Russia, as it has never been studied before.

Another prominent feature of major Russian online markets is that they support interface features that allow us to measure each review's helpfulness: both “upvotes” and “downvotes”. After the policy change somewhere in 2016 Amazon.com no longer offered its users to leave a “Not Helpful” feedback on a review. As Amazon.com is the leading Western online marketplace In our research we are only concerned with Amazon's core business it is a very common data source for the studies on reviewing behavior. The lack of negative feedback leaves researchers, who choose Amazon.com or any other platform, similar in functionality with no ability to generalize observations and assess helpfulness at all. With Russian online markets we will not face this problem.

For our research we will propose review qualities unique to the issue of review helpfulness. We will explore the influence of specific linguistic choices, perceived review genuineness, reviewer's authority, review informativeness and threshold levels of review oversaturation along with controlling for the underlying product's technical complexity. “Linguistic choices” will elaborate on works of Cao et al. (2011) and (Krishnamoorty, 2015), who forgone exploring specific words affecting review helpfulness. Instead of mindlessly factoring the entire dictionaries we will approach the contents of reviews from the perspective of different “sections” containing very specific types of information that can be identified by using a very limited set of keywords. New credibility indicators of review genuineness and reviewer's authority aim to explain and expand on the contradictory findings of Cheng & Ho, (2014) and Huang (2015). Review oversaturation levels aim to explain unintuitive findings about the impact of posting recency (Gao, Duan & Gan, 2011). Informativeness factor is included to compete with current extensiveness measures, which rely on mechanical word counts: we expect measures of informativeness to perform better than word counts. Product's technical complexity is an entirely new factor and is unique to this paper. All these factors are given theoretical justifications in the following subsection.

New explanatory factors

Review genuineness

Before a review is read and its contents are evaluated it has to be pre-screened by the customer. During this selection process he wants to make sure that the review is worth his time based on surface criteria of the review and that person's intuition. The ideal review has to be (1) informative, (2) concise and (3) factually correct. Customers use surface-level indicators to filter uninformative, too long and/or incorrect reviews right away, although not always appropriately.

Factual correctness, or review accuracy, is a measure of how accurate is the information that a review conveys. Oftentimes, people are not savvy or knowledgeable enough to tell a legitimate claim from a fraudulent or incorrect one, especially if the reviewed product is technically complex. Many claims about the product may be made that are supported by user's experience alone and don't rely on any previous knowledge a viewer of such a review might possess. People may understand that the current paradigm can be shifted and an unreasonable claim about the product might prove correct in the future and assume that they just don't know enough themselves.

In situations like this people might want to have some evidence for the fact that the product was even fairly tested and that the reviewer has actually had the experiences he describes (and customers question). So, customers try to distinguish between genuine and forged reviews.

“Genuine reviews” are reviews that were created by the actual product/service users - people, who have purchased or received the product without any incentives from the third-parties. The purchase of the product was made of their own volition. Their claims might be false due to human error but never fraudulent. Only these reviews are desirable for the viewers.

In contrast to genuine reviews, “forged reviews” are reviews that are not supported by the actual user experience. These reviews could be commissioned by the companies that want to sell the product from freelancers or could be left by companies themselves. Notice, however, that forged reviews don't include paid promotional material by influential media figures and big-name reviewers just because they were commissioned by the product seller to market the product, because even though they may downplay the negatives and inflate the positives, the reviews are still usually written after the reviewer had tested the product by themselves. Forgery is when the reviewer writes exactly what was ordered by the commissioner and might not even have seen the product in real life, let alone test it. These reviews tend to be overly praiseful and are always fraudulent. These reviews are undesirable by the viewers.

Fortunately, customers have options in evaluation of review genuineness. Big online market platforms, for example, monitor what product pages their customers visit. If they register a purchase (not in case of redirection to the seller's website, but when the purchase is made directly from the platform) a customer will be classified as “verified customer” and an according label will appear at the top of their review, in case they decide to write it. This label leaves no doubt of monetary transaction happening and thus makes it less likely that a review is forged, because customers primarily purchase products for their own use, thus testing the product. Another giveaway of genuineness is inclusion of amateur user photos of the product or unboxing process. Aside from providing extra informativeness they signal that the reviewer really possesses the product and thus at least has an easy opportunity to test it before writing a review. Even if the review turns up to be commissioned its genuineness will still be greater than for the commissioned review without photos, because the author is expected to be less tied down to the script provided by the company and explain some of his claims with actual product experience.

Reviewer's authority

Reviewer's expertise can play a role in providing extra weight to his arguments and it is no surprise that people tend to trust specialists more than any other regular user. If the viewer perceives review author's claims as possessing some level of authority behind them, he will find that review more credible (and therefore making it more helpful).