The main objective of this study is to identify the relevant service characteristics of online shops. However, a regression analysis including all 15 different service characteristics is not appropriate due to the high multicollinearity between the regressors. Table 2 lists the correlation coefficients for the different customer evaluations of the retailers’ service characteristics; most characteristics are correlated with a coefficient of correlation larger than 0.7.
To transform the number of highly correlated variables into fewer unobserved variables, we use factor analysis, whereby an underlying unobserved variable (called a factor) is constructed as a linear combination of the observed ones. We then exclude the customer evaluations of price because Rel. Price and the Price level capture this attribute in a more direct manner. Since we want to identify the main categories that can be developed from the fourteen different characteristics, we have applied the “Principal Factor Analysis” (see Backhaus et al. 2008 or Basilevsky 1994). As far as the number of underlying factors is concerned, the literature offers various methods, which might lead to different results.Footnote 21 According to the “Kaiser-Criterion”, the optimal number of factors should be determined by the number of eigenvalues higher than one. For our data, the “Kaiser-Criterion” would retain only one factor. In addition the “Scree Test”, searching for a sharp change in the first differences of the eigenvalues, indicates that only one factor should be used. This is obviously not viable. A third approach—preserving as much information as possible in the potential factors—is to extend the number of factors until 90 or 95 percent of the variance can be explained. This would lead to four to seven factors. Given the spread of potential factor numbers, we have opted for the interpretability criterion (see Hatcher 1994): All solutions between the extreme values of the other selection criteria should be calculated and the most interpretable solution selected. In our case, five factors emerge as the optimal ones.
Table 3 shows the rotated factor loadings of the different service characteristics, based on a principal component analysis with orthogonal varimax rotation for five distinct factors. Factor loadings below 0.4 are usually considered as low and contribute little to the principal factor (in the table, factor loadings above 0.5 are printed in bold). The last column “Uniqueness” is the proportion of variance of the variable (e.g. Navigation) that is not accounted for by all of the factors used. Note that in 9 out of 14 cases, our chosen principal factors explain a communality (calculated as one minus uniqueness) of more than 90 percent of the original variance in the service characteristics. For the remaining five variables (Navigation, Product Info, Service before Buy, Website Performance, and Order Transaction), at least 85% of the original variance is explained.
The following reduced set of factors can be calculated as linear predictions of the rotated factor loadings: The first factor Order comprises all the relevant aspects of the ordering process. Besides the choice options in form of the retailers’ assortment,Footnote 22 aspects such as the convenience of the ordering process, order confirmation and package tracking as well as the delivery time influence this unobserved factor. Examining the factor loadings, it emerges that the delivery times, in particular, as well as the individual’s information needs (accuracy of information and feedback on the order and shipment process), are the driving forces behind this factor. Legal conditions such as the terms of business and the terms of payment (e.g. payment options) load up to the second factor Terms & Conditions. The functional aspects of the retailer’s website, such as convenience of navigation, the speed and response rate of the web-server, and the provided product information are highly correlated with the third factor Web Performance. Post-delivery Service, covering all aspects of the ordering process after arrival at the delivery address is also an important area; the quality of packing and service after purchase (e.g. handling of warranty claims) contribute to this factor with factor loadings of above 0.5. Finally, the Shipping factor channels the consumers’ satisfaction with shipping cost. As shipping costs also relate to payment conditions it is not surprising that the variable payment conditions have the second highest factor loading for this factor.
These factors can be compared to the service dimensions found by Parasuraman et al. (2005), who developed a 22-item scale with four dimensions: efficiency, fulfillment, system availability, and privacy. While their scales are based on focus group discussions, our scales are pre-defined by www.geizhals.at and cannot be manipulated. Still, there is strong correlation between these dimensions and ours; only privacy issues are absent in the items www.geizhals.at offers on the customer evaluation form. Bauer et al. (2006) identify five service qualities in online shopping: functionality and design, enjoyment, process, reliability, and responsiveness. Again, there are considerable analogies with our dimensions; only enjoyment and responsiveness are not explicitly present in the www.geizhals.at catalogue.
Impact of service characteristics on online demand
Table 4 contains the results of our negative binomial regressions. We show the marginal effects of the relative price and some other control variables, and two ways to measure service quality for an online shop; whereas the first two columns include the overall measure of the retailers’ quality calculated by the average across all the different service characteristics, the remaining four columns use the more detailed principal factors from our factor analysis. In columns (1) and (3), the number of referral requests from the www.geizhals.at site to the retailer’s shop (all clicks) is used as a dependent count variable. Columns (2) and (4) show the results for the LCT count. Owing to the large number of observations, all the variables are significantly estimated and all have the expected sign, although we see some variation in the significance levels.
Impact on referral requests
In columns (1) and (3) we use clicks as dependent variable to measure the attention of customers for online-shops. Obviously, perceived service quality is an important determinant of attention in column (1). A decline in service quality by one grade in the 1-5 scale generates a log-difference in clicks by 0.100, which represents 9.5% of the average amount of referral requests. Note that firm evaluations vary only with a relatively low standard deviation of 0.36 around the mean of 1.79. Nevertheless, we measure a substantial and significant effect of change in consumers’ firm valuations; Thompson and Haynes (2017) show that the possession of one additional star in reputation reduces the discount in price relative to the manufacturer’s recommended selling price by only 1%. It seems that quantity reacts stronger than prices on changes in service quality—this may be due to the strong visibility of prices in the explicit price ranking on www.geizhals.at.
Instead of the overall measure for service quality, Firm Evaluation, column (3) uses the more disaggregated principal factors. All factors influence the customers’ attention in a highly significant and expected manner. Better service quality increases demand. However, the factors can be classified in two groups of very important and less important service characteristics.Footnote 23 It turns out that the factor Order (summarizing the customers’ valuation of the convenience with the ordering process within the website) together with the factor Website Performance (showing the customers satisfaction with the navigation process, the quality of product infos, as well as the reaction rate of the shop software) are decisive features influencing the customers’ attention. The other factors Terms & Conditions, Post-Delivery Service, as well as Shipping play only a minor role.
Inspecting our other control variables, it is not surprising that the relative price of an offer is a very important variableFootnote 24; an increase in the relative price by 10% would decrease the log difference of clicks by − 0.0598 clicks. Given the mean of 0.44 clicks per firms’ product offers, this results in a 5.8% reduction in demand. Other firm-specific characteristics corroborate theoretical predictions and also have a substantial impact on demand; shops located in Germany attract 10.8% less demand from Austrian customers. Presumably, customers fear warranty or delivery problems across borders. Immediately available product offers have 4.8% higher referral requests. Offers with an additional pick-up possibility—the online shop is part of a brick-and-mortar store—receive 5.8% more clicks.
A comparison with older Austrian data in Dulleck et al. (2011) shows decreasing importance of customer valuations and increasing impact of relative price over time. It seems that customers’ confidence in e-commerce transactions increases over time, which makes price differences between shops relatively more important. With a similar argument, we can explain the decline of discrimination toward German shops. However, the service feature “pick up possibility” becomes more important over the years.
At first glance, the fact that relative shipping cost has a slight positive effect on demand is surprising; doubling the shipping cost increases referral requests by 1.2 percent. This result may be explained with the successful working of obfuscation strategies in online markets; attract customers with low product prices and generate profits with high shipping costs or more expensive complements.Footnote 25 Hamilton and Srivastava (2008) analyze price partitioning on the internet with examples of price and shipping cost. It seems that customers only start looking at shipping cost once they become interested in buying from the shop. The shipping cost amount is the only variable that we have to parse from a text field. Therefore, some 10% of the cases cannot be properly coded. These missing cases are included with the mean shipping cost and are accounted for with a missing flag variable. As expected, the firm-specific general price level, representing the average relative price of all other goods offered by the respective retailer, has a negative effect on attention; an increase in the price level of the other offered goods by 10% reduces attention by 2%. Hence, retailers with a reputation for relatively low prices attract more consumers, a result that is quite remarkable in highly transparent online markets.
The higher the number of inspections of detailed customer evaluations, the higher is the number of clicks. Apparently, #Inspections acts as an indicator of the attractiveness of an online shop. However, as a high number of quality evaluations is a prerequisite for inspection by customers,Footnote 26 the coefficients of #Inspections and #Evaluations have to be interpreted jointly. If the #Evaluations enters our estimation without the #Inspections, we observe a robust positive and highly significant coefficients for the number of evaluations, as expected (results not shown in the tables). Obviously, an increasing number of firm evaluations has a positive effect on demand, because customers might trust the reliability of the shop and the evaluations further. Although the inclusion of #Inspections changes the sign of #Evaluations, we should not oversee the positive aggregate effect—as the negative coefficient of #Evaluations is relatively small, the total effect of a higher number of evaluations on the number of clicks is unambiguously positive considering the relatively strong and positive correlation between #Inspections and #Evaluations. For all our explanatory variables, the results concerning our control variables are robust across all estimations with different dependent variables.
Impact on Last-Click-Through clicks: a proxy for actual purchases
Thus far we have discussed the influence of service characteristics on referral requests. Unfortunately, the actual act of purchasing a product is unknown, because actual purchases occur at the e-tailer’s own website, which cannot be observed by www.geizhals.at, and therefore, by us. We use the Last-click-Through concept as a proxy for actual purchases. Columns (2) and (4) show the results for our fixed-effects negative-binomial regressions.
The results from LCT regressions behave as expected. Theoretically, all the reliability and quality aspects of the e-tailer should become more important in an actual purchasing decision. This is exactly what we observe—with a falling mean of the dependent variable, practically all coefficients increase.Footnote 27 To mention one example, for our LCT measure, a deterioration of the customer valuation by one grade would reduce the amount of LCTs by − 11.9% (compared with -9.5% for referral requests).
The use of LCTs as dependent variables also increases the importance of our service characteristics in column (4). The service characteristics Order and Web Presence remain the decisive factors. Although we see that Shipping, and to a lesser degree, the Post-Delivery Service gain relatively in importance compared to the main indicator Order, the remaining factors Terms & Conditions, and especially, Website Performance, lose in relative importance. These results are consistent with the expectation that factors are more relevant if it comes to actual purchases.
In addition, our other controls behave as expected, the relative price and the availability of the product offer become more important if we reduce our analysis to referral clicks with a higher purchase probability. The homeward bias of Austrian consumers remains relatively constant. As expected, shops with a reputation for low prices lose some of their advantage if it comes to actual purchases.
Interested and informed customers
Estimates in columns (1) and (3) of Table 4 use the clicks of all consumers as dependent variable, irrespective of whether the customers have informed themselves with detailed customer evaluations.Footnote 28 Hence, the last two columns of Table 4 use an alternative approach; Column (5) only uses clicks from customers who have at least once inspected the detailed firm valuation of the different service characteristics for any firm in our sample (call them Interested Customers). Column (6) goes one step further and only counts clicks from customers who inspected the detailed customer valuations from the offering firm at least once (we call them Informed Customers).Footnote 29 The negative binomial regressions for Interested Customers and Informed Customers show statistically robust and expected results. Although the last two columns show smaller coefficients, they refer to substantially lower means of the dependent variables.Footnote 30 Interestingly, for those subgroups of consumers, the quantitative impact of service characteristics lessen somewhat. Whereas a deterioration of the factor Order by 1 would reduce the amount of all clicks by − 4.4%, the values for columns (5) and (6) are − 2.8% and − 0.4%, respectively. The dramatic reduction in sample size and the resulting sample selection to certain offers as well as the imbalance of information could be an explanation for this result. As these customers have only inspected certain shops, it does not necessarily mean that they are better informed. As we do not observe clicks from informed and interested customers for all of our products and firms, we lose different numbers of observations in the last two columns, whereby a direct comparison of the effects for informed and interested customers does not make sense.Footnote 31
Importance of service in different markets
Online shops specialize in a certain branch of products (e.g. web shops specializing in the video and photo business) as well as general stores with a very broad assortment of products. Table 5 repeats our analysis separately for specific product categories: “Audio/Hifi”, “Games”, “Hardware”, “Household articles”, “Software”, and “Video/Photo/TV.” Table 5 and the following uses the clicks of all consumers as the dependent variable. The results for the aggregated firm evaluations are shown in Panel A in each table, and the marginal effects of disaggregated principal factors are included in Panel B.
Product Categories: Examining the results in Table 5, we see that the broad lines of our argument have been confirmed. On an aggregate level, firm evaluation is important for all markets, with the highest impact recorded for Audio and Hifi products.Footnote 32 If we consider the different service characteristics in Panel B in detail, by and large, Order and Web Performance are again the most important service characteristics for all product categories. Noticeable is the relatively strong effect of Order in the Audio/Hifi category. Shipping conditions are more important than Web Performance for only the “Software” category. The specificity of information products is also demonstrated by the relatively small economic effects for our factors in the category “Games”. Our results also show that the category of household appliances are specific, as both Terms & Conditions and Shipping are statistically not significant. Apparently, other variables are more important in the decision for a specific shop offer than these two factors. The fact that household products only entered the price search engine in the recent past might explain this result.
In terms of the other firm-specific variables, relative price has the highest impact on the demand for household and Audio/Hifi products. For the IT-related product categories of Hardware and Software, the country of origin plays a very subordinate role. Almost all variables come up with expected signs and comparable coefficients across the different product groups. We only obtain contradicting coefficients for the general price level of a firm, which may be easily possible, because being a cheap shop may have quite different demand effects for high- or low-quality goods. In our classification, a more expensive shop increases demand for Audio/Hifi and Games, but reduces it for Hardware, Household appliances, and Software.
Price of products: The first two columns in Table 6 stress another classification—the price level of products. While column (1) shows the regression results for low-price products, column (2) presents the results for products with price above the mean. Note that, on average, high-price products are clicked nearly three times as often—obviously, consumers inform themselves better on expensive products. We expect that a purchasing decision in the case of more expensive products is more thoughtful; hence, we would expect higher coefficients, especially for the relative price. Panel A confirms these expectations. Both the relative price and customer valuations become more important for expensive goods. The comparison of service factors in Panel B indicate an even greater shift toward the factors Web Performance and Order for high-price products.
Information versus non-information goods: columns (3) and (4) compare information goods (e.g. games and software products) and non-information goods (e.g. Audio/Hifi, Hardware, Video/Photo, TV, appliances). Overall these product groups are similar, although there exist some differences; consumers of information goods react stronger to Web Performance, while consumers of non-information goods give more relative importance to Shipping, a fact that might be related to the transport of bulkier goods.
Main products versus accessories: Table 7 juxtaposes the main products against accessories, that is products that are not used alone, but only in combination with a main product. We refer to the classification system of www.geizhals.at, which assigns all products into a hierarchical system of categories, subcategories, and subsubcategories. Products within the subsubcategories are typically substitutes. We have classified each subsubcategory into main products in column (1) (e.g. single-lens reflex cameras SLR), expensive accessories in column (2) (e.g. lenses for an SLR), and cheap accessories in column (3) (e.g. SD-cards for the SLR). The remaining subsubcategories, for which no classification was possible, are subsumed as “Others” in column (4). Most of the products can be found in main products and cheap accessories. Again, our expectation would be that consumers make well-considered decisions for main products and expensive accessories. For cheap accessories, the purchase decision is often made within the shop’s website and not at the price search engine. Saving shipping cost is one motivation for this behavior. A glance at Panel A confirms our expectations. For main products and expensive accessories, we see higher coefficients for the general quality evaluation of firms than for other categories. As expected, the relative price of cheap accessories has the lowest effect on customer clicks. Unsurprisingly, we again observe the shift toward Order and Web Performance for main products and expensive accessories. The same is true for the category “Others”.
Firm types: Table 8 compares the impact of service characteristics for different firm types. In columns (1) and (2) the median of the shops’ total number of clicks on all offered products is used to distinguish large and small firms. In columns (3) and (4), the median of the total number of shops’ evaluations separates firms into samples of few and many evaluations. By separating these groups, we can analyze the impact of different service characteristics for firms that are “widely known”, compared to other small stores. We would expect stronger impacts of service characteristics for larger firms, which typically have more customer evaluations. The systematically higher coefficients of service characteristics for larger shops with more customer evaluations demonstrate that customer reviews may help overcome information asymmetries.
Sum up: We observe a recurring pattern that customers react positively to Order and Web Performance, especially for high-priced or main products. Online shops typically want to sell these products because they bring higher profits and sales (e.g. main products also sell complementary products). To attract customers for these products, our results demonstrate the importance of investments into the convenience of the ordering process, order confirmation, package tracking, and delivery time. Moreover, firms should invest in functional aspects of the retailer’s website, such as convenience of navigation, the speed and response rate of the web-server, and the provided product information. We measure a systematically higher impact of service characteristics, especially in large firms with a high number of customer evaluations.