Memoire Online - The effect of raising searching obstacles on online purchasing behavior: proof from field experiment

THE EFFECT OF RAISING SEARCHING OBSTACLES ON ONLINE PURCHASING BEHAVIOR: PROOF FROM FIELD EXPERIMENT

The effect of raising searching obstacles on online purchasing behavior: Proof from field

Abstracts

While online retail allows consumers to obtain goods or services directly from a seller via an additional channel, operating margins are often lower in online stores than in physical stores. There are well-known reasons for this disparity: price comparisons are easier to do online, coupons and codes are more widely adopted, and marketers often bear the cost of shipping products to buyers. Most online stores are designed for frictionless shopping, with few barriers to finding and buying discounted products. We propose that the intentional addition of search frictions - barriers to locating discounted items - may improve online retailers' margins by allowing shoppers to choose between «paying with money» (low discount) or «pay with effort» (high discount). In a series of field experiments carried out with an electronic commerce platform specializing in diasporas connecting buyers and sellers, we show that getting customers more difficulties in finding products at a reduced, reduced price the average discount on items purchased without reducing the impact of purchases or the average selling price. By using transaction information from existing customers, we show that price-sensitive buyers are more likely to make efforts to locate heavily discounted items. Our results posted that adding research frictions can be used as a self-selected price discrimination tool to offer high discounts to price-sensitive consumers and reduce the number of price-insensitive consumer subsidies.

Keywords: e-commerce, online purchasing, obstacles, search costs, price discrimination

1. Introduction

Online retailing expands business access to consumers through an addition channel, but operating margins are often lower in online stores than in physical stores. Afrimarket, the largest online reseller in Benin, achieved average operating margins of 2.8% between 2013 and 2018, while its traditional counterparts earn between 4% and 8% (Insae, 2018). The reasons for this difference are well known: price comparisons are easier online, coupons and codes have higher adoption, and sellers often bear the cost of shipping products to buyers.

Consumers shop online for products that are also available in brick-and-mortar stores because it is generally easier to browse a large selection of goods and fulfill transactions online (Teixeira and Gupta 2015). Online retailers like Afrimarket, Odjala, and mymotherlandstuffs are continually striving to lower search, transaction, and delivery costs for consumers. Afrimarket is the best example of an electronic commerce platform that systematically reduces the obstacles of online search.

This trend contrasts sharply with the practice of physical stores that have long accepted the deliberate use of research frictions to improve store revenues. By making it more difficult to locate discounted or otherwise less expensive items, by placing the sales section at the back of the store or in a separate store, physical stores can induce self-selection among the consumers who stand out. by their sensitivity to price and their willingness to search.

In this paper, we seek to challenge the prevailing assumption that minimizing search frictions, i.e., facilitating consumer search across a retailer's entire assortment, is the optimal strategy for online retailers selling searchable branded goods (Bakos 1997, Brynjolfsson and Smith 2000). We argue that, just as in physical sales contexts, careful integration of research frictions can facilitate price discrimination in online retailing.

The existing literature has typically conceived of search costs as the time, effort, and money required to physically identify and consider additional options before making a purchase decision (Bell, Ho and Tang 1998). Given the ease and immediacy of online shopping, it is not surprising that equivalent search costs have not been studied as tools that a company would use to implement discrimination based on price. We identify and explore the power of search costs in online settings: the effort of clicking an additional link, displaying an additional page, scrolling through a catalog of articles, or mentally calculating the discount percentage on

We assume that, under certain conditions, an online retailer can improve its gross margins by increasing the search costs associated with the search and purchase of discounted items on its website. The first condition is that there is a negative correlation between price sensitivity and sensitivity to research costs: consumers who are more concerned about getting good deals are less concerned about making efforts to locate them. The second condition is that by encountering these additional research frictions, price-insensitive consumers would replace very small items with smaller items. The third condition is that price-sensitive consumers would make the extra effort required

To test our hypothesis, we conducted a series of field experiments with online kitchen items and tableware retailer. This category is particularly attractive for our purpose as it has a moderate frequency and purchase value. Consumers are broadly aware of price points for kitchen items and tableware but not completely certain of item prices at any given purchase occasion. And, luxury brands notwithstanding, item prices are material but not exorbitant to most shoppers. Lastly, it is common practice for kitchen items and tableware retailers to frequently offer sizeable discounts to acquire and retain customers.

As part of the first trial, we randomly distributed new visitors on an online platform for a week in one of the reference groups or one of the three treatment groups. The supply and price of the products remain unchanged under all conditions. Each process means increasing search friction in some way: (1) deleting the direct contact with the point of sale because of the large reduction of the point of sale; (2) deleting the order by delivery option; (3) deleting the delivery mark unique to this article. We note that in each of these cases, the average discount rate for purchases is much lower than the control state without reducing the conversion rate. These results demonstrate the power of the treatments we have chosen and support our hypothesis.

In a follow-up analysis, we aim to establish the mechanism underlying the results of our first experience. We use the historical transaction data of existing customers to pre-classify them according to their price sensitivity. To do this, we downgrade the last basket update to demographic and past purchase variables and then use the expected values as an approximation of price sensitivity. We endorse this classification by showing that buyers we identify as price-sensitive are much more likely to click on random price electronic newsletters (rather than on

In our second experiment, we randomly assigned all visitors, new and existing, to the online store for two weeks, to a control group or one of four treatment groups. Once again, product availability and prices are kept constant under all conditions. We are resuming the treatments of our first experience and include the replacement of discount banners with no discount banners as an additional condition. The goal is to identify the presence of self-selection among all of the company's customers, as price-sensitive consumers are relatively immune to additional search costs, but price-insensitive consumers are not.

We find that, as in the first experiment, the average discount on purchased items is lower in the treatment groups than in the control group, while the conversion rates are not negatively affected by the addition of research. Also, these gains are attributed to the fact that price-insensitive consumers purchase a disproportionately larger number of full-price items in the treatment groups. These results imply that price-insensitive consumers turn to cheaper items when severely reduced items are harder to find. They also show that the key effects we capture are stable under varying demand conditions, as our second experience, which included new and existing customers, was conducted more than a year after our first experience.

The rest of the document is as follows. We review the literature on online retail, research costs, and price discrimination in Section 2. We formalize our hypothesis on the role of research costs in online purchases of discounted products. Section 3. We describe our empirical framework in Section 4. Sections 5-8 explain the experimental framework and provide details on execution. We summarize our findings and suggest future directions for research in Section 9.

2. Review of Related Literature

This article focuses on two areas of literature: marketing and economics. The first examines the influence of research costs on consumer decision-making, particularly in the context of online sales. The second is price discrimination at the second level, where monopolists determine their product lines so that higher-value consumers choose higher-quality, more expensive products themselves. We discuss each literature in turn.

Existing literature on the costs of research in online environments generally explores the

effect of reducing friction between information on consumer choice, well-being, and market structure. The first articles associated the decrease in research costs with an increase in competition, especially on prices (e.g. Bakos, 1997). Subsequent work has shown that online research may be more expensive than originally thought, perhaps because online shoppers have higher search costs than offline shoppers, and there may be considerable heterogeneity in search costs among online shoppers (Brynjolfsson and Smith 2000).

Lynch and Ariely (2000) found conditions in which increased transparency of online quality could reduce price competition, confirming that «in a competitive environment, the strategy of maintaining certain high search costs is likely to fail». Others have shown that some companies have found it advantageous to artificially increase research costs to promote price comparison. Ellison and Ellison (2009) show that firms in a market have an incentive to mask actual price and quality information sold online to reduce the tendency of search-sensitive consumers to comprehensively compare prices.

Ellison and Ellison (2005) show, as examples of obfuscation, how online retailers cost-effectively offer complex price menus, products that appear to be grouped but are no hidden prices, descriptions complex products, and other tactics designed «to allow the tendering process to take sufficient time». They argue that many advances in search engine technology that are supposed to facilitate the collection of information from consumers have subsequently been accompanied by business investment to hinder research.

Our work differs from these documents in that, rather than analyzing the competitive incentives that drive firms to impede research, we focus on the potential benefits of price discrimination within firms by increasing the costs of doing business. research. Although previous work has recognized the importance of heterogeneous research costs for consumers, none of our knowledge has examined how this heterogeneity can be exploited by a monopolistic enterprise in an online context.

Varian (1980) formulates a price discrimination plan that allows a company to extract surpluses of uninformed consumers through high prices and simultaneously sell to informed consumers through low prices using a mixed price balance strategy. Recent models with heterogeneous research also find equilibrium in which sellers adopt a mixed strategy

(Ratchford 2009). For example, Stahl (1996) finds a mixed strategic balance that, when implemented by two retailers, creates a market divide so that fully informed consumers always buy from the low-cost business while Uninformed consumers stop at less than a complete search and end up paying more.

Previous research has shown that differences in learning, provider switching costs, and risk perception when shopping online lead to heterogeneity in consumers' willingness to search (Ratchford 2009). These differences in consumer propensity to search allow retailers to selectively target price or product promotions with different margins among customers (Kopalle et al., 2009). The majority of research cost literature assumes that multiple firms have different levels of research costs (Ratchford, 2009) and thus separate the market based on the heterogeneity of consumers who are willing to pay for additional research costs. In this article, we focus on a single multi-product company that exploits this heterogeneity of search costs to separate the market to offer different segments to different products and prices.

We identify ways for a company to increase its profits by simultaneously offering high prices and low prices, the latter in the form of discounts, to attract informed and uninformed consumers by using search costs as a segmentation mechanism. As Ellison and Ellison (2009) predict about obfuscation, we also expect this approach to increase increments. But unlike their result, we expect the company to not only capture the fraction of customers who are willing to incur research costs but also to realize higher profits from customers who are unwilling to incur high search costs. Besides, there is no explicit or implicit intention to confuse customers as to the quality or actual price of an item in the online store. All items and prices are easily accessible to all consumers on the same website.

Ngwe (2016) shows that fashion stores use travel costs, away from shopping centers, as a way to encourage shoppers to choose to buy in stores - which leads to high travel costs to obtain low prices, or buying in regular downtown stores - resulting in lower travel costs but higher prices. This form and other forms of deliberate addition of research frictions as a tool of price discrimination are well known to brick and mortar store managers and can lead to concrete action. Unfortunately, online retailers cannot charge travel costs by physically placing their stores. The question is whether managers of online stores can add non-traditional and «virtual» search costs to the online shopping process of consumers to obtain the same benefit of price

discrimination. To make matters worse, online store managers often sell new items and liquidation items through the same channel. Also, it is not certain that the increase in search costs for a multi-brand non-proprietary online retailer may be a promising strategy when consumers can buy cheap substitutes from a competitor within a few clicks from distance only.

To practice price discrimination, companies must first be able to identify and distinguish sensitive customers or prospects from the high prices of low-price customers or prospects. The literature has generally defined three sets of characteristics associated with price sensitivity: the characteristics of a person, e.g., demographic characteristics, current behavioral characteristics, e.g., shopping characteristics, and past purchase characteristics. Besides, previous literature has shown that demographic characteristics are primitives of behavioral and purchasing characteristics (Kim, Srinivasan, and Wilcox 1999). From the variations in each of these characteristics, companies have tried to offer different prices to different price-sensitive customers.

Discrimination of prices by demographics: Previous research has linked demographic variables such as age, sex, income and geographic location, and consumer price sensitivity (Kim, Srinivasan and Wilcox, 1999). Traditional retailers and other businesses have used this knowledge to offer different prices based on demographic profiles. Supermarkets offer special discounts to retirees, cinemas offer students lower prices, grocery stores offer discounts depending on location, and educational institutions tend to offer scholarships and discounts to low-income families. And because many companies believe that men are more price-sensitive than women in certain categories, market comparisons have revealed considerable price differences between virtually identical products for men and women.

The use of price discrimination based on online demographics, on the other hand, has had mixed success. One obvious reason is that once consumers know who benefits from the rebate and who does not, they are incited to misrepresent or hide their identities. As a result, it becomes difficult for online businesses to distinguish between different demographic characteristics without requiring proof of identity, which would hinder many clients.

Price discrimination through buying behavior: Monitoring current buying behavior has also been a ubiquitous tool for brick and mortar retailers to assess a customer's price sensitivity. The way consumers browse, search, compare options, and negotiate prices has been widely

observed and used by sellers of expensive products that require more thought. These include consumer electronics, automobiles, and homes.

The use of behavioral variables observed during online shopping has had mixed results. On the one hand, and unlike demographic variables, search, click, and online navigation patterns are easily observable without the client having to reveal themselves. Hannak et al. (2014) showed that among 16 retailers and online travel providers, half of them showed that they offer different products and/or prices to online shoppers based on the users' cookies, clicks recorded, whether they have logged in or not, and the use of the mobile phone or PC to search for products and services. Yet, online, there are some disadvantages of behavior-based price discrimination if consumers discover the rules and can emulate the behavior that will allow them to get the lowest price. Amazon discovered it in 2000 when it offered the same DVD for fewer buyers who did not have a cookie in their browser, to offer new customers a cheaper price (CNN 2005). This encouraged Amazon's customers to delete their cookies or not to log in to the website to obtain lower prices. In summary, while behavioral price discrimination may be more observable than demographics or personal characteristics, it is also «playable» because clients are encouraged to learn to behave in a way that will allow them to get low prices.

Price discrimination based on purchase history: The third category of variables that have been associated with price sensitivity is related to previous purchases. Looking at five categories of groceries, Kim, Srinivasan, and Wilcox (1999) found that the characteristics of a person's purchase history, such as previous purchase frequency, quantity purchased, loyalty, and incidence cheaper items, are strongly associated with individual price sensitivity, far more than demographic characteristics. Offline grocery retailers and non-food retailers have successfully used this method by tracking past consumer purchases using loyalty cards (Lal and Bell 2003) and using this information to selectively target coupons, the prototype price discrimination method used. Online, pre-purchase couponing is also ubiquitous. It is observable and difficult to be playable because it is unlikely that consumers will buy items that are not their best options at the moment just so they can get future discounts.

The practice of price discrimination by the history of online shopping poses two major challenges. First, online retailers that tend not to have frequent or repeat purchases, such as auto, furniture, or apparel resellers, do not have good data to reliably estimate the sensitivity

of new customers to price. Second, while many industries such as groceries and fast food tend to offer fidelity-based discounts, the more you buy, the less expensive, some do the opposite. It has been shown that the insurance industry generally charges higher prices, not lower, for customer loyalty (Hughes 2008). The reasoning is that people who tend to stay with their insurance company for many years are opposed to the stress and hassle of shopping for better insurance quotes and are therefore less price sensitive. Insurance companies take advantage of this to increase loyalty customers' premiums year by year, a «loyalty tax» by charging more than new customers (Bankrate 2016). In the latter case, higher prices for loyalty, consumers are encouraged to move from one company to another. The reason most consumers do not do this is probably because of the considerable effort they put into doing it. In these cases, purchases that are infrequent or discouraging redemption, price discrimination from past purchases can have serious negative effects on firms.

In summary, there are three generalizable categories of variables associated with an individual's price sensitivity, according to previous literature. Furthermore, no variable is perfect for price discrimination in the sense that there are tradeoffs inherent in observability and «playability». In particular, some variables are better than others for the implementation of price discrimination policies, depending on the category of product or industry. Therefore, a good model for estimating price sensitivity should incorporate as many variables as possible of the three types, demographic, behavioral, and past purchases. In section 5, we try to estimate a price sensitivity model specific to each person in our context, taking into account the constraints linked to the data available to the company. Previously, we develop the research plan used in this paper.

3. Overview of Research Design

We use both field experience and historical purchase data analysis in our research plan. All data is provided by an online kitchen and tableware retailer in Benin. The online retailer sells branded products as well as articles under its brand. The company offers the largest choice of kitchen and tableware in the country. The articles are listed on the website under three catalogs: main catalog, sales catalog, and point-of-sale catalog. The main catalog contains all full price offers as well as some items at slightly reduced prices. The sales catalog contains moderate priced items, while the retail catalog contains very low-priced items, where the precise

thresholds between «light», «moderate» and «deep» vary over time. All products offered by the company are first listed in the main catalog and then progressively updated and listed in the other catalogs as new products are introduced.

Our research strategy has four components. The first is an exploratory field experience in which new visitors to the online store are randomly exposed to additional research frictions. In the second part, we use a regression model to classify existing consumers according to their price sensitivity. Third, we validate our classification by measuring the response rates to randomly assigned electronic bulletins. Fourth, we expose new and existing customers at the online store to additional research frictions and measure how treatment effects affect purchase rates, discounts and margins differently, depending on the sensitivity to the price estimated by the buyers. We cover the design and results in the following four sections.

4. Field Experiment I

In this experiment, we are looking for preliminary evidence that minor changes in website design can have significant effects on buyer behavior and purchasing results. We vary the presence of website features that may facilitate or hinder buyers from finding discounted items. We only include new visitors to the desktop version of the online store because they have relatively little information about the distribution of available products and prices. In assessing the results, we are particularly interested in the effects of processing on the discount levels of the transactions concluded and on the overall conversion rate. We anticipate that increased search costs will decrease the likelihood that price-insensitive customers will make the extra effort required to find products at very low prices, and that these customers will replace the full-priced items instead.

We experimented on the retailer's website for 15 days. During this period, all new website visitors were randomly assigned to the control group or one of three treatment groups with equal probability. New visitors are defined as customers who do not have the retailer's cookies on their machines and who register for a new account before making any purchase. Only visitors who used a desktop computer, laptop, or tablet were included in the study. A total of 195,806 customers were included in the experiment. Also, only consumers who access the site through the main home page have been included, excluding consumers who have visited the site using an electronic coupon, newsletter or link from a third-party website. During the

experiment, no other changes were made to the website. Descriptions of the testing and processing conditions follow. In each of the treatment conditions, neither the assortment of products available nor the prices of the products were different from those of the control condition.

The control condition was simply the website as it was at the time of the study. The website has elements designed to make it easier for consumers to find discounted items. Customers have three ways to find discount items: by clicking on a prominent link from the home page to the store catalog, sorting products by discount level in each catalog using a drop-down option, and by consulting the markers which highlight the discounts greater than 40%. In each of the processing conditions, we eliminate these elements to increase the effort required to locate items at reduced prices.

In this condition, we are eliminating the simplest path to discounts: the exit link from genre-specific landing pages. Other links to the outlet catalog can be found in the «selling» section of the website, requiring an additional click from a buyer to access the outlet catalog compared to those of the comparison group. This treatment represents a very slight increase in research friction which constitutes a solid test of our hypothesis.

Here we are removing the second easiest way to find discount items after the point of sale link: the ability for consumers to order product listings based on the level of discount. We are also removing the accompanying discount markers, which provide visual cues to identify high discounts. These items are widely used by online retailers to facilitate the search and navigation of buyers.

In this last treatment, we implement the largest increase in search frictions by deleting all the elements of the site deleted piecemeal in the first two treatments. This is an effort to significantly increase the magnitude of the research costs charged to buyers looking for discounts.

Our main goal is to find out whether obstructing consumers' search for high discount products leads some of them to buy fewer discount products and replace them with regular priced items to improve the retailer profitability. To assess the effects of each treatment, we take into account several variables:

Table 1 shows that clients in the three treatment groups purchased items at significantly lower discounts on average (9.5 to 11.3% off versus 15.5% off) and that they purchased more items at full price (65.5% to 68.7% versus 60.8%). Consequently, the average selling prices of the items purchased in the three treatments were considerably higher than in the control state, which confirms our initial hypothesis.

A natural concern is that if the search frictions for finding discount items are too large, then the expected result of reducing discount purchases could also be accompanied by lower conversion rates. This is of particular concern for new price-sensitive buyers. However, we did not find any significant decrease in conversion rates, measured by the number of transactions carried out. There was no significant difference in conversion rates between treatment group 3 and the control group. And the conversion rates were even slightly higher in treatments 1 and 2.

To verify the robustness of our main finding, we carried out a comparison between the treatments and the control groups at the level of the shopping basket (compared to the article) to compare the differences in purchasing behavior compared to the size and composition of the basket. Confirming the main results at the item level, Table 2 shows that the average discount on baskets purchased by consumers in two of the three treatments is significantly lower than that of the control group (12.1% to 13.2% compared to 14.8%). For treatment 3, it is slightly lower. The average basket size in all treatment groups was not significantly smaller than in the control group.

These results strongly support our hypothesis and demonstrate the effectiveness of the manipulations we have chosen. Our results show that online retailers can increase their margins without sacrificing conversion by slightly increasing the search frictions associated with their

reduced offers. In an environment without search frictions, price-insensitive consumers can locate discounted options «for free». By adding search frictions, online retailers can provide a semi-permeable way to close these consumers down to full-price options while offering discount options available to price-conscious buyers extra efforts to find them. In the following sections, we examine in more detail the mechanism underlying our main results.

5. Measuring Price Sensitivity

A full test of our hypothesis requires assessing the differential impact of research frictions on price-insensitive customers compared to price-sensitive customers. The expected impact on purchasing behavior is expected to disproportionately affect the former. In this section, we develop a sparse empirical model of price sensitivity for buyers in our context. This serves two purposes. First, it uses a new and unique set of data to identify the relevant determinants of price sensitivity. Second, it provides us with a way to pre-classify consumers based on their expected price sensitivity. We then use the predicted values from this model as an approximation of a buyer's price sensitivity. After estimating the model, we assess its predictive accuracy by comparing the behavior of groups of pre-classified buyers in a validation experiment in the field.

The data in this analysis consists of historical sales registers at the retailer's transaction level since its creation in October 2016 until October 2019. Over 2.1 million individual items were sold, 318,050 consumers have made more than a million transactions during this period. Each record (an item sold) contains buyer attributes, product attributes, and transaction attributes. Tables 3 and 4 describe the available data and basic summary statistics on operations.

We estimate a simple price sensitivity model to determine, through field monitoring experience, whether price sensitive (rather than insensitive) buyers are more willing to bear the search costs of finding discounted items online. As the primary objective of the estimation is not to identify the primitives of consumer behavior, but to distinguish consumers who are price-insensitive from price-sensitive consumers, we adopt a parsimonious model that aims to explain the discount at the level from the basket of transactions made. The underlying assumption is that very price-sensitive buyer is more likely to buy discounted items than price-sensitive buyers. The categories of variables explaining the preference for discounts include demographic characteristics, past transaction behavior, and purchasing conditions known to be associated with the search for discount behavior. For each buyer I in month t, we have:

discount_preference?? t=?? (consumer_attributes??, shopping_behavior??, t-1, shopping_conditions?? t)

We carry out a series of Tobit regressions of the average basket discount on these covariates and present the estimates in Table 5. We use the Tobit model since the markdowns on the baskets are a censored approximation to the left (at zero) of the price sensitivity, our conceptual

variable of interest. To assess the relative importance of demographics, past transaction behavior, and current purchasing conditions, we estimate separate regressions for each subcategory of explanatory variables in addition to the full model.

In general, we find that the relationships between consumer attributes, past purchasing behavior, current purchasing conditions, and current purchasing behavior are strong and robust when using different choices of covariates. Each category of explanatory variables improves the model's ability to predict preference for discounts. The observed discounts are lower for men and older customers. They are higher for customers who have previously purchased products at higher prices, used coupons, and more store-branded items. In the meantime, markdowns are lower for consumers who redeem coupons for an ongoing purchase, who use the store credit and who have more transactions already made.

We use the empirical model estimated in this section to pre-classify buyers according to their level of price sensitivity to articulate the mechanism behind our main result in Field Experience 1. Indeed, we use all the information available on consumers to obtain this classification, by assigning weights to each variable according to its estimated coefficient. We consider this to be superior to an ad hoc classification, for example, by grouping buyers according to the average discount in their purchase history. However, we also recognize the shortcomings of this approach due to the aggregation of information, the absence of purchasing or purchasing decisions, and the evolution of the assortment over time. To increase our confidence in the resulting classification, we seek to establish its external validity. In the next section, we describe how we validate our classification model using electronic newsletters as part of a field experiment.

EQUATION	VARIABLES	(1) basket_markdown	(2) basket_markdown	(3) basket_markdown	(4) basket_markdown
model	male	-0.0254***			-0.0197***
		(0.000865)			(0.00108)
	cust_age	-1.26e-06***			1.11e-06***
		(1.06e-07)			(1.26e-07)
	prev_markdown		0.396***		0.392***
			(0.00200)		(0.00199)

	prev_coupon		0.0325***		0.0458***
			(0.000907)		(0.000948)
	order		-0.000645***		0.000500***
			(4.90e-05)		(4.92e-05)
	prev_sb_ratio		0.0170***		0.0125***
			(0.00125)		(0.00124)
	cust_hist		1.74e-05***		1.49e-05***
			(2.05e-06)		(2.04e-06)
	wallet			0.0150***	-0.00323**
				(0.00143)	(0.00147)
	coupon			-0.0336***	-0.0544***
				(0.000766)	(0.000956)
	Constant	0.141***	-0.0166***	0.0704***	0.0278***
		(0.00748)	(0.000760)	(0.00125)	(0.0100)
sigma	Constant	0.331***	0.310***	0.329***	0.307***
		(0.000338)	(0.000392)	(0.000336)	(0.000388)

	Billing region FE	yes			yes
	Month FE				yes

	Observations	1,112,297	698,456	1,112,298	698,456
	Pseudo R2	0.00324	0.0576	0.0105	0.0729

6. Validation Experiment

In this section, we test the external validity of our empirical model of price sensitivity. We sent email newsletters about discounts and rebates to randomly assigned consumers and we checked to see if the classification determined by the Tobit model in the previous section was associated with a higher response rate for emails. discount (as opposed to non-discount) price-sensitive consumers (as opposed to price-insensitive). We find that consumers whose model predicts that they will be price sensitive are more receptive to messages that include a discount element. Experimental Design

We include in this experience the entire mailing list of the company, which has 246,688 consumers. A consumer signs up for the mailing list by providing their email address to the business by creating an account, signing up for updates, or requesting a coupon. Consumers were randomly assigned to two groups, now Group 1 and Group 2, and each group received a schedule of newsletters with and without discounts, as shown in Table 6.

The control bulletins sent on Sunday were not focused on discounts and were identical from one group to another, whereas within each day, only the discount and discount messages were different from one group to another. For each newsletter sent, we observe if the email has been opened and which link in the email, if any, has been clicked by the recipient. We also observe all transactions on the website, which we can link to consumers participating in the experience by their email address.

The product categories in electronic newsletters varied from day to day but remained constant between control and treatment groups within a few days. We have also ensured that all the creative elements of the newsletters remain constant so that only the messages at a reduced price (for example «up to -40%») and all the price information vary during execution. This variation in messages was reflected in the subject of the message.

The frequency with which customers on the mailing list choose to receive newsletters varies. 63.34% of subscribers receive them every day, 4.32% three times a week, and 35.35% once a week. The schedule for the newsletter presented in Table 6 was designed to obtain maximum variation among consumers, regardless of their frequency, as well as to minimize the effects of the day of the week.

To establish the validity of the classification on the Tobit model, we generate predicted values from the model according to the purchase history of each consumer before the experience of the newsletter. Our model can be taken as an indication of price sensitivity if consumers we expect to be more (rather than less) price-sensitive have a greater propensity to open and click on discount emails. Since the final experience will compare the average purchasing behavior of one group of consumers to another, price-sensitive, or insensitive, the

classification model should be precise only at the group level rather than at the level individual. Table 7: Regressions on newsletter response variables

VARIABLES	(1) open	(2) click	(3) click open
disc	-0.0149*	-0.00103	0.0188
	(0.00811)	(0.00417)	(0.0181)
price_sensitivity	0.0402	0.0248	0.0868
	(0.0297)	(0.0153)	(0.0671)
disc*price_sensitivity	0.125***	0.0505**	0.0728
	(0.0420)	(0.0216)	(0.0933)
Constant	0.207***	0.0328***	0.152***
	(0.00630)	(0.00324)	(0.0141)

day dummies	yes	yes	yes

Observations	106,534	106,534	22,043
R-squared	0.002	0.001	0.003

We regress the variables of the results of the bulletin according to our variables of interest. Each record in the following regressions is an email-client pair. The dependent variables are binary, where success is either an open email or a clicked email. The independent variables are an email discount dummy, the price sensitivity predicted from the empirical model, the interaction between email discount and price sensitivity, and the day of the week.

We find that our measure of price sensitivity is positively correlated with the probability of responding to a discount email compared to an email without discounts. Table 7 examines the relationship between price sensitivity and three response variables: (1) if the email was opened, (2) if a link in the email was clicked, and (3) if a link was clicked. provided the email is opened. We find that price sensitivity is positively associated with the first two measures, whose sample size was almost five times that of the third measure.

The exogenous allocation of newsletters with and without discounts provides us with an additional means of pre-classifying consumers. Rather than relying on purchase history to infer price sensitivity, we can project the probability of responding to a newsletter on consumer attributes versus a newsletter without discount and thus have a means based on a model for

predicting price sensitivity. Now that we have established the validity of our price sensitivity classification approach, we can use it to characterize how purchasing behaviors differ in online retail environments by price sensitivity.

7. Field Experiment II

In our last experiment, we study how the response of buyers to additional search frictions varies depending on price sensitivity. As with Field Experience I, we expose consumers to different versions of the online store, each with an additional search friction element. We compare the results against the control over retailers' performance measures and use our predictive models from the previous section to characterize the heterogeneity of consumer response.

We carried out this experience from June 1 to 15, 2019 on the desktop and tablet versions of the online store. In the following analysis, we use the data from June 2 to 14 to eliminate the possibility of contamination from the start and end of implantation. All consumers were randomly assigned to the control group or one of the four treatment groups with equal

probability. While in Field Experience I,we only included new visitors entering through the

main landing pages, here we include new consumers as well as returning consumers, regardless

of which page they are viewing in first. The processing conditions are as follows:

Treatment 1: Removal of links from main pages to points of sale and sales sections of the

Unlike experiment 1, we separate the removal of the discount flags and the sorting options into two different treatments for reasons of completeness. We are also adding a fourth treatment, the use of banners without discounts throughout the site because this communication approach is the equivalent on the website of the email treatments used in the previous validation experience.

Before examining the impact of price sensitivity on consumers' propensity to find and buy

discounted items, we perform the same analysis as in Table 8, which groups all types of consumers. By further validating the main conclusions of experiment I, this time by including current customers rather than only new customers, we find that the removal of the discount flags, sorting by discount and discount banners (treatments 2 to 4) decreases both the average discount of items purchased and the impact of purchasing items on sale. As in the past, this objective is achieved without reducing conversion rates. An exception to this rule, and contrary to the conclusions of experiment I, is the null effect of the removal of the links to the points of sale and the sales from the home page (treatment 1). It appears that current customers were not as dissuaded as new customers from finding the high discounts at the point of sale section of the website. This is not so surprising since many buyers probably already knew of the existence of the point of sale and only had to go through an additional click to find it. In summary, except for this processing, the addition of research costs has a very similar qualitative impact on new and existing customers.

A more precise test of our forecasts is to show an interaction between a buyer's price sensitivity and their willingness to incur search costs to find discounted items. The use of regular customers, while changing the «navigability» of the website is, in our opinion, a very rigorous test of this prediction. First of all, customers have memories and we expect them to remember that very discounted items exist on the platform. Second, our manipulations are quite subtle (that is, they slightly increase search costs) and do not cause any change in sales prices or the assortment of products. Third, fashion retail is a category in which buyers have a pretty

good idea of when prices are high or a good deal and maybe more motivated to leave the website if they can't find a discount. Despite these challenges, we find that price sensitivity still plays a moderating role in the impact of research frictions on the likelihood of purchasing items at reduced prices.

In Table 9, we group consumers into three quartiles based on their price sensitivity, as set out in Section 5. We find that price-insensitive consumers are more likely to buy items at full price across the entire market (see the first row). In three of the four processing conditions, we observe a statistically significant increase in the proportion of full-price items purchased by customers with little price sensitivity. Equally remarkable, this is not the case for consumers sensitive to average or high prices, who willingly incur research costs to benefit from discounts. This result provides additional evidence, by including current users and adding other forms of search costs to the website, online retailers can improve margins and, therefore, their profitability, by deliberately adding costs friction.

8. Conclusion

Online retail represents a rapidly growing proportion of all retail sales. However, margins in online retail can often be lower than in offline retail. One of the guesses behind this discrepancy is that online sellers are less able to discriminate by price than offline sellers. In this article, we explore how charging research costs to online buyers can improve gross margins by serving as a sorting mechanism among customers.

We find encouraging evidence that minor changes to the design of an online store (i.e. friction) can significantly improve its margins and profitability. By simply increasing search frictions - by removing selected links, reducing product sorting options, and limiting visual markers - online sellers can make more full-price sales to price-sensitive buyers who have higher search costs. Although significant enough to influence the purchasing behavior of price-insensitive buyers, these search frictions are minor enough for price-sensitive buyers to make the extra effort to locate discounted items. As a result, the average selling price increases as the number of items sold remains stable.

Our results have direct implications for online sellers. Without changing prices or the product mix, online sellers can improve their margins by making subtle changes to their website design. We note that this is essentially free manipulation with low data requirements, but which can lead to significant margin gains. Our point of view implies that by unduly favoring the ease of research and purchase, online sellers give up their gross margins by granting discounts to price-insensitive consumers.

Adding search frictions, we expected some decrease in conversion: by making high discounts more difficult to find, some price-sensitive consumers might choose not to spend more time and effort on the website and leave. We find it rather surprising that the conversion rates are not lower, and even higher in some of our processing conditions. This can be explained by branding effects: by hiding high discounts, new customers can perceive the seller as being of better quality, which increases the number of purchases. Another possible explanation can be explained by the overload of choice associated with very large assortments so that a more limited set of considerations can improve the expected utility (e.g., Iyengar and Lepper, 2000). For our research question, the point to remember is that any improvement in unit margins does not necessarily have to be at the expense of weaker conversions. This is a speculative statement

and more research is needed in the online retail field to validate it and provide boundary conditions.

Our research also suggests that certain online browsing behaviors can be painful enough or take enough time for buyers to pay more. We consider it useful to conduct future research to determine the specific properties of online interaction that consumers find most demanding. This can provide useful advice for a wide range of applications, from online store design to digital advertising.

Bakos, J.Yannis. 1997. Reducing buyer search costs: Implications for electronic marketplaces. Management Science, 43(12), pp.1676-1692.

Bankrate. 2016. Insurance `loyalty tax' in the crosshairs. By Jay MacDonald, published January 7, 2016. Accessed on January 7, 2016, at

Bell, David R., Teck-Hua Ho, and Christopher S. Tang. 1998. Determining where to shop: Fixed and variable costs of shopping. Journal of Marketing Research, pp.352-369.

Brynjolfsson, Erik, and Smith, Michael D. 2000. Frictionless commerce? A comparison of Internet and conventional retailers. Management Science, 46(4), pp.563-585.

Eisingerich, A.B. and Kretschmer, T., 2008. In e-commerce, more is more. Harvard Business Review, 86(3), pp.20-21.

Ellison, Glen, and Ellison, Sara F., 2005. Lessons about Markets from the Internet. Journal of Economic Perspectives, pp.139-158.

Ellison, Glen and Ellison, Sara F. 2009. Search, Obfuscation, and Price Elasticities on the Internet. Econometrica, 77(2), pp.427-452.

Ellison, Glenn and Wolitzky, Alexander. 2012. A search cost model of obfuscation. The RAND Journal of Economics, 43(3), pp.417-441.

Hannak, Aniko, Gary Soeller, David Lazer, Alan Mislove, and Christo Wilson. 2014. Measuring price discrimination and steering on e-commerce web sites. In Proceedings of the 2014 Conference on Internet Measurement Conference, ACM. pp. 305-318

Hughes, Ian. 2008. Customer loyalty in the UK general insurance market. Journal of Direct, Data and Digital Marketing Practice, 10(1), pp.29-42.

Iyengar Sheena S., Lepper Mark R. (2000), «When Choice Is Demotivating: Can One Desire Too Much of a Good Thing?» Journal of Personality and

Kim, Byung-Do, Srinivasan, Kannan and Wilcox, Ronald T. 1999. Identifying price-sensitive consumers: the relative merits of demographic vs. purchase pattern information. Journal of Retailing, 75(2), pp.173-193.

Kopalle, Praveen, Biswas, Dipayan, Chintagunta, Pradeep K., Fan, Jia, Pauwels, Koen, Ratchford, Brian T. and Sills, James A. 2009. Retailer pricing and competitive effects. Journal of Retailing, 85(1), pp.56-70.

Lal, Rajiv, and David E. Bell. 2003. The impact of frequent shopper programs in grocery retailing." Quantitative marketing and economics 1.2, pp.179-202.

Lynch Jr, John G., and Ariely, Dan. 2000. Wine online: Search costs affect competition on price, quality, and distribution. Marketing Science, 19(1), pp.83-103.

Monga, Ashwani, and Saini, Ritesh. 2009. The currency of search: How spending time on

search is not the same as spending money. Journal of Retailing, 85(3), pp.245-257. Ngwe, Donald. 2016.Why Outlet Stores Exist: Averting Cannibalization in Product Line

Ratchford, Brian T. 2009. Online pricing: review and directions for research. Journal of Interactive Marketing, 23(1), pp.82-90.

Rigby, D., 2011. The future of shopping. Harvard Business Review, 89(12), pp.65-76.

Stahl, Dale O. 1996. Oligopolistic pricing with heterogeneous consumer search. International Journal of Industrial Organization, 14(2), pp.243-268.

Teixeira, Thales S., and Sunil Gupta. 2015. Can Retailers Win Back Shoppers Who Browse then Buy Online? Harvard Business Review 93(9).

Varian, Hal R. 1980. A model of sales. The American Economic Review, pp.651-659. Wilson, Chris M. 2010. Ordered search and equilibrium obfuscation. International Journal of Industrial Organization, 28(5), pp.496-506.

Zettelmeyer, Florian. 2000. Expanding to the Internet: Pricing and communications strategies when firms compete on multiple channels. Journal of Marketing Research,