The (underestimated) role of product data for winning online retail

(1)

IN

DEGREE PROJECT INDUSTRIAL MANAGEMENT, SECOND CYCLE, 30 CREDITS

STOCKHOLM SWEDEN 2020,

The (underestimated) role of product data for winning online retail

JOHN BOLMGREN

HENRIK LINDSTRÖM

(2)

(3)

The (underestimated) role of product data for winning online retail

by

John Bolmgren Henrik Lindström

Master of Science Thesis TRITA-ITM-EX 2020:365 KTH Industrial Engineering and Management

Industrial Management

(4)

Den (underskattade) rollen av produktdata för att vinna e-handeln

av

John Bolmgren Henrik Lindström

Examensarbete TRITA-ITM-EX 2020:365 KTH Industriell teknik och management

Industriell ekonomi och organisation SE-100 44 STOCKHOLM

(5)

Master of Science Thesis TRITA-ITM-EX 2020:365

The (underestimated) role of product data for winning online retail

John Bolmgren Henrik Lindström

Approved

2020-06-15

Examiner

Lars Uppvall

Supervisor

Pernilla Ulfvengren

Commissioner Contact person

Abstract

As E-commerce continues to take market share from traditional brick and mortar businesses, there are few choices left for managers apart from migrating their sales online. While the topic of online adoption has been studied extensively, this thesis attempts to investigate one of the major drivers of complexity within the industry - the role of structured product data. The study was performed on a major Nordic online retailer, and identified a set of six guiding propositions on the topic of structured product data in e-commerce from interviews with industry professionals. Contemporary data science literature contributes to the body of evidence suggesting a strategically prioritized focus on creating and maintaining structured product data is the way of the future for e- commerce, aligning with much of the interview results. Furthermore, the propositions were thoroughly examined through multiple linear regression analysis on data from the same firm. The study gives empirical support for significant positive impact on most studied metrics from having structured product data available on the website as well as within the internal systems, with slight discrepancies across product categories.

Key-words

(6)

Examensarbete TRITA-ITM-EX 2020:365

Den (underskattade) rollen av produktdata för att vinna e-handeln

John Bolmgren Henrik Lindström

Godkänt

2020-06-15

Examinator

Lars Uppvall

Handledare

Pernilla Ulfvengren

Uppdragsgivare Kontaktperson

Sammanfattning

I takt med att e-handeln fortsätter att ta marknadsandelar från traditionella fysiska butiker finns det få alternativ för ledningsgrupper förutom att migrera sin försäljning online.

Online-migrering som ämne har studerats i stor utsträckning tidigare, men denna uppsats försöker utforska en av huvuddrivarna till branschens komplexitet – rollen av strukturerad produktdata. Studien gjordes på en större nordisk e-handlare, och identifierade sex ledande teman inom ämnet för strukturerade produktdata i e-handel genom intervjuer med experter på bolaget. Kontemporär litteratur inom datavetenskapen bidrar till belägg för att ett strategiskt prioriterat fokus på att skapa och managera strukturerad produktdata är vägen framåt för e-handeln, vilket ligger i linje med resultaten från intervjuerna inom studien. Vidare analyserades de identifierade temana genom multipel linjär regression genom data från bolaget. Studien ger empiriska belägg för att strukturerad produktdata på e-handlarens hemsida samt i de interna systemen ger signifikant och positiv påverkan på de flesta responsvariabler, med vissa diskrepanser mellan produktkategorier.

Nyckelord

E-commerce, Product data, Structured product data

(7)

List of Figures

2.1 Illustration of structured vs. unstructured product page . . . . 13 2.2 Table of success factors from Varela et al. (2017) . . . 19 3.1 Example QQ plot for the pageviews model of category bath . 34 4.1 Example of different types of searches . . . 43

(11)

List of Tables

4.1 Summary of the image count attribute regression coefficient

per category . . . 55

4.2 Summary of the category specific attribute regression coefficient per category . . . 56

4.3 Summary of the base attribute regression coefficient per category 57 4.4 Summary of the standard attribute regression coefficient per category . . . 57

4.5 Summary of the dimensions attribute regression coefficient per category . . . 58

4.6 Summary of the title length attribute regression coefficient per category . . . 59

4.7 Summary of the description length attribute regression coefficient per category . . . 59

A.1 Regression table for category full . . . 86

A.2 Regression table for category bath . . . 90

A.3 Regression table for category construction . . . 94

A.4 Regression table for category floor . . . 97

(12)

A.5 Regression table for category int . . . 102 A.6 Regression table for category kitchen . . . 106 A.7 Regression table for category garden . . . 110

(13)

Chapter 1 Introduction

E-commerce has won tremendous ground in the past thirty years and its growth has been accelerating even further in recent years. Today, there is no debate on whether e-commerce will account for a significant share of the consumer retail industry going forward, the question is rather how large that share will ultimately become. Strong structural trends such as digitization, online-adoption, demographic shifts and most recently the consequences of the Covid-19 pandemic all support continued growth of the e-commerce industry. Amazon has become one of the world’s most valuable companies with a significant part of its revenues attributable to its e-commerce business.

In light of these structural trends, many (if not most) brick-and-mortar businesses have been forced to adapt to the new market conditions by taking their business online while new online-native businesses have entered the market. The competition for consumers’ online spending has become fierce and the dynamics have shifted significantly as retailers, search engine com-

(14)

panies, online aggregators and manufacturing companies all want their piece of the seemingly ever-growing e-commerce sector. The tough competition and rapid changes have triggered researchers to ask the question of how a successful e-commerce business is built, what the key success factors are and how technological developments can be leveraged in order to win over the hearts and minds of online shoppers.

This paper asks the question of what role data generally, and product data specifically, plays in the realm of e-commerce. When products are not on physical display but presented through images, descriptions and attributes and when stores are not visible from the street but accessed through specific entries on a keyboard or smartphone - companies must adapt all parts of its business, from marketing to purchasing, in order to find ways to survive and thrive. In this study, we investigate how product data is used in all parts of an e-commerce business, what role it plays, how it should be treated and prioritized as well as how it relates to a company’s ability to prevail in a harshly competitive landscape.

Since the theme of data specifically applied to e-commerce has not been widely discussed, we approach the topic with an open mind and simply ask the question of what role it plays in creating a successful online business.

This is done in the context of a case study involving interviews with several industry professionals working at different functions in a large Nordic e- commerce company, henceforth referred to as ”the Company”. This part of the paper will henceforth be referred to as the ”proposition analysis”.

Insights and conclusions from the proposition analysis are then consolidated to form a set of propositions about the role and significance of data in e-

(15)

commerce that are further investigated and benchmarked against previous research and tested using statistical methods on company data.

1.1 Scope and delimitations of the paper

For purposes of clarity we begin by giving a definition of how we define e-commerce. A general definition was proposed by Frost et al. (2018): “E- commerce refers to the online transactions: selling goods and services on the internet, either in one transaction (e.g., Amazon, Zappos, Ebay, Expedia) or through an ongoing transaction (e.g., Netflix, Match.com, Linkedin etc.)”.

Given that this paper focuses on the trade of physical goods over the internet, we narrow the definition used in this study to: E-commerce refers to the transaction of physical goods over the internet.

We will also clarify what we mean by data. There are many different kinds of data in the e-commerce space. The different kinds of data are collected and used for different purposes and while some of the data can be seen as generic for all businesses, sales data being the obvious example, other kinds of data exist more or less uniquely in the e-commerce sector. Akter and Wamba (2016) divides e-commerce data into four categories:

(a) Transactional data (b) Click-stream data

(c) Data in the form of video (d) Voice data

(16)

Yet again, given our narrowed scope for this paper and the subsequent defi- nitional difference on the term E-commerce, only the first two types in Akter and Wamba (2016) categorization applies to our definition in a meaningful way. We suggest a different approach to the classification of data types based on the source of the collected data as follows:

(a) Transactional data: Refers to data collected from transactions with the customer. This data type includes sales, profitability, pricing and return rates to name a few.

(b) Behavioural data: Refers to data collected from the customers’ online behaviour and interactions with the e-commerce platform. This data type includes conversion rates, site visits, session lengths and points of entrance among others.

(c) Logistical data: Refers to data collected from the process of shipping products to customers. This data type includes delivery times, delivery methods, stock levels etc.

(d) Product data: Refers to data collected from the products themselves.

This type of data includes product features, images, titles and descriptions.

The focus of this paper going forward will be mainly on the impact of product data.

(17)

1.2 Setting the stage for discussing e-commerce data

For purposes of clarity, some key concepts are defined as they relate to e- commerce websites. While they are not commonly used in the literature on e-commerce, these concepts play an integral role in understanding the role of product data and will be referred to throughout this paper.

• Home page: The home page is the webpage that a customer is di- rected to if they enter the store using only the website’s domain without additions. Commonly, the home page in the e-commerce context is the first point of contact with the customer and can be used to browse the website’s assortment. A real-life analogy is to the entrance of a large mall where a visitor is guided by signs to the appropriate store or department.

• Landing page: Landing pages display several products within the same category or with other kinds of similarities. These pages are often used as interstages between the home page and the product page. Here, customers can browse through a subset of the website’s assortment, often with the help of filters. A real-life analogy is to the entrance of a store within the mall that sells a specific kind of product.

• Product page: The product page is where the customer can make the actual purchase of a product. The product page is dedicated to a specific product and contains information and images relating to that product.

(18)

This paper makes use of both qualitative and quantitative methods of study in order to answer the research questions (see section 1.3). Firstly a proposition analysis was conducted as a case study of a large Nordic e-commerce company - as proposed by Baxter and Jack (2008). Interviews were conducted with the aim of deriving propositions from leading professionals at the company relating to product data and its role in e-commerce. Furthermore, a rigorous exploration of current literature on the subject was conducted to give perspective to the data gathered from the proposition analysis. Given these propositions (see below), proposition validation was dedicated towards testing their legitimacy in the context of this specific company in the form of a quantitative analysis on company data (see section 4.3). While we recognize that a single company cannot be used as a generalization for the industry as a whole, since it is bound by its specific circumstances, we consider the company a good subject for study given its presence in many different product categories as well as its size and market share. The conclusions might not be upheld in the general case, especially the conclusions from quantitative analysis drawn from company data, however we will try to argue as generally as possible since it is the sense of the authors that the assumptions laid out in the hypotheses are broadly considered to be true, even outside of the Company and their subsidiaries.

For reference, the following propositions were extracted from the proposition analysis, and are elaborated on in section 4.1:

• Proposition 1: Structured product data, in contrast to its unstructured counterpart, is significantly more valuable in terms of its potential application in all parts of the e-commerce value chain

(19)

• Proposition 2: Structured product data improve navigation

• Proposition 3: Structured product data is crucial in search engine optimization

• Proposition 4: Optimizing product titles is very important for long-tail SEO, and structured product data makes them seamless to create

• Proposition 5: High quality product images are important for selling products online

• Proposition 6: Structured data is highly valuable for business intelli- gence and on-site curation

1.3 Research questions

The following research questions are proposed for the study, and are inti- mately linked to the identified propositions:

1. What role do online retailers place on structured product data?

2. How well does the online retailers’ appreciation of structured product data align with measurable outcomes?

(20)

Chapter 2 Theoretical background

The purpose of Chapter 2 is to give the reader an introduction to contemporary academic literature in the field of e-commerce in general, and product data for the former in particular. Furthermore, this chapter provides a critical academic reference for discussing the identified propositions defined in section 4.1. Given that the interviews that were conducted within the study were confined to a single company, this literature review is deemed necessary in indicating whether the findings from the proposition identification have the potential of being considered valid also in the generalized case.

2.1 The current state of academic e-commerce literature

The role of data in e-commerce has been studied from multiple perspectives.

Little has been written in the field of management on the necessity of placing

(21)

data at the heart of every e-commerce business. A surprising fact given that daily operations in these businesses has data management as a core struggle, taking up the vast majority of all operational activities. However, a lot has been written from a technological perspective ranging from the potential analytical values that could be extracted from e-commerce data to the field of data science that have extensively studied methods for mining, deploying and enriching product data as well as the potential of that same data for search engine- and UX applications.

From the perspective of business and management the topic has mainly been approached by more generally studying critical success factors in e-commerce and also the potential of Big Data Analysis (BDA) in the e-commerce setting given its native stance as an industry with great access to many kinds of data in tremendous volumes. An extensive positional paper on the current stance of research on BDA in e-commerce is offered by Akter and Wamba (2016) from which we have drawn several references for this paper. The overall conclusion from the study of BDA-applications in e-commerce is nicely summed up by Loebbecke and Picot (2015) as “the platform for growth of employment, increased productivity, and increased consumer surplus”.

The data science field has approached the topic of data in e-commerce from a more practical standpoint. The value of having high quality data is seen as axiomatic and much of the research is centered around how data on product specifications, reviews and prices can be mined, structured and leveraged to fit applications such as search engine optimization, product catalogue creation and product matching. The topic of product feature extraction from unstructured data sources online has made significant progress in recent

(22)

years and the most successful methods from the area are summarized by Rao and Sashikuma (2016). Methods for solving the not at all trivial problem of matching identical products from different sources has been proposed by Ristoski et al. (2018) and a method for synthesizing product catalogues from unstructured data sources was given by Ristoski et al. (2018).

The common theme of the data science papers on e-commerce data has been that proposed applications are rarely aimed at the e-tailers themselves, but rather in favour of platform-type applications such as product search engines and other recommendation engines for consumer use. This approach is taken by Nguyen et al. (2011) who describes a method for synthesizing products for online catalogues using novel methods in computer science with the explicit aim of developing methods for creating generalized product catalogues that draw data from many e-commerce websites with the aim of consolidation.

On the same general topic, Ristoski et al. (2018) lay out a method for both categorization of products and matching of products using neural language models and deep learning. The paper mentions Google Product Search ex- plicitly as a target use-case for their methods, but implicitly makes the same assumption as Nguyen et al. (2011), namely that e-commerce companies have already solved the problem of data quality and reliability internally and that the next natural step in the data-accessibility-value-chain is democratizing the data through consolidation of data from all e-commerce actors.

The aim of the following sections in the literature review is to provide an overview of recent academic efforts adjacent to the topic of data in e- commerce. Publications in the field are dominated by data science papers which we will try and summarize in understandable language for those not

(23)

versed in the field. The key point is to stress two important facts that become evident from the literature:

1. There is a vibrant discussion in the data science community on methods for, and applications of, e-commerce data driven and financed not primarily by the e-commerce sector but by the technology giants and search engine companies. The value of structured product data is axiomatic and much of the research rests on the assumption that high quality data is already “out-there” and the problem to be solved thus becomes 1) collecting the data, and 2) structuring the collected data.

2. Regardless which field of study we turn to, there is little emphasis on the value of data for the e-commerce companies themselves. Very little is written on topics such as management priorities, operational challenges and marketing opportunities in e-commerce in general. Particularly, none of that research has the same axiomatic conviction on the value of data that permeate the data science community.

2.2 E-commerce from the perspective of Data Science

Sticking to our categorization of e-commerce data it becomes evident that the focus of data science research is on product data. Keep in mind that much of this research is aimed at finding solutions for consolidated product databases such as price comparison sites and product search engines, or to steal an expression from Krys and Bagheri (2016): the research is set on

(24)

finding solutions for “online aggregators”. The interest in product data has emerged as the growth of e-commerce has continued to accelerate (Nguyen et al., 2011). We will focus this part of the literature review to text form data, meaning that media is left for a later part of the discussion.

2.2.1 Product data come in many shapes

An important distinction that is often made in the data science community (but rarely if ever made in the business community) in terms of e- commerce data is whether a set of data is unstructured, semi-structured or structured (Rao and Sashikuma, 2016). Unstructured data is difficult to use in its original form for applications ranging from BDA (Kang et al., 2003) to search engine optimization and product catalogue creation (Nguyen et al., 2011). Nguyen et al. (2011) conclude on the topic of structured data that

”This structured data is fundamental to drive the user experience: it enables faceted search, comparison of products based on their specifications, and ranking of products based on their attributes.”. To shed some light on the distinction between structured and unstructured data we refer the reader to Figure 3.1. In the case of the unstructured product page the data is in free-text format and even though the reader can get a sense of the product, the ability to leverage this data is very limited for most applications. A basic example relates to on-site-navigation: if there is no product level structured data, then there is no possibility to create filtering functionality that the user can apply to find relevant results among large assortments of products. Other examples can be applying AI/ML-algorithms to unstructured data generally yields inferior results compared to structured data (Shimada and Endo, 2005)

(25)

Figure 2.1: Illustration of structured vs. unstructured product page and the ability to generate relevant search results is significantly improved by searching in a structured database compared to an unstructured (Duan et al., 2013).

The important distinction between the different kinds of product data and consequently the necessity of structured data, preferably in the form of key- value-pairs (i.e. a key along with a connected value, where “Color” is an example of a key associated with the value “Blue”) has emerged as an integral component for achieving better customer experience (Ristoski et al., 2018) as well as improved search performance (Nguyen et al., 2011).

2.2.2 Towards the mighty concept of a structured prod- uct catalogue

As such, the task of the data science research in the area can be thought of as three-fold, remembering it’s desired application for “online aggregators”:

1) Collect the (unstructured or structured according to unknown structure)

(26)

raw data from publicly available online sources on the web. 2) Make the unstructured data structured by a) categorizing the products along a predefined “category-tree” and b) extract key-value-pairs according to a predefined schema of keys associated with the chosen category from the unstructured product data. 3) Aggregate the products into a product catalogue. (Rao and Sashikuma, 2016). Several methods have been proposed for achieving these three tasks including web-mining via crawler-scripts for collection, regular expressions and/or machine learning for structuring data and finally other machine learning methods and feature comparison for aggregation. Worth noting is that all of the efforts in this area are done with the objective of building fully automated systems for achieving all of the steps above.

In light of this paper, where emphasis lies on the e-commerce sector, representing the data source for this field of research, the same three-fold process can be successfully applied if the data source in 1) is exchanged to the e-tailers supplier. Effectively moving the whole process one step backward in what can be thought of as the “data-supply-chain” or “layers of data consolidation”. Here it should be recognized that many suppliers of the retail industry in general and the e-commerce sector in particular haven’t got sophisticated websites making the full set of raw data publicly available. However, substi- tuting a supplier website to a supplier product database and the comparison still holds true. Efforts has been made to use supplier websites as the source of raw data, though to a significantly lesser extent then using the e-commerce websites directly (Walther et al., 2010).

Given the similarities in approaches for the data-supply-chain between the use cases it is relevant to adress some key challenges faced by the data science

(27)

community adjacent to these tasks. Rao and Sashikuma (2016) describe the major hurdles in structuring data faced by researchers. These include the volatility of the source data (i.e. the e-commerce websites), the challenge with different data formats from different sources (i.e. structured tabular formats vs. unstructured text formats) and the incompleteness of the source data with regard to the target schema och keys. It is not far fetched to assume that e-commerce companies face similar challenges in their relation towards their suppliers.

2.2.3 The value proposition of structured product data

To conclude the review on data science progress in this field we’ll address the topic of value-creation to try and answer why working toward complete and structured data is important for e-commerce actors and “online-aggregators”

alike.

Considering the main objective of the research, that is, creating a structured product catalogue, Nguyen et al. (2011) says ”The product catalog is to online shopping what the Web index is to Web search” and elaborates by

”[...] structured data is fundamental to drive the user experience: it enables faceted search, comparison of products based on their specifications, and ranking of products based on their attributes.”. Thus Nguyen et al. (2011) regards structured data as an important enabler for a wide range of further applications. Petrovski and Bizer (2017) make a similar analysis and argues

”The central challenge for many tasks within the domain of e-commerce, including product matching, product categorization, faceted product search, and product recommendation, is extracting attribute-value pairs with high

(28)

precision from unstructured product descriptions or semi-structured product specifications.”. Ristoski et al. (2018) takes the perspective of the e- commerce customer and argues that as the aggregated online assortment of products has expanded it has become increasingly difficult for customers to find and compare products online. Investigating the cause of this experi- enced hardship on the part of the customer, Ristoski et al. (2018) find that the majority of products for sale online is presented only in terms of a title and description, meaning that unstructured product data dominate the online retail environment. Looking at e-commerce websites input feeds of product information, where target schemas for “online aggregators” product catalogues are clearly stated, the authors find that the data is often incomplete in comparison to the input schema - making the search performance of those products orders of magnitude less effective than products fulfilling the schema requirements. Staying in the customer perspective, Walther et al.

(2010) argues that structured product specification are the most valuable data for the online consumer as it creates a comprehensive understanding of the product and allows comparison with other similar products.

We have briefly addressed the underlying assumption of completeness in the data that is prerequisite for the success of aggregation systems of product data. To fully address the problems of the assumption we turn to Walther et al. (2010) who’s thesis is built on using supplier websites as source for raw data collection given the flawedness in e-commerce data. On e-commerce data they argue that “The information in individual online shops is restricted to only the sold products and often error prone and not comprehensive”

and drive the thesis that supplier data is in contrast “complete, correct and

(29)

up-to-date”. Along with Rao and Sashikuma (2016) identification of data incompleteness as a core obstacle in the journey towards building comprehensive product catalogues, we conclude that e-commerce website cannot be considered a reliable source of complete product information.

Lastly, the value of complete and structured data is evident in terms of machine learning applications. Having incomplete data generates substantially weaker classifiers in from ML-algorithms (Shimada and Endo, 2005) and structured data works better in creating strong ML-based systems than it’s unstructured counterpart.

2.3 Building a successful e-commerce busi- ness

Given the technological developments in recent decades, many businesses have had to rethink traditional ways of conducting commerce and adopt their business to emerging technologies. Online commerce has been one such example where, particularly brick-and-mortar retail businesses, have been forced to go online to stay competitive in a new market environment. Given these developments, transitioning business online and adapting them to the digital era has become a major research area. E-commerce in particular has been the target for much of this research to address the challenges companies face during this transition.

Transitioning brick-and mortar business online appears to be easy. However, constructing a profitable online based model including everything from prod-

(30)

uct presentation to fulfillment of logistical promises and after-sale activities is evidently a big challenge (Atchariyachanvanich et al., 2008). While the online supply of products and destinations where they can be purchased has grown tremendously in the past decade, E-commerce has not evolved at the same rate in quality and the possibility of setting up an online store without huge initial investments has driven many without domain knowledge to in- vest in this area (Varela et al., 2017). The strong trend of internet adoption on part of the consumer has forced companies online rapidly in order for them to stay relevant, but winning online takes more than presence and as the competition has grown stronger, the need for domain knowledge to create competitive advantages has become painfully evident for market participants.

2.3.1 Critical success factors in E-commerce

Varela et al. (2017) summarize the research on success factors for e-commerce companies and find that the mainstream of the studies identify five categories that need addressing to stay competitive online: technology acceptance factors, social factors, cognitive factors, ethical factors and environmental factors. Technology acceptance factors aside, the critical success factors relate to organizational challenges that emerge from the effort of transitioning a business from offline to online as well as behavioural challenges in getting the consumers to adapt to online purchasing. Breaking down the larger themes laid out by the categories, Varela et al. (2017) suggest twelve critical success factors that must be addressed for building a successful e-commerce website. These are presented in Figure 2.2.

While the success factors are often discussed in general terms in the littera-

(31)

Figure 2.2: Table of success factors from Varela et al. (2017)

(32)

ture without touching the topic of key enablers for the different dimensions of building a website some have touched upon the topic of data completeness and quality. Burgess and Karanasios (2008) and Cebi (2013) identify information quality as a main factor in building a competitive e-commerce business and Chaudhuri et al. (2019) argue that ”In e-commerce, content quality of the product catalog plays a key role in delivering a satisfactory experience to the customers”. The most widely discussed factor relates to website usability and Varela et al. (2017) discuss on-site navigation as a critical problem in terms of usability. Moreover, the aspect of trust has been discussed at length within this research area as it relates to both social and ethical success factors (Lee and Lin, 2005), (Machado, 2011). Trust is important in every aspect of e-commerce, from describing products objectively and honestly to practicing solid privacy policies (Ngai, 2003).

Other examples of research on the topic that has been done on a higher level of abstractions is provided by Choshin and Ghaffari (2017) who investigate important factors for small- and medium-sized companies in creating online businesses and finds statistical proof for customer satisfaction, cost, technological infrastructure and customer awareness and knowledge being integral factors for success. Furthermore, Nisar and Prabhakar (2017) find perceived value, customer expectations, perceived quality and loyalty to be important.

To summarize, the research done in the realm of business and management has accurately depicted the broad strokes of the many factors that are necessary to keep in mind when pursuing the e-commerce space. However, the field has yet to discuss the connection between these general factors and the underlying data that is needed to support many of them.

(33)

2.4 The different kinds of data affecting cus- tomer experience

So far in our discussion on e-commerce in general and the aspect of data in particular, we have focused on data in textual format. An important note is that selling products online demands complementing data for ensuring a good customer experience. Product pages today contain reviews, comments, images and videos along with the textual product data which all contribute to the customer experience. The impact of these different forms of product data on how a product is perceived online has been discussed individually.

Chaudhuri et al. (2019) discuss the impact of product images and argue that

”Images play a key role in influencing the quality of customer experience and the customers’ decision-making path in e-commerce transactions. Images provide detailed product information that helps the customer build confi- dence in the product quality and fulfillment promises.” and further argue that bad or incorrect images can have a significant negative impact on the customers willingness to purchase a product online.

Similar studies have been made on the impact of product reviews by for example Singh et al. (2017) and Wan et al. (2018). We want to highlight that complete and structured data goes beyond the realm of textual data and end on an important point made by Chaudhuri et al. (2019): “Human errors in compiling product information and limitations of software systems severely hinder the ability to provide a homogeneous content experience across categories to the customer.”

(34)

Chapter 3 Method

This chapter aims to give the reader an understanding of the methodology used, and methods applied, when conducting the study. On the highest level, a qualitative method using a case study approach was used in order to evaluate the first research question: what role does an online retailer place on structured product data? The findings from this analysis resulted in a set of six key propositions. These propositions were used as input for a validity analysis in the form of a multiple linear regression model, where data from the subject company was used in an attempt to validate each of the propositions.

3.1 Proposition analysis

The proposition analysis is structured as a single case study with embedded units as described by Baxter and Jack (2008). In this case, the embedded units are the subsidiaries of the Company, and the analysis will largely be

(35)

considered a cross-case analysis. The results of the interviews are analyzed and consolidated to a set of propositions, in this methodological context they can directly be related to the propositions in the case study framework put forward by Yin (2003). The design choice of linking data to propositions has been decided in order to create a solid foundation for the latter part of the study. The use of pattern matching Yin (2003) is deemed appropriate in order to determine patterns observed from individuals close to, or within, the data management teams at the Company and its subsidiaries. This would require interviews as the main data collection method, which will be discussed in greater detail below (Yin, 2003).

The proposition analysis encompassed 15 exploratory interviews with em- ployees and management at a large Nordic e-commerce company. The main purpose of this analysis was to gain insights into the role of data in e- commerce. This was done by identifying themes where the importance of data is prevalent, these themes then acted as input to the proposition validation analysis. The interviews were conducted in January and February of 2020.

3.1.1 Interviews

The guiding question of the role of data in e-commerce will be analysed through interviews using a qualitative lens as outlined by Creswell (2009).

The process can, in short, be described in the following steps:

1. Collecting raw data (transcripts, notes etc.) 2. Organizing and preparing data for analysis

(36)

3. Reading through the data 4. Coding the data

5. Identify themes and descriptions for themes 6. Interrelating themes/descriptions

7. Interpretation

8. Validating accuracy of information (through cross-validation)

The interviews serve the main purpose of acting as input data for the formu- lation of the propositions. The interviews were semi-structured in the sense that they related to the guiding theme, while allowing the interviewees the freedom to potentially add propositions of their own, which may or may not be included in an extended scope.

The interviews were conducted in ten separate sessions either in person or via video-conference. Interviewees were picked from multiple organizational levels and categorized by organizational functions are listed below:

• Management

– Chief Operating Officer

– Head of Business Development & Strategic Projects

• Merchandising

– Head of Merchandising – Merchandiser (x2)

• Product management

(37)

– Senior Category manager – Junior Category manager (x2)

• Online marketing

– Head of online marketing

– Online marketing specialist (x2)

• Business controlling – Controller (x2)

• Content & marketing – Content curator (x2)

3.2 Proposition validation

The following propositions (refer to section 4.2) were deemed appropriate for a quantitative analysis given the data available: Proposition 1, Proposition 2, Proposition 3, Proposition 4, Proposition 5. These propositions crucially relate to tangible response variables in the form of internal traffic, external traffic and quantity of orders. These response variables are described below.

The focus of the proposition validation lies in conducting a quantitative analysis of the propositions from section 4.1 in order to evaluate their legitimacy connected to actual sales and product data within the scope of the specific company. Note, again, that this single company is not to be used as a direct generalization, but is considered an adequate subject for the scope of the the

(38)

study as a whole.

• Quantity of orders, in the models denoted as ”quantity”, is defined as the number of orders placed from a single product page. This response facilitates the evaluation of propositions 1, 2 and 5, as we can evaluate the impact that our meta-attributes and images has directly on sales.

• External traffic, in the models denoted as ”sessions”, is defined as the number of times a user has started their session on the e-commerce website on a specific product page. That is, a session is only counted where the user enters the e-commerce website from an external link on e.g. a search engine. This response variable is thus suited to quantify the external traffic that a single product page generates. This response facilitates the evaluation of propositions 3 and 4, as we can evaluate the impact of our chosen meta-attributes and the product title on the external traffic that they generate.

• Internal traffic, in the models denoted as ”pageviews”, is defined as the number of times any user has visited a product page, but not started their session on that specific product page. This response variable is thus suited to quantify the internal traffic that a single product page generates. This response facilitates the evaluation of propositions 1 and 2, as we can evaluate the impact of our chosen meta-attributes on the internal traffic that they generate.

The only proposition left out of the quantitative analysis is thus Proposition 6. This proposition captures the value of structured data on business intel- ligence, and the benefits of exploiting such assets are not as direct as with

(39)

the former propositions.

3.2.1 Multiple linear regression

In order to evaluate how rich and structured data on product features is a driver of online sales and traffic, a multiple linear regression model is proposed as it is widely used for this kind of problem (see e.g. Ye, Law, Gu 2009).

This method of analysis allows us to not only evaluate whether there is a significant impact on sales, but also to control for differing product/retailer contexts in the analysis.

The full quantitative analysis will be made on the aforementioned response variables on the company subject to study. The analysis consists of different product categories which will be the main analysis in investigating the legitimacy of propositions 1-5, but will also strengthen our analysis towards a generalized conclusion. Data for the analysis will be made available to us by the company and will be drawn from internal ERP-systems, PIM-systems as well as from Google Analytics.

As mentioned, there are three response variables of interest. Each of the response variables are to be modelled individually:

1. Number of visits to the product page from external links 2. Number of visits to the product page from internal links 3. Quantity of orders on a product page

These variables were modelled using essentially the same predictors where the predictors were different measures of the data quality of the product page

(40)

in question. These measures included (but were not limited to): quality of product title, length and quality of product description, number and type of product attributes, number of high quality images and classification of the product. The construction of the model and the choice of predictors has been careful and deliberate, drawing from the interviews with industry professionals from the proposition analysis as well as the theoretical background in Chapter 2. Furthermore, a number of control variables that are well es- tablished to correlate with the responses were used in order to limit model variance.

In summary, the multiple linear regression model will not try to predict sales or traffic, since we are aware that the aspect of product data is only one theme among many that impact these variables. Instead, we want to investigates the aspect of product data as it relates to sales and online traffic to see 1) whether they have a significant role in predicting how well a product sells online and thus further validate the propositions, and 2) how big of an impact the different aspects have individually and in relation to one another.

3.2.2 Data

This section mainly aims to describe the quantitative data collected through the Company’s various databases, but will also give a brief discussion on the format of the interviews conducted.

For the data compiled from the Company’s internal databases, the chosen time span ranges over two years - from 2018-01-01 through 2019-12-31.

(41)

Product data

The product data set is compiled from multiple exports from the Company’s own PIM (product information management) system. The complete data set contains all the relevant information on the SKU (stock keeping unit) level of the product that is presented on the website. This is crucial for the analysis, as we can utilize the category groupings on multiple levels to infer different rules in the analysis.

A full list of parameters used in the models of analysis will be provided in the appropriate section.

Sales data

The sales data is collected from the Company’s ERP (Enterprise resource planning) system. In practice, the data describes the sales on SKU level, both in terms of total revenue and number of SKU sold.

Traffic data

The traffic data set is generated from the Google Analytics platform, and provides us with information on page hits, the customers’ journeys through the website and conversion rates on the level of web pages. I.e., we can utilize this data to track where the visitor entered the site, and how the journey towards a specific product is conducted in order to model the importance of certain data features.

(42)

Product attribute metadata

The product attribute metadata data set is a consolidated set on the product data, where we define units of analysis relevant to the propositions. Firstly, the key objective is to find measures indicating to what extent the products have structured product data. Our approach is to consolidate the data in different groupings, and count have many structured data points are present for the different products in the data set. Secondly, we want to measure other aspects of the data in one way or the other relating to the propositions. We try to find measures for the quality of the product titles and to what extent the underlying structured data have been leveraged in their creation and also seek measures for images and descriptions. The following set of metadata attributes have been carefully selected:

• Number of populated base attributes

– These attributes include, but are not limited to: product brand, method of delivery, country of origin and unit type

• Number of populated standard attributes

– These attributes include, but are not limited to: design series, material and model number

• Number of populated Dimensions attributes

• Number of populated category specific attributes

• Number of high-quality images on the product page

• The length of the product description (number of words)

(43)

• Whether or not the following attributes are present in the title:

– Design series – Colour – Brand – Material

• Number of available colour attributes

• Length of the product title when adjusted for automated title creation

3.2.3 Model specification

From the reasoning above, the following multiple linear regression equations are proposed for each category c:

log(pageviewsc,i) = βc,0+ β^T_cxc,i+ i (3.1) log(sessionsc,i) = βc,0+ β^T_cxc,i+ i (3.2) log(quantityc,i) = βc,0+ ˜β^T_cx˜c,i+ i (3.3)

For the models specified in equations 3.1 and 3.2, the vector of regressors is

(44)

defined as:

x_c,i =







base_c,i standard_c,i dimensions_c,i imagecount_c,i categorySpecif ic_c,i shortDescription_c,i longDescription_c,i

intitleSeries_c,i intitleM aterial_c,i

colour_c,i log(averageP rice_c,i)

adjustedT itle_c,i

log(averageP rice_c,i) × longDesc_c,i log(averageP rice_c,i) × imagecount_c,i log(averageP rice_c,i) × adjustedtitle_c,i







(3.4)

where the vector β_c is then simply the corresponding coefficient vector for the regressor vector x_c,i. For equation 3.3, the vector of regressors ˜x_c,i is identical to x_c,i, but appended with an interaction term with the delivery time delivery_c,i, pageviews_c,i as well as an interaction term between delivery and log(averageP rice).

(45)

3.2.4 Validity of assumptions

Homoscedasticity

One key assumption of the multiple linear regression model is the homoscedasticity assumption – that the error terms of the regression have a constant variance across the sample. To ensure that the model yielded no heteroskedastic error terms, quantile-quantile plots were evaluated for each model. Figure 3.1 illustrates an example for the bath category. To ensure homoscedasticity, the empirical and theoretical quantiles should match as closely as possible, as shown in the figure.

In order to achieve homoscedastic error terms, however, the response variables had to be log-transformed in all cases. This is a common transformation technique used for this type of problem.

Multicollinearity

While the existence of multicollinearity in the model is only a violation of the model assumptions in the case of perfect multicollinearity, high levels can still cause some issues. A common approach to detect potential multicollinearity in the model is to utilize the variance inflation factor (vif). Each of the models run were checked using vif, resulting in no highly correlated regressors – with the exception of the interaction terms, which should be expected.

Omitted variable bias

One crucial point in the estimations of the regression models is the issue of omitted variable bias. For a model to be biased through omitted variables,

(46)

Figure 3.1: Example QQ plot for the pageviews model of category bath two conditions must hold:

x_i is correlated with the omitted variable x_o for some i x_o is a determinant of the response variable y

In the construction of the models, significant care was taken in order to reduce the risk of bias from omitted variables. Since the models are not aimed to be predictive by construction, this issue is largely simplified.

(47)

3.3 Research ethics

The study was conducted with great regard to current research ethical con- siderations. Specifically, the study utilized the four principles for ethical research proposed by the Swedish Research Council (Vetenskapsr˚adet, 2002).

These four principles, or criteria, are presented below and discussed in relation to the study.

The criterion of information states that the researcher shall inform the people included in the study about its aim. Specifically, the researcher shall inform them about their role in the study, that participation is optional and the terms which are at play. In order to accommodate for this set of rules, all interviewees were asked whether or not they wanted to participate in the study, leaving full disclosure of the terms of personal anonymity. The interviewees were also informed of the aim of the study either via e-mail, a workplace instant messaging application or verbally. All interviewees com- plied in full.

The criterion of consent states that any participant in a study has the right to control their own contribution. That is, the researcher shall collect the participant’s consent (and possibly the consent of a legal guardian). Fur- thermore, the participant has the right to independently decide the terms of their involvement and be able to abort their involvement without any negative consequences. Finally, the participant shall not be the subject of any undue pressure. As stated previously, consent was collected from every in- terviewee in the study, and they were informed that they should only convey information that they deem appropriate for sharing. Furthermore, as the

(48)

interviews were recorded, consent was asked for (and approved) before the start of each interview.

The criterion of confidentiality concerns information about the research participants. Any information on the participants shall be given as much confidentiality as possible, and any personal data shall be stored so that none other than the researcher has access to them. During the interview process, no personal data was stored in the transcripts except the first name and function of the participant. The first name was collected in order to facilitate discussions between the authors. When presenting findings, the interviewees were simply referred to by their function at the Company. While some of the employee functions only employ a few people, leaving the Company anonymous throughout the thesis aids in keeping confidentiality.

The criterion of good use states that any information collected on single participants shall only be used for the purpose of research. In the study, no data was passed on from the researchers to any function of the Company apart from the finished thesis. This means that the interview transcripts were only seen by the authors, and any information relating directly to a participant was thus ensured not to be used for other purposes.

(49)

Chapter 4 Results

4.1 Proposition analysis

The overall aim of the proposition analysis was to explore the topic of product data, its application and potential, in e-commerce with an open mind.

In the pursuit of achieving an understanding as complete as possible we in- terviewed people in most parts of the organization and let them explain their thoughts and daily struggles relating to product data. A high-level take away that became evident from our sessions was that the value ascribed to data differed significantly between people from different organizational functions which we will explore further below. In terms of structure, we present our findings under six headlines representing the most common themes discussed in the interviews. Moreover, all of the interviewees were in agreement when discussing the value of product images with the message that images are integral for successfully selling products online. As such, the findings below

(50)

refer to textual product data.

4.1.1 Structured vs. unstructured product data

It is evident that product data cannot be discussed without making the initial distinctions between unstructured and structured data. The terms are assigned by the authors with inspiration from data science literature but were referred to in the interviews as “tabular data” instead of structured, and “free text data” instead of unstructured. The consensus from all parts of the organization was that structured data is preferable given the many applications in the e-commerce value-chain. However, there is a significant trade-off between working towards structured data formatting and the cost of pursuing that structured data (in terms of time, effort and quality).

The teams working with assortment onboarding, including category managers with the responsibility for supplier relations, pricing and marketing within categories and merchandisers with responsibility for data curation, both stressed the value of structured data, and the onboarding process has been tailored to achieve it by the best means available. When onboarding new assortment, the suppliers must structure their data according to a template defined by the category manager. The template represents a “blueprint” or a “schema” for what data is necessary depending on which product category it belongs to. The main purpose of a pre-defined schema is that it ensures that products in the same category are presented in a consistent way, allowing the customer to compare products across suppliers. A consistent set of structured data within a category also allows for sitelist filtering, for example on color or width, to allow the customer better on-site navigation in large

(51)

assortments.

The argument against working towards achieving structured product data is that it consumes a lot of time. Suppliers are seldom capable of quickly packaging their data to a pre-defined format. Instead, each supplier has their own blueprint for how they store their data in different categories. This forces suppliers to, often through manual effort, re-structure their data to fit the mandated format, a process that often takes significant amounts of time.

When the data arrives to merchandising, it is re-packaged and enriched to ensure optimal site-presentation and compliance with the existing assortments packaging. Moreover, many suppliers lack parts of the mandated data internally which creates a difficult situation, the supplier can be pressured into

“creating” the mandated data, but more often than not the suppliers lack the willingness to do so, forcing the onboarding team to regularly make ex- ceptions with regards to the blueprint. While the process generally achieves the desired result of consistency, it is painfully manual for everyone involved, has significant lead times and is prone to errors.

Interviewees from functions not involved in the process of assortment onboarding were in agreement over the necessity of structured data for multiple reasons. Considering an assortment of products with unstructured data, the possibilities for automated applications decrease significantly. Optimiz- ing on-site navigation through filtering functionality was considered to be near impossible, and the ability to understand the in-house assortment in terms of white-spots and weak-spots would only be possible in terms of the structured data available (namely the product categorization). Furthermore, the ability for search engine optimization of the assortment would be very

(52)

limited without significant manual effort.

The most important finding from our discussions on structured versus unstructured data was that all organizational functions are in agreement on the necessity of structuring product data but from many different angles. Most interviewees mentioned the obvious application in filtering functionality, but other perspectives and levers of structured data were only raised by specific organizational functions indicating that even though the value is appreciated by everyone, there is a knowledge gap between internal functions in their understanding of how product data is leveraged throughout the organization.

Going forward, we discuss our findings relating to the current and potential applications of structured product data, that is, taking the perspective of an e-commerce business where the data is perfectly structured and complete.

4.1.2 Structured data in online marketing

Results from this section are derived from interviews with two online marketing experts within the company.

Online marketing encompasses several channels and methods but the overwhelming majority of online traffic arriving at the e-commerce website from marketing efforts enter either from search engines such as Google or from social media platforms such as Facebook. Social media marketing was only discussed briefly since it was not the interviewees’ day-to-day responsibilities, but search engine optimization was discussed at length and particularly how structured data can be leveraged for ranking higher on the organic search results for the company’s target keywords and categories.

(53)

Those familiar with SEO (Search Engine Optimization) recognized that the overarching target in the e-commerce context is to get one’s website listed as high up in the search results as possible in searches using specific keywords that are related to one’s products. How the underlying ranking algorithms used by the search engines work is proprietary, there are however some intu- itive basics that experts in the field agree are the most important for making a website rise in the search engine rankings and two of the three directly relate to product data.

The first method for achieving a good search engine ranking relates to keywords used in search queries. Words that relate to products in different categories are referred to as keywords, and the main concept here is that content on the e-commerce website should include the same keywords that potential customers might use when searching for products in relevant categories.

Consider the scenario where a potential customer enters a search engine with the intention of finding a suitable sofa, that customer will likely use keywords such as e.g. ‘sofa’, ‘couch’, ‘settee’ or ‘divan’. For the e-commerce website selling sofas, it is important that those keywords are present in the website content to indicate to the search engine that this is a relevant website for a consumer searching for sofas.

The second method relates to content relevance. The idea is that a website yielded by a specific search query or keyword should have content directly related to that query or keyword. The more specific results the better. Con- tinuing with the same example, a result that links directly to a landing page containing an assortment of sofas will rank higher than a result that links to a homepage for a website selling a variety of furniture. The relevance is

(54)

measured by customers’ tendency to stay on a website after entering from a search engine and also how many clicks a customer must use to navigate to achieve a desired result.

The last method has to do with linking to a landing page, this method is somewhat more technical and is excluded from this result as it does not relate to product data.

Search queries can be categorized into general, specific and long-tail depending on their level of specificity as demonstrated in Figure 4.1. General queries have the highest competition and is as such the hardest to achieve good rankings for. Just imagine how many websites would like to be the preferred results for queries such as ‘nice clothing’, ‘cheap furniture’ or ‘buy laptop’.

These queries are generally used by individuals wanting to explore assortments and options and as such relates to broad categories of products. Rel- evant results for these queries are often e-commerce homepages or category landing pages. Given the fierce competition and the fact that the number of pages at each e-commerce website that are relevant for general queries are generally very few, the content on these pages is curated manually by SEO experts.

However, with increasing specificity in search queries, the number of landing pages in need of content curation and optimization increases exponentially, and with the increase in number of landing pages follows an ever growing burden in manually managing the content on thousands or even millions of landing pages. In this context, structured product data can play an integral role for success in the online marketing space.

(55)

Figure 4.1: Example of different types of searches

Keeping in mind the important concepts of keywords and relevance, take the example of an e-commerce website offering a large assortment of furniture and the specific search query ‘green sofa’. The website in question likely has several other product categories besides sofas, including tables, chairs, beds and storage furniture and all of these categories likely contain products of different colors. Furthermore, all of the mentioned categories likely have one, two or even three levels of subcategories resulting in hundreds of cumulative categories on a single website. To present the most relevant results relating to the query ‘green sofa’ the website would naturally want to refer to a landing page containing all of the website’s green sofas (and no products that are not both green and sofas to maximize relevance) and would further want that landing pages’ content to include the keywords ‘green’ and ‘sofa’.

Here, the first use case of how structured data is a core prerequisite for

(56)

online marketing becomes evident: The only way one can easily, scalably and without manual effort create a landing page containing all of the website’s green sofas is if all sofas have a structured attribute where the key refers to color and the associated value is green. Effectively using a category along with an attribute filter for that same category. While this can be done manually, but with 15-20 different colors and hundreds of categories, the landing pages for the set of relatively simple search queries containing a product type and a color is counted in the thousands. To make matters worse, color is only one key, or attribute, relevant for the assortment. Customers could use simple queries such as ‘leather sofa’ or ‘vintage sofa’ relating to the keys material and style respectively implicating the addition of thousands of more necessary landing pages to maximize search engine relevance. The manual effort in creating this volume of pages and content is overwhelming, calling for automated solutions. With a complete set of structured product data, these pages and the related keywords can be created automatically by combining categories and keys using simple algorithms without need for manual efforts.

So far, specific search queries have been considered as they relate to landing pages and concluded that thousands of landing pages are necessary for relevance optimization in the SEO-context. Intuitively, thousands could be exchanged for several millions depending on the size of the assortment and the level of detail as well as the number of dimensions in the structured data.

Using the same example of a website selling furniture, we instead consider the example of a long-tail query, namely ‘green velvet chesterfield sofa’. De- pending on the depth of the assortment, the website could have none, one

The (underestimated) role of product data for winning online retail

The (underestimated) role of product data for winning online retail

JOHN BOLMGREN

HENRIK LINDSTRÖM

The (underestimated) role of product data for winning online retail

by

John Bolmgren Henrik Lindström

Den (underskattade) rollen av produktdata för att vinna e-handeln

av

John Bolmgren Henrik Lindström

Contents

List of Figures

List of Tables

Chapter 1

Introduction

1.1 Scope and delimitations of the paper

1.2 Setting the stage for discussing e-commerce data

1.3 Research questions

Chapter 2

Theoretical background

2.1 The current state of academic e-commerce literature

2.2 E-commerce from the perspective of Data Science

2.2.1 Product data come in many shapes

2.2.2 Towards the mighty concept of a structured prod- uct catalogue

2.2.3 The value proposition of structured product data

2.3 Building a successful e-commerce busi- ness

2.3.1 Critical success factors in E-commerce

2.4 The different kinds of data affecting cus- tomer experience

Chapter 3

Method

3.1 Proposition analysis

3.1.1 Interviews

3.2 Proposition validation

3.2.1 Multiple linear regression

3.2.2 Data

3.2.3 Model specification

3.2.4 Validity of assumptions

3.3 Research ethics

Chapter 4

Results

4.1 Proposition analysis

4.1.1 Structured vs. unstructured product data

4.1.2 Structured data in online marketing