A Large-scale Analysis of the Marketplace Characteristics in Fiverr
Suman Kalyan Maity, Chandra Bhanu Jha, Avinash Kumar, Ayan Sengupta, Madhur Modi, Animesh Mukherjee
AA Large-scale Analysis of the Marketplace Characteristics in Fiverr †Suman Kalyan Maity, ‡Chandra Bhanu Jha, ‡Avinash Kumar, ‡Ayan Sengupta, ‡Madhur Modi and †Animesh Mukherjee†Dept. of CSE, IIT Kharagpur, India - 721302 ‡VGSOM, IIT Kharagpur, India - 721302Email: † { sumankalyan.maity,animeshm } @cse.iitkgp.ernet.in; ‡ { cbj,avinashk2017,ayan.sengupta007,madhurmd } @iitkgp.ac.in Abstract
Crowdsourcing platforms have become quite popu-lar due to the increasing demand of human computation-based tasks. Though the crowdsourcing systems are pri-marily demand-driven like MTurk, supply-driven mar-ketplaces are becoming increasingly popular. Fiverr is afast growing supply-driven marketplace where the sell-ers post micro-tasks (gigs) and users purchase them forprices as low as $
5. In this paper, we study the Fiverrplatform as a unique marketplace and characterize thesellers, buyers and the interactions among them. We findthat sellers are more appeasing in their interactions andtry to woo their buyers into buying their gigs. Thereare many small tightly-knit communities existing in theseller-seller network who support each other. We alsostudy Fiverr as a seller-driven marketplace in terms ofsales, churn rates, competitiveness among various sub-categories etc. and observe that while there are certainsimilarities with common marketplaces there are alsomany differences.
1. Introduction
In recent years, there has been a huge growthin crowdsourcing platforms due to rising demands ofcrowdsourcing micro-tasks like doing surveys, prepar-ing resumes and other similar tasks that require man-ual labor. As the popularity of these crowd-basedmicrotasks increased, various crowdsourcing platformsemerged like the Amazon Mechanical Turk, CrowdFlower, Microworker, ShortTask etc. These platformsare primarily demand-driven. Another range of plat-forms drift from this demand-driven based approach tosupply-driven approach where the workers try to show-case their skills and talents by offering various micro-tasks. These marketplaces are primarily used by thefreelancers. Upwork, Elance, Fiverr, oDesk, Freelancer,TaskRabbit are some of the very prominent supply-driven marketplaces where the suppliers post advertise-ments on services like creative writing, programming,software development, graphics etc. with hourly ratesor a fixed fee.Since 2010, Fiverr has grown into one of the largestand popular online freelancing marketplace. As of 11 th June, 2016, most of the Fiverr traffic (17.9%)comes from India followed by US (17.6%) and Pakistan(12.1%) . Fiverr facilitates the buying and selling of“gigs” (micro-tasks) online starting from as low as $5per gig. The sellers in Fiverr, apart from selling, canalso buy gigs from other sellers; similarly, buyers canalso become sellers whenever necessary. This dual roleplayed by the members of this community and the in-teraction among them has made Fiverr a unique market-place.
2. Related works
There have been several studies on crowdsourc-ing based online labor marketplaces, predominantly onAmazon Mechanical Turk (AMT). Ross et al. [1] pre-sented a user demography study on AMT. Berinsky etal. [2] described issues related to recruiting subjects onAMT platform. Paolacci et al. [3] reviewed the strengthsof MT in recruiting participants. Kosinski et al. [4] mea-sured crowd intelligence. Mason and Suri [5] performedbehavioral analysis of Turkers and Suri et al. [6] stud-ied their honesty. Halpin and Blanco [7] used machinelearning to identify spammers in AMT. Allahbakhsh etal. [8] discussed quality control schemes while Hey-mann and Garcia-Molina [9] proposed a novel analyt-ics tool for crowdsourcing systems. Apart from AMT,Wang et al. [10] performed analysis of the tasks on ZBJand SDH (Chinese crowdsourcing sites) and estimatedthat 90% of all tasks were crowdturfing tasks.
Supply-driven marketplace:
Though there havebeen several studies on demand-driven marketplaces,supply-driven marketplaces, which are also a grow-ing business in crowdsourcing based marketplaces, areunder-studied. [11] performed measurement studies onFiverr focusing on the quality of the gigs. [12] analyzedabusive tasks on Freelancer. Several black-hat mar-ketplaces like HackBB, SilkRoad, Agora, SEOClerks,MyCheapJobs, Gigbucks, Gigton, TenBux have alsoemerged which facilitate sale of fraudulent products andillicit goods [13, 14, 15, 16]. [16] studied crowdturfingtask detection on Fiverr. [17] analyze marketplace char-acteristic in SEOClerks and MyCheapJobs. [18] studied a r X i v : . [ c s . S I] S e p he reward effect on submissions in Taskcn. Our studyis different from the above in the following ways - wepresent a first comprehensive picture of the behavior ofthe buyers, the sellers and the seller-seller interactions and demonstrate how all these together impact Fiverr asa marketplace from the economic , the sociological andthe strategic perspective.
3. Dataset description
We crawled Fiverr data using preferential crawlersin R . The crawler ran for a period of 20 days and datafor 59,786 gigs from 103 subcategories belonging to 10categories were collected. For each gig, we collectedprice, reviews, seller’s response to reviews, and mutualratings. We then separately crawled the member profileinformation (average response time, level of seller, userlocations and conversant languages etc.). Next we re-moved the gigs which have incomplete information forsome fields. Consequently, we were left with 41,473gigs which we use for our study. The dataset contains21,767 unique sellers and 5,31,841 unique buyers. Thetotal reviews reached a count of 34,43,381 and are usedas proxies for business transaction between the reviewer(also buyer) and the gig owner. The prices of the gigsrange from $5 to $995 indicating the diversity in the ser-vice quality offered. Gigs, Sellers and Buyers:
Gigs are services providedby sellers in Fiverr and are grouped into categories andfurther subcategories. The
Graphics Design categoryhas the highest number of sellers with 32% of the totalshare. ‘Level of seller’ is assigned by the website to sell-ers based on the total time spent, the volume of sales, theratings, the cancellation rates and the order count thresh-olds amongst many other things. The categories here are new , level 1 , level 2 and top rated .Table 1. Categories based on no. of gigs and buyers. CategoryGroups
Group 1
Group 2
Group 3
Group 4
Group 5
Total 21767 531841 41573 $8.69 78.22 $757.17
Categories and product differentiation:
We organizeall the categories in Fiverr into 5 broad groups – Graph-ics & Design, Digital Marketing, Writing & Translation;Video & Animation, Music & Audio; Programming &Tech; Advertising , Business; Lifestyle, Gifts, Fun &Bizarre – and analyze them separately to find whichamong them are more popular in terms of the number ofgigs, high supply and demand. We also calculate aver-age gig price in each category, average sales of the gigsand revenue generated by the gigs (see Table 1). We findthat the average price is maximum for the creative cate-gories i.e., group 1 ($31.02) whereas the average price in the lifestyle related category (group 5) is just $5. The av-erage sales per gig varies from 471.45 (Audio and Videocategory) to 24.41 (group 5). We also observe that manycategories have significant sales in terms of total salesbut show mediocrity in terms of average sales per gig.Therefore, the product differentiation is quite apparentin this marketplace.
4. Fiverr under the economic lens
From the economic viewpoint, we analyze behaviorof various marketplace entities like buyers, sellers andthe review patterns.
We find that more than 40% sellers have only earned$5-$100 over the last nine months whereas the propor-tion of sellers with revenue > $50 ,
000 is only about0.4% (see fig 1). Fiverr provides a lot of informationabout sellers and gigs e.g., average response time, or-ders in queue, number of reviews, number of positivereviews, average response time of the seller and manyrecently added new features like - a rating based on theservice as described on a scale of 0–5, a rating based onthe recommendation of the gig by a user. We investigateseveral of these features in order to unfold the behaviorof the sellers. Based on different attributes, Fiverr clas-Figure 1. Revenue generation of the sellers.sifies the sellers into three levels - level 1 , level 2 and top rated . There are many sellers who have not yet metthe criterion to be in any of these levels and are called new sellers by Fiverr. We observe that about 22.56%of the whole seller community are in level 1 , more than62.89% are level 2 sellers whereas only 4.8% are toprated sellers and the rest of them are new . Table 2 showshow sales and revenue varies for the different levels ofsellers. Apart from the fact that the top rated sellers areselling more number of products, it is also quite clearthat they charge more for their product than others. A top rated seller charges $9.56 on average, where a level2 seller charges $8.96. This premium charging showstheir bargaining power over others.Table 2. Performance of sellers. Level Avg. Sale Avg. Revenue Avg. Gigs
Level 1
Level 2
Top rated ales and top sellers
First we observe the performanceof sellers based on the revenue earned in the last ninemonths. The revenue generated by the seller varies from$555,715 to only $5. If we consider the number of unitsales without considering the price, then it can be as highas 38,302 to as low as 1. From the fig 2, we can clearlyobserve that most of the top sellers prefer selling a smallnumber of gigs in few subcategories rather than provid-ing diverse type of gigs. Hence, they are mostly profi-cient in certain type of services which indicates a goodlevel of professionalism in a service driven marketplace.Figure 2. Sales of top sellers. The sizes of bubblesdenotes the amount of total sales. The color coding isbased on the number of categories they provide servicesin.
Pricing of gigs
In Fiverr, 87.73% of the gigs arepriced at $5. 0.02% of the gigs are highly priced andmore than $500. From the dataset, we observe thatmost of the highly priced gigs are from the subcategories
Buy Photos Online Photoshop , Banner Ads and
ContentMarketing . Figure 3 shows how different levels of sell-ers price their gigs. The underlying distribution withineach seller level appears to be the same with the onlydifference being the percentage of total sellers in level 2is much high compared to other levels.Figure 3. Price of gigs of different sellers in each level.
Buyer ratings and seller response times
Buyersprovide ratings to the sellers for each purchase on a scaleof 0-5. Hence, average rating for each seller can be a pa-rameter for judging his/her performance. After analyz-ing the data, we believe that the rating system is not verydiscriminatory for Fiverr. More than 60% of the sellershave an average rating of 5 and more than 96% havemore than 4.7. This distribution is also valid categorywise (see fig 4(a)). In fig 4(b), we show the distributionof average response time of sellers. The graph showspower-law behavior. More than 95% of the sellers usu- ally respond within a day and ∼
35% of them respondwithin an hour. (a) (b)
Figure 4. (a) Average buyer rating of sellers for differentcategories (b) Average response time of seller.
Order queue
Another new feature that Fiverr has re-cently added is the order queue count which tells ushow many incomplete orders are present for a partic-ular seller for a given gig. Low value of order queuecount can be either because of less popularity or excel-lent customer handling. Similarly, high value of orderqueue count can indicate less proficiency in customerhandling or large number of orders inflow. In fig 5,we show the distribution of no. of orders in queue foreach level of sellers. Most of the level 2 sellers are veryprompt, however, a non-negligible percentage of themalso delay their services.
Gain of new customers
The customer base of sell-ers keeps changing over time and the inflow of new cus-tomers can be taken as a proxy for growing popularityand hence an indicator of good services. We propose ametric to measure this inflow. For a particular seller, wedefine
Gain ( t ) = | customer t ∩ customer t − || customer t − | where, customer t is the set of all customers of the sellerat time t and customer t is the set complement of the set customer t . customer t − represents the set of customersone day before the t th day. Fig 6 shows the average gainof top sellers of five subcategories. We observe that av-erage gain decreases over time. Globally, the percentagegain of customers for top sellers is around 40% and isgreater than the preceding time intervals for one month.Some subcategories show different trends. For example- in Custom Handmade Jewelry the largest gain ( > Extremely Bizarre showed the maximumgain of new customers which is around 75% (fig 6).
Buyers are the driving force of any marketplace. Inour dataset, we have a total of 5,31,841 buyers. Never-theless, if we only want to observe the profitable or re-igure 5. Distributionof order queue count ofsellers. Figure 6. Gain of newcustomers by sellers ineach subcategory.turning customers, then the number would be much lessthan this. ∼
45% of the buyers are one-time (only onepurchase across all gigs) buyers. Only 18% of the buyershave purchased more than 5 times in last nine months,with the maximum being 2939. Fig 7(a) shows the dis-tribution of number of purchases of gigs indicating apower-law behavior. In fig 7(b), we present the distribu-tion of purchases from different subcategories. 68.62%of the buyers buy from one subcategory and only ninebuyers have bought from more than 30 subcategories. (a) (b)
Figure 7. Distribution of (a) no. of gigs (b) no. ofsubcategories from which the buyers buy gigs.
Purchase and top buyers
As we have seen that mostof the buyers are just one time buyers, we consider onlythe top buyers in our studies, whose total purchases aremore than 200. In fig 8 we show that the amount of pur-Figure 8. Purchase plot of buyers. Size indicates thenumber of gigs a buyer has bought. Color coding isbased on the number of subcategories the buyer is buy-ing from.chases is not always directly proportionate to the volumeof purchases. Moreover, most of the top buyers preferto buy small number of gigs from a few subcategories, rather than buying different types of products across dif-ferent subcategories. This also shows that these buyersare not just stray visitors to the website; instead they re-peatedly buy the same products over time. Hence, it islegitimate to assume that loyal customers (although lessin number) do exist in supply-driven marketplaces likeFiverr.
Returning time of buyers
As we have seen the ex-istence of loyal and repeated buyers in Fiverr, the nextquestion that arises is - how often does a returning cus-tomer buy a particular gig? To measure this, we definea statistic for each of the repeated buyers for each of thegigs they are buying. returning time = ∑ Ni = | T i − T i − | N − where, { T i } Ni = is the time periods of purchase of a par-ticular gig. In other words, this statistic measures the av-erage day difference between any two consecutive pur-chases for each of the repeated buyers. In fig 9 we showthe distribution of the returning time of the top buyersin Fiverr. It clearly shows the peak at time zero and al-most uniform frequencies up to 3rd day followed by adecrease. We can therefore infer that most of the fre-quent buyers usually return within a day , which is an-other unlikely event for other e-commerce and productdriven marketplaces. This observation can greatly helpthe sellers to incentivise returning customers. This alsoshows the extent of heavy usage of services by the fre-quent customers which is overall an encouraging signfor Fiverr.Figure 9. Distribution of returning time of buyers. Reviews have increasingly become one of the guid-ing foundations for buyers in online marketplaces. Agood review system allows for a discriminative analy-sis among the different choices and helps the consumerdetermine the usefulness and ingenuity of the productunder question in the different stages of the purchase de-cision making process. Fiverr’s review system, thoughprobably intended to be means for insights into gig’sperformance, does seem to be quite different from theusual e-commerce platforms. The buyer and the sellercan engage in a mutual review exchange and rate eachother based on their interaction. Initial analysis showsthat the linguistic structure has evolved along a very con-strained framework which involves a high degree of mu-ual appreciation, well-wishing and goodwill exchange.We perform general analysis of gig reviews and try tolook at the commenting behavior of the users across allthe gigs. Though we consider comments on a gig as aproxy for purchase, we study the repetitiveness of com-ments in Fiverr. The normalized figures for 5,31,841users show us that ∼
45% of the users comment onlyonce. This is followed by 18% of the users placing thesame comment twice. ∼ .
9% of the users have com-mented the same thing thrice. This distribution followsa power-law behavior (see fig 10(a)). We also observe (a) (b)
Figure 10. Distribution of (a) number of times the samecomment is used by different users (b) number of timesthe same comment is used in different gigs.how comments are being repeated across all the gigs.Figure 10(b) shows the distribution of the comment rep-etitions which shows a power-law curve with heavy tail.80% of the comments are non-repetitive, 10.98% haveoccurred twice. Notably, the single comment “Outstand-ing Experience!” has appeared 51,451 times across allcomments. Close variations of this comment are alsoobserved frequently. This is followed by 12,406 repeti-tions of the comment “thanks”.
Sentiment analysis:
We perform sentiment analysis ofthe reviews using the standard sentiment dictionary [19].Most of the reviews have positive and encouraging con-structs. This can be attributed to the structure which hasdeveloped over time and is peculiar to Fiverr. “Outstand-ing experience”, “Thanks” and “Great to work with” aresome of the most common phrases encountered and thisis reflected in the skewed sentiment scores. However,a difference in the polarity distribution can be observedacross categories with different inherent characteristics.The difference is more pronounced across different priceranges. Sellers are found to be more positive in their re-sponses to the comments by the buyers. A cumulativeplot of Creative Logo Design subcategory demonstratesthis fact (see fig 11(a)). Comparisons across differentsubcategories (see fig 11(b)) also gives very convincinginsights. For trivial services like the ones contained inExtremely Bizarre, the reviews have higher proportionof positive sentiments than more serious services likethe ones in
Wordpress and
Creative Logo Design . Comparison with Amazon’s reviews:
Since reviewing (a) (b)
Figure 11. (a) Comparison of sentiment of buyer andseller text exchange for the Creative Logo Design sub-category (b) Review sentiment comparison across threecategories.patterns are a key representative of any marketplace, weuse these to differentiate a supply-driven and a demand-driven marketplaces. We collected Amazon’s reviewdata for the
Electronics category and randomly sampledthe reviews. These are in high contrast to that of Fiverr.The sentiment distribution is more inclusive of negativepolarities and the linguistic structure looks more realis-tic upon primary inspection. Comparison of top 10 mostfrequent words (see table 3) in reviews of Amazon
Elec-tronics products and Fiverr’s most sold
Creative LogoDesign shows that Amazon’s products get feedback re-lated to product specifics with some words expressinggratefulness while Fiverr is dominated by mostly emo-tional responses. The average number of characters perreview for Amazon is around 381, while for Fiverr, it ismerely around 61. This can be attributed to the elabo-rate and explanatory nature of the former and a genericcongratulatory nature of the later.Table 3. Top 10 most frequent words in reviews
Marketplace Top 10 words
Amazon case phone one great works like just well good useFiverr experience great work outstanding thanks logo jobwill good excellent
5. Sociological/network perspective ofFiverr
We create two types of seller-seller networks (Tier-Iand Tier-II) and study the interesting properties of thesenetworks.
We create a Tier-I seller-seller network G = ( V , E ) asfollows. The set of nodes V = { s , s , ... } are the set ofsellers in Fiverr and if a seller s buys k times from otherseller s , then we add a directed edge from s to s in G with weight (proportional to no. of purchases) weight ( s , s ) = k / total sales ( s ) Graph density:
The network formed using the abovemethod has 49.13% of the full sellers community butwith a graph density of only 0.001. This low number in-icates that, although a number of sellers together formthe network, they are loosely connected to each other.Thus they usually have a very small set of sellers in theirneighborhood from whom they make a purchase.
Degree distribution and power law properties:
It isa known fact that degree distribution in most of the realworld networks follows a power law p ( k ) ≈ k − α withthe exponent α lying between 2 and 3. The Fiverr seller-seller network also exhibits the same characteristics with α value of 2.27 for in-degree and 2.41 for out-degree(see fig 12). We further observe that percentage of nodeswith zero in-degree is 43.53% whereas percentage ofnodes with zero out-degree is 29.45%. This shows thatthe number of sellers who are only buying from othersellers is much higher compared to those who are onlyselling. Transitivity:
Transitivity, also referred to as the clus-Figure 12. Degree distribution of the network.tering coefficient, is the measure of how well the neigh-bors of a node are connected among themselves. Thisalso indicates the existence of cliques or dense mod-ules within a network. The Tier-I seller-seller networkof Fiverr has very low clustering coefficient of 0.037(the direction of the edges are ignored in this computa-tion). The low clustering coefficient shows that there isno tightly connected module existing between the sellersin the Fiverr market which is in contrast to the generalmarket where it is usually known to exist.
Connected components and communities:
The net-work consists of 526 weakly connected components and11,077 strongly connected components. The number ofstrongly connected components is almost same as thenumber of nodes which show low reciprocity in the net-work. By reciprocity of a network G ( V , E ) , we under-stand the proportion of occurrences where both ( v , v ) and ( v , v ) belongs to the edge set E . Using Louvain al-gorithm [20], we determine the community structure ofthis network. 625 communities are discovered with anoverall high modularity value of 0.941. Therefore, thenetwork shows many small communities and the com-munities themselves are tightly-knit. Geography of sellers:
We collected the geographicallocation based data of the sellers associated with theTier-I seller-seller network. We find that 31.57% of the whole community belongs to United States, followed byPakistan (9.27%) and India (7.73%). Moreover, the or-der remains same for individual sellers and buyers in thecommunity, where United States is leading followed byIndia and Pakistan. We also observe that in 26.15% ofthe cases both the buyer and the seller belong to the samecountry.
Characterizing the Tier-I seller-seller network:
Weobserve that in 15.53% cases a seller buys from anotherseller who sells gigs in the same category itself. Now, ifa buyer, who himself sells gigs in the same subcategorybuys something similar to his product, then the expla-nations behind this could be either promoting the othersellers’ business or an anomalous behavior. There can beseveral other reasons like - supporting someone’s busi-ness, who is personally very close, or, buying similarproduct from some expert and again selling to someoneelse at a higher price. In 36.84% of the cases, we findthat buyer’s total sales is more than the seller. In 7.81%cases they offer gigs in the same category. These casesmay be those where a high income seller supports otherless popular sellers by promoting them or, by just pro-viding good reviews for their gigs. If we consider thelevel wise distribution, then higher level sellers buy gigsfrom lower level sellers in 24.25% cases, whereas, lowerlevel sellers buy gigs from higher level sellers in 15.09%cases. The remaining 60.66% cases are purchases withinthe same level.
Here, we test the hypothesis whetherthe average revenue difference between a seller and abuyer in a community decreases as the reciprocity in-creases. We calculate reciprocity for each of the com-munities of the network and then calculate the aver-age difference of revenue for each of the communi-ties. So, if a community c i of the network contains { ( s , b ) , ( s , b ) , · · · , ( s N , b N ) } with s j being the sellerand b j being the buyer, we define average revenue dif-ference of c i as R i = ∑ Nj = ( | revenue ( s j ) − revenue ( b j ) | ) / N Our test cases then are - H : Avg. difference of revenue in a community de-creases as reciprocity increases. H : Avg. difference of revenue does not decrease withincreasing reciprocity.We apply linear regression to check the dependence be-tween these two factors. The correlation between thetwo factors comes out to be − . p -valueof 0 . H ).Hence, our study shows that if reciprocity increases i.e.,if more people indulge into a two way traffic, then theaverage revenue difference decreases i.e., their sales val-es become almost identical. This means that when twosellers in the market almost perform similarly, they tendto support each other. This type of mutual uplifting canpromote both their businesses. Intu-itively one would expect that in a global marketplace, iftwo people belong to same geographical location, thentheir interaction would be much higher on average thanwith other people from different locations. To checkthis hypothesis for Fiverr, we calculate the average vol-ume of sales happening in a particular community andcheck its dependence on the average geographical loca-tion similarity. For each community c i in the network,we define average location similarity as L i = ∑ Nj = ( G s j , b j ) / N where, G s j , b j = s j and buyer b j (bothof them are from c i ) belong to the same geographicallocation and 0 otherwise. The correlation coefficient be-tween average volume of sales and L comes out to be0 . p -value of 0 . Next, wecalculate the average similarity between the gig namesof the corresponding gigs provided by the seller and thebuyer. Following this, we calculate the average gig sim-ilarity in a community. We apply linear regression tocheck the relationship between average gig similarityand reciprocity. Our results seems to strongly favor ourclaim since, as the average similarity between gigs in-creases, so is the propensity to reciprocate. The regres-sion coefficient is 0 . p -value of 0 . We observe that in ∼
60% cases, seller-sellerinteractions happen between same level sellers. There-fore, the next question that arises is - if two same levelseller interact, what is the probability that they recipro-cate?To answer this, we calculate average seller similarity be- tween people in each of the communities. For each com-munity c i we define avg. seller level similarity as SR i = ∑ Nj = ( S s j , b j ) / N where, S s j , b j = s j and b j belong to same level ( level1 or, level 2 or, top rated ) and 0 otherwise. After ap-plying statistical test of significance, we find that thecorrelation coefficient to be 0 . p -value of0 . We form another network on the basis of number ofbuyers common between two sellers and denote this net-work as Tier-II network. We take the top sellers anddraw an edge between a pair of sellers if they share anycommon buyer. We define the weight of the edge as - weight ( a , b ) = | buyer a ∩ buyer b | / | buyer a ∪ buyer b | with buyer i being the set of buyers of the seller i . An-alyzing this network gives us insight about loyalty ofbuyers. For the subcategory Creative Logo Design , theaverage degree of the 282 nodes in the network comesout to be 110.28 with only one strongly connected com-ponent. This is indicative of divided loyalty or no loyaltyof buyers in Fiverr market. Moreover, the network fol-lows small world properties with a diameter of 3. Wealso find that the network is highly connected with anaverage clustering coefficient of 0.34. These results alsoindicate the fact that Fiverr is a marketplace where thecompetition is less monopolistic, i.e., most of the sellerssell similar types of gigs in order to gain revenue andcustomers do not discriminate well among these gigs.
6. Fiverr under strategy lens
In this section, we evaluate Fiverr as a marketplacefrom strategy perspective focusing on the inter-entity de-pendence. For this study, we only choose those sellerswhose sales are more than 500 per gig ( ∼ .
45% of thewhole seller community).
Dependence between no. of gigs sold and revenue
Intuitively, a seller selling more gigs seems more likelyto generate more revenue. We perform statistical testof significance on data of top sellers. We apply linearregression with number of gigs as the independent vari-able and obtain a regression coefficient of 0 . p -value of 2 . × − . This shows that the relationshipbetween the two is weak for Fiverr. Customer churn rate and sales
We define customerchurn rate for a seller for a particular gig as churn rate = ∑ Nt = | customer t ∩ customer t − || customer t − | Here, customer t is the set of the customers of the sellerat time period t and N is total number of time periodshe gig has been selling for in the last nine months.We calculate the average churn rate for all gigs offeredby a seller (fig 13). We believe that the lesser the cus-tomer churn rate, the better the customer handling andability to satisfy a customer, and more the sales. This isof course not true for products with long life-cycles. Forexample, a customer who bought a mobile phone froma seller in Amazon is less likely to return for anotherphone purchase. In low value service driven market-places, customers returning can be a very frequent phe-nomenon. We studied the relationship between churningrate and total sales of Fiverr. Churn rate is somewhat af-fecting the total sales, and the two are positively related(correlation coefficient of 0.12). One reason could bethat most of the customers of Fiverr are just one timebuyers. Average churn rate is higher for many of thesellers. Dependence between similar gigs and sales
We ob-serve that many top sellers sell more than one gig andin many cases they provide their services in the samesubcategory. So, the next question that arises is - howsimilar are their gigs or do they provide different type ofservices? We create tf-idf matrices for 1-gram, 2-gramand 3-gram similarity of the gig names of the gigs theseller offers. As most of the gigs start with the phrase“I will do” or “I will make”, we remove these phrasesand other stop words from the gig names. After that, wetake the average of the three tf-idf matrices to get onematrix M . Standardized rows of MM T gives us the av-erage similarity between one gig and other gigs. Hence,average of all the off diagonal elements of the matrix MM T gives the average gig similarity of the particularseller. We then study the distribution of the similarityvalue (fig 14) and check the significance of similarityvalue over number of sales made by a seller. The re-sult we obtain from the statistical test of significance isa coefficient of − .
216 with a low p -value of 0 . Similarity with other gigs and sales
In the previ-ous subsection, we define the similarity between gigsof each seller. Here we define a similar concept, be-tween a gig and other gigs. Therefore, we focus onthe average similarity between a single gig with othergigs in the same category offered by other sellers. Ina competitive market, we usually believe that the prod-ucts which are somewhat different from the other prod-ucts stand out and make good sales. For example, mostof the niche products like Apple iPhones or automobilecompanies like BMW provide a unique experience totheir customers which creates a strong brand awarenessamong the customers. In case of Fiverr, we wish to ex-amine whether a unique gig that has very low similar-ity with other products in the category is having moresales, or the other way round. The correlation coefficientbetween average similarity with other gigs and salesis 0 . p -value of 0 . t -statistics and a positive correlation clearly showsthat there is a significant dependence between these two.Moreover, in Fiverr, higher similarity with others meansmore prospects of being successful. This tells us thatmost of the products offered in Fiverr are homogeneousin nature and the number of alternative products is alsovery high. Assessing competition in subcategories
We define asimilar concept of similarity between all gigs in eachof the subcategories. If the average inter-gig similar-ity is high for a particular subcategory, that means mostof the sellers tend to sell similar types of gigs in thatcategory. This indicates that the prospect of being suc-cessful in that category is high, if one seller goes withthe trend. This implies that very high competition existsin that subcategory. We observe whether this measureof competition affects the sales of a category. The top-most competitive subcategory came out to be
Video PostProduction Editing with average sales per gig of 1,141.We standardize the total number of sales by the num-ber of gigs in that subcategory to get the average salesper gig in that subcategory. After applying linear regres-sion analysis on the two standardized variables, averagesales per gig and average similarity between gigs, weget a high correlation coefficient of 0 . p -value of 3 × − . This result clearly shows thatas the competition or the similarity between gigs in asubcategory increases, the average sales also increases.This type of phenomenon is very common in an idealscenario of perfectly competitive market. Also, largenumber of buyers and sellers, accurate and necessary in-formation about all the products, very flexible and bar-rier free market structure has made Fiverr a very good https://en.wikipedia.org/wiki/Perfect_competition xample of perfectly competitive marketplace. The av-erage gig price which is very close to the actual marketprice of $5 shows that every entity is a “price taker” inthe market. Evaluation of growth prospective of seller in market
Evaluation of performance of sellers is vital for anycrowdsourcing platform. We use the classical methodof Boston Matrix (BCG matrix ) to evaluate the perfor-mance of the different top 2.5% sellers. The BCG ma-trix allows organizations to identify product positioningin the market and thus they could allocate resources ac-cordingly. We also use the same concepts in order toclassify the sellers in the similar lines. The differentgroups in BCG matrix are1. Stars : The best performing products or, sellers whichgenerate most of the revenue and have positive growthrate and use most of the resources of the organization.2. Question Mark : The products or, sellers having veryhigh growth prospective but currently low market share.They are the potential stars.3. Cash Cow : The market leaders with low growth rate.People invest in them to maintain their status in market.4. Dogs : The products or, sellers with low market shareas well as low growth prospects.Based on the matrix (fig 15) we also divide top sell-Figure 15. Boston Matrix showing sellers in differentquadrants of matrix.ers in four groups - Stars, Cash cows, Question marksand Dogs based on their market share (i.e., proportionof revenue generated) and their growth rate measuredas the slope of gain of revenue in consecutive months.Therefore, if a seller had revenue { r , r , · · · , r N } respec-tively in the time period { t , t , · · · , t N } , then the growthof revenue for the time period { t , t , · · · , t N − } wouldbe { r − r , r − r , · · · , r N − − r N } . We normalized thegrowth by the number of days and applied linear regres-sion to get the slope of change of revenue growth. Wethen use the regression coefficient as the growth rate foreach of the sellers. More negative the slope is, moreis the growth rate of the seller. The below table showsthat there is very low percentage (4.3%) of stars which isa common observation in any market place. Moreover,we see that the top performers in the Stars have very https://en.wikipedia.org/wiki/Growth-share_matrix Quadrant of BCG matrix percentage of top sellers
Stars 4.3%Cash cow 16.85%Question Mark 10.11%Dogs 68.74% high growth rate (slope is close to − ◦ ). Similarly,16.85% of the sellers have a growth rate close to zerobut with high revenue. We further observe that 37.9%of these people are from top rated level. We believe thatthese sellers may have reached their saturation point andpossibly will be taken over by level 2 in the next fewmonths. Similarly, among the Questions Marks level 2 sellers who have a very high growth prospectwith relatively low revenue in the current period. Pro-viding proper incentive to these sellers can help them inbecoming stars. Thus, Boston matrix provides the in-sights about incentivizing the right set of sellers in orderto boost their businesses.
7. Discussions and conclusions
In this paper, we study the characteristic propertiesof Fiverr marketplace from various points of view - eco-nomic, sociological and strategy and make the followingkey observations.The Fiverr marketplace is a unique in that the buyerswho purchase gigs from sellers can convert themselvesto sellers at any point in time. Most sellers are profi-cient in certain type of services which indicates a goodlevel of professionalism in the marketplace. Hence mostof the top sellers prefer selling small number of gigsin a few subcategories rather than offering diverse cat-egory of gigs. From the viewpoint of the average gain oftop sellers, the trends decrease over time. Globally, thepercentage gain of customers for top sellers is around40%. However, some subcategories also show differ-ent trends. The gains sometimes reach as high as 75%.The Fiverr marketplace is dominated by one-time buyerswhich consists of 45% of the total buyers. An interest-ing and unique characteristic in a supply-driven market-place is the existence of loyal customers. It is observedthat many buyers buy the same product repeatedly froma few categories. This also results in the amount ofpurchases not always being directly proportionate to thevolume of purchases. We observe that sellers are moreappeasing in their interactions and try to woo their buy-ers into buying their gigs. Serious categories have com-paratively more reviews with near zero and negative po-larities as compared to the trivial categories.On analyzing the Tier-I seller-seller interaction net-work, we observe that there are many small tightly-knitseller communities existing in the network; howeverthe overall network structure indicate loose connectiv-ity with density of 0.001 among 49.13% sellers. Thevery low clustering coefficient in the network shows thathere are very few or no transitive links in the networkwhich is quite uncommon for general markets. There isalso low reciprocity in the network.We also observe that some sellers buy from othersellers within the same subcategory. The possible rea-sons for this behavior include promoting which is sup-plemented by the buyer having more sales than the sellerand providing positive reviews to the seller. Most of thetop sellers prefer to offer multiple gigs but in the samesubcategory. An analysis of the gigs of the top sellersindicates that the more diverse type of products a sellersells, the more sales he/she can make. In terms of gigsimilarity on Fiverr, higher similarity with others meansmore prospects of being successful. This tells us thatthe products offered in Fiverr are homogeneous in na-ture and the number of alternative products is also veryhigh. Similarly, as the competition or the similarity be-tween gigs in a subcategory increases, the average salesalso increases. This type of phenomenon is very com-mon in an ideal scenario of perfectly competitive mar-ket. Boston Matrix analysis shows that top performersin the ‘Stars’ group have very high growth. At the sametime, about th of all the sellers have almost no growthbut high revenue. Similarly, most of the sellers in the‘question mark’ bracket can potentially become stars ifprovided with the right incentives.
8. References [1] J. Ross, L. Irani, M. Silberman, A. Zaldivar, andB. Tomlinson, “Who are the crowdworkers?: shift-ing demographics in mechanical turk,” in
CHI’10extended abstracts on Human factors in computingsystems , pp. 2863–2872, 2010.[2] A. J. Berinsky, G. A. Huber, and G. S. Lenz, “Eval-uating online labor markets for experimental re-search: Amazon. com’s mechanical turk,”
PoliticalAnalysis , vol. 20, no. 3, pp. 351–368, 2012.[3] G. Paolacci, J. Chandler, and P. G. Ipeirotis, “Run-ning experiments on amazon mechanical turk,”
Judgment and Decision making , vol. 5, no. 5,pp. 411–419, 2010.[4] M. Kosinski, Y. Bachrach, G. Kasneci, J. Van-Gael, and T. Graepel, “Crowd iq: Measuring theintelligence of crowdsourcing platforms,” in
Proc.of WebSci ’12 , pp. 151–160, 2012.[5] W. Mason and S. Suri, “Conducting behavioral re-search on amazons mechanical turk,”
Behavior re-search methods , vol. 44, no. 1, pp. 1–23, 2012.[6] S. Suri, D. G. Goldstein, and W. A. Mason, “Hon-esty in an online labor market.,”
Human Computa-tion , vol. 11, p. 11, 2011.[7] H. Halpin and R. Blanco, “Machine-learning forspammer detection in crowd-sourcing,” in
Work- shop on Human Computation at AAAI , pp. 85–86,2012.[8] M. Allahbakhsh, B. Benatallah, A. Ignjatovic,H. R. Motahari-Nezhad, E. Bertino, and S. Dust-dar, “Quality control in crowdsourcing systems:Issues and directions,”
IEEE Internet Computing ,no. 2, pp. 76–81, 2013.[9] P. Heymann and H. Garcia-Molina, “Turkalytics:analytics for human computation,” in
Proc. ofWWW ’11 , pp. 477–486, ACM, 2011.[10] G. Wang, C. Wilson, X. Zhao, Y. Zhu, M. Mo-hanlal, H. Zheng, and B. Y. Zhao, “Serf and turf:crowdturfing for fun and profit,” in
Proc. of WWW’12 , pp. 679–688, 2012.[11] H. Ge, J. Caverlee, and K. Lee, “Crowds, gigs, andsuper sellers: A measurement study of a supply-driven crowdsourcing marketplace,” in
ICWSM ,2015.[12] M. Motoyama, D. McCoy, K. Levchenko, S. Sav-age, and G. M. Voelker, “Dirty jobs: The role offreelance labor in web service abuse,” in
USENIXSecurity , 2011.[13] N. Christin, “Traveling the silk road: A measure-ment analysis of a large anonymous online market-place,” in
Proc. of WWW ’13 , pp. 213–224, 2013.[14] E. De Cristofaro, A. Friedman, G. Jourjon, M. A.Kaafar, and M. Z. Shafiq, “Paying for likes?: Un-derstanding facebook like fraud using honeypots,”in
Proc. of IMC , pp. 129–136, 2014.[15] G. Wang, C. Wilson, X. Zhao, Y. Zhu, M. Mo-hanlal, H. Zheng, and B. Y. Zhao, “Serf and turf:crowdturfing for fun and profit,” in
Proc. of WWW ,pp. 679–688, ACM, 2012.[16] K. Lee, S. Webb, and H. Ge, “The dark side ofmicro-task marketplaces: Characterizing fiverr andautomatically detecting crowdturfing,” in
ICWSM ,2014.[17] S. Farooqi, M. Ikram, G. Irfan, E. De Cristo-faro, A. Friedman, G. Jourjon, M. A. Kaafar,M. Z. Shafiq, and F. Zaffar, “Characterizing seller-driven black-hat marketplaces,” arXiv preprintarXiv:1505.01637 , 2015.[18] T. X. Liu, J. Yang, L. A. Adamic, and Y. Chen,“Crowdsourcing with all-pay auctions: a field ex-periment on taskcn,”
Management Science , vol. 60,no. 8, pp. 2020–2037, 2014.[19] B. Liu, “Sentiment analysis and opinion mining,”
Synthesis lectures on human language technolo-gies , vol. 5, no. 1, pp. 1–167, 2012.[20] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, andE. Lefebvre, “Fast unfolding of communities inlarge networks,”