A Little Too Personal: Effects of Standardization versus Personalization on Job Acquisition, Work Completion, and Revenue for Online Freelancers

Gig work has been defined as “electronically mediated employment arrangements in which individuals find short-term tasks or projects via websites or mobile apps that connect them to clients and process payment” (Kuhn and Galloway, 2019). In short, online labor platforms act as an intermediary (providing boundary resources such as communication channels, evaluation metrics and automated transactions (Jarrahi et al., 2020)) between freelancers who seek jobs and clients who look to hire professionals to complete various forms of work. In this study we focus primarily on virtually-deliverable and knowledge-intensive work, as opposed to physical services such as furniture assemblage offered by gig workers on platforms such as TaskRabbit, ridesharing services that are now commonly provided by Uber/Lyft, or microwork such as those found on Amazon Mechanical Turk. The digital nature of gig work suggests the prospects of greater work flexibility and independence, and many workers report pursuing gig work with these benefits in mind (Hall and Krueger, 2018). However, such affordances comes at the cost of several unique challenges. For example, gig workers may face professional isolation and atomization over the long term (Yao et al., 2021), causing them to obtain fewer networking and advancement opportunities (reduced lateral and upward mobility, respectively) as well as limited social support while enduring fierce competition (Kost et al., 2020). Beyond social isolation, workers must also contend with unique day-to-day difficulties, such as income instability and the need to self-manage, e.g. coordinating their time and resources, maintaining productivity, self-advertising, proactively seeking out new work, building reputation, and maintaining client relationships (Sutherland et al., 2020).

The unique structure of online labor platforms introduces elements of uncertainty and information asymmetries (sometimes intentionally), which can create power imbalances that favor clients and enable platformic management (Jarrahi et al., 2020; Newlands et al., 2017; Evers et al., 2018; Sutherland et al., 2020; Rosenblat and Stark, 2016; Jarrahi and Sutherland, 2019; Kingsley et al., 2015). For instance, clients on Upwork are not required to disclose their identities (a privilege not afforded to its freelancers) and they may also leave private reviews for workers they’ve hired; freelancers, on the other hand, cannot even access the other bidders of a project they apply to, nor can they see who the ultimate winner is (Jarrahi et al., 2020). Specific components of platform structure, such as calculated ratings, have also been found to increase power asymmetries and worker precarity (Sutherland et al., 2020).

Compounding on their already precarious job situations, freelancers can be highly susceptible to volatility in the marketplace. For instance, Huang et al. (Huang et al., 2020) found that, during the 2008 recession, an unemployment increase of 1% was associated with a 14.9% increase in project bidding and a 21.8% rise in the number of active workers. More recently, Sutherland et al. (Sutherland et al., 2020) discovered that the COVID-19 pandemic caused a decrease in worker-controlled flexibility, along with increased competition, exploitation and workload intensity. As we gradually transition from this era of work from home, many employees face the dilemma of whether to remain remote (Barrero et al., 2020). Some companies hesitate to offer such long-term remote work options for their employees, which urges many workers to turn to freelancing alternatives. This looming wave of novice online freelancers, who likely intend to remain for the long term, poses many questions about successful strategies for online freelancing.

In a systematic review of the sharing economy in computing research, Dillahunt et. al. focused on the computing (HCI in particular) community’s contributions toward the sharing economy as well as underexplored and unexplored topics for future research. (Dillahunt et al., 2017). This prior literature review suggests that existing HCI studies on the sharing economy has been largely descriptive and qualitative. To diversify the range of HCI approaches applied toward the gig work context, we present here a quantitative study that leverages the aggregate past experiences of workers. The literature review also suggests a need for human-centered optimization that increases the decisional autonomy and long-term performance of workers, while minimizing overheads such as reduced availabilities, monetary cost, and worker burnout (Dillahunt et al., 2017). Within this empirical investigation we offer an initial identification of effective project-level and long-term self-management strategies, to inform novice freelancers about what, when, and how much to communicate with their clients, and more generally about the overall marginal benefit (or cost) of ‘personalizing’ service delivery, e.g., tailoring communications or work schedules to one’s client.

Compared to the more organization-centered employees of the traditional labor market, workers of the gig economy are individually-organized and experience many of the same challenges as entrepreneurs at the beginning of their gig career. With this in mind, we draw on literature in entrepreneurship to identify relevant strategies that could be applied to online freelancing. For resource-constrained entrepreneurs, prior work (Busch and Barkema, 2021; Vanevenhoven et al., 2011; Atarah et al., 2021) identified the strategy of bricolage - the act of creatively working with available, limited resources, and adapting them toward new or important purposes; or as Levi-Strauss put it: make do with “whatever is at hand” (Lvi-Strauss, 1966). Online freelancers are also bricoleurs when they leverage available resources (such as messaging systems, client reviews and job descriptions to learn more about the requirements of a gig) to tailor their pitches toward employers accordingly.

One particular way of circumventing resource constraints is by engaging in network bricolage - where workers utilize their network resources in a manner that is different from the original basis for the connection, thereby creating new opportunities (Chang and Webster, 2019). In the online gig economy, such resource-creators may find relational support by engaging with offline networks, developing a mentorship relationship with senior freelancers, or cold-emailing potential clients. Because the “infrastructure supporting individuals’ careers in the gig economy is deeply relational in nature” (Ashford et al., 2018), it is important that freelancers accrue portable human capital. After amassing such social capital, workers may maintain their reputation using various strategies such as keeping a high and positive rating, reaching out to past clients or cultivating relational agility by productively forming, maintaining and dissolving work relationships (Sutherland et al., 2020; Blaising et al., 2021).

To overcome information asymmetries in online labor markets, workers may engage in prosocial network bricolage behaviors to build connections, so that successful freelancers can share experiential knowledge and novice ones can gain from the collective advice of more veteran peers. Social media groups, for instance, serves as a key resource for informational peer support for rideshare drivers, helping them alleviate the burdens of atomization of being geographically dispersed (Yao et al., 2021). In freelancing platforms such as Upwork, novice workers may leverage the advice of more senior and successful freelancers (Jarrahi et al., 2020). To circumvent power asymmetries and platformic management, workers might take courses to gain algorithmic literacy about the platform (Jarrahi et al., 2020) or experiment with it themselves to develop strategies such as “saving searches” to improve recommended jobs or asking clients to report multiple hours of work as one condensed hour to improve their hourly rates (Jarrahi and Sutherland, 2019).

For digitally-mediated work, many of these network bricolage behaviors involve some means of direct communication between the client and worker, such as emailing or messaging. While there are workers who use opt to use external communication tools to deal with technical inefficiencies, unreliability and monitoring concerns (Jarrahi et al., 2020; Jarrahi and Sutherland, 2019), most online labor platforms provide a form of direct messaging system to mediate textual exchanges between workers and employers. In fact, one of the platform’s core functionalities is to facilitate communication between transacting parties (Jarrahi et al., 2020). Thus, we focus primarily on communication strategies that freelancers commonly engage in when chatting with potential clients.

While we intend to study strategies that are applicable during all periods of a freelancer’s career, it is important to distinguish different points of a project cycle. In this paper, we will consider both the initial, pre-contract (bidding) stage and the project execution stage:

1. (1)

   Bidding stage: Client may interact with multiple freelancers prior to offering the job to selected candidate(s), both clients and freelancers may negotiate and clarify the scope of work before finalizing on price

2. (2)

   Execution stage: After the client makes the job offer and sends deposits promised compensation via an Escrow, the worker begins work to complete job demands

At the initial bidding stage (1), freelancers may attempt different techniques to garner the attention of a potential employer. These may include stylistic techniques such as the use of custom signatures and uppercase words for emphasis, as well as content curation strategies such as using templates to quickly initiate conversations with multiple employers. On the other spectrum, some freelancers may also choose to personalize the content of their (bid) messages to accommodate job demands. Following initial introductions, the freelancer and client negotiate to settle on a price and review contract terms to clear up points of confusion. At this point, if the project is ill-matched, either party may choose to reject the collaboration. During the execution stage (2), the worker may provide progress updates, request additional clarifications, or ask for milestone payments while the client can ask for updates to monitor progress. Note that after the successful completion (or abandonment) of the project, users may request reviews, provide reminders about payments, or bring up opportunities for future collaborations.

Evidence abounds that direct communication benefits sellers on digital marketplaces, including Alibaba, Amazon, and Travelweb (Lim et al., 2006; Tan et al., 2019). For instance, the use of live chat on Alibaba can increase purchase probability of tablets by 15.99% (Tan et al., 2019). In online freelancing, it has been found that workers are 8.9% more likely to be hired if they initiate a direct message to a potential employer when submitting their bid (Hong et al., 2021). In our own sample data, we qualitatively observed comparable patterns of benefit. Specifically, we see that freelancers who employ high-quality templates (e.g., containing examples of past work and self-promotional messages) tend to receive more responses from clients. Further, we observe that freelancers who proactively provide progress updates to their clients during the course of a project are more likely to successfully complete the work and receive payment.

With the exception of (Hong et al., 2021), the present body of literature has yet to systematically investigate the communication strategies employed by workers in the context of online labor platforms. We thus currently have a limited understanding of the alternative work strategies that gig workers employ during job search and project execution, and the relative efficacy of each. Hence, we explore those questions here.

According to De Vries, standardardization is defined as “the activity of establishing and recording a limited set of solutions to … problems directed at benefits for the party or parties involved balancing their needs … expecting that these solutions will be repeatedly or continuously used during a certain period by a substantial number of parties for whom they are meant” (Münstermann and Weitzel, 2008). For independent contractors, it is certainly expected that their services will be used among multiple parties. Meanwhile Lehr (Hanseth et al., 1996) considers standardization to be the “social and technical process of developing the underlying artifact related to [information infrastructure] - … standards that govern the communicative patterns”. Such procedures for developing standardized communication brings us closer to process standardization, involving “the development of a common approach to such activities as establishing (and evaluating) a distributor network … the underlying approach to relationship development strategies”(Griffith et al., 2000). In (Griffith et al., 2000), Griffith et. al. explores communication strategies applied across different cultural contexts, and finds that standard processes may beneficial when applied to nations of similar cultural types, but not necessarily on a global scale. But to the extent of our knowledge, there exists no prior establishment of standardization measures for communicative practices in global, online freelancing platforms.

In corporate contexts, communication has been found to benefit organizations during challenging or exciting times, while ill-conceived or incomplete communication caused by poorly constructed and delayed messages has the potential to turn small issues into major crises (Newman, 2016). In business contexts (service sectors in particular), process standardization offers profitable outcomes by helping define clear and precise output objectives for the service provider, and by better facilitating communication and coordination between exchange partners through increased uniformity of process activities (Ramakumar and Cooper, 2004; Münstermann et al., 2010; Wuellenweber et al., 2009). For the freelancing context, we consider content standardization to be the process where workers repeatedly use messages constructed from templates to promote their services toward multiple clients or job postings, and temporal standardization to be the practice where freelancers message around a fixed time of the day across various projects.

In May 2020, Upwork (a leading online freelancing platform) presented a set of proposal templates as resources to guide beginner freelancers. But by June of 2021, the use of templates is no longer recommended and instead it is suggested that freelancers should “focus more on [specific] project needs”, suggesting that the benefits and harms of template use in bidding could be complicated (noa, 2021).

There’s no shortage of existing frameworks for personalization, especially within marketing literature and persuasive (mobile & e-commerce) technologies (Fan and Poole, 2006; Zhang, 2007). For technology, Blom defined personalization as “a process that changes the functionality, interface, information content, or distinctiveness of a system to increase its personal relevance to an individual” (Blom, 2000); meanwhile the Personalization Consortium defined it as “the use of technology and customer information to tailor electronic commerce interactions between a business and each individual customer.” (Li and Karahanna, 2015). In business contexts, personalization may entail “Customizing some feature of a product or service so that the customer enjoys more convenience, lower cost, or some other benefit” (Peppers and Rogers, 1997), and in internet marketing it has been considered “A specialized form of product differentiation, in which a solution is tailored for a specific individual” (Hanson, 2000).

The definitions of personalization presented so far addresses the specialization of content or a service for an individual, which can be achieved through the presentation of curated options based on known information about their target user or customer - a process that Churchull refers to as outcome personalization (Churchill, 2013). But consider process personalization (which occurs in service encounters) where information is collected about a customer through realtime interactions, and instead of focusing solely on the outcomes, increasing the quality of interaction and delivery are also a part of the objective (Churchill, 2013). Reflecting this more interactive definition, the Personalization Consortium expands on their previous definition: “Using information either previously obtained or provided in real-time about the customer, the exchange between the parties is altered to fit that customer’s stated needs as well as needs perceived by the business based on the available customer information” (Li and Karahanna, 2015). Dyche and Robert’s respective definitions are also more process-oriented “the capability to customize customer communication based on knowledge preferences and behaviors at the time of interaction” (Roberts and Zahay, 2012), and “The process of preparing an individualized communication for a specific person based on stated or implied preferences” (Dyche et al., 2002).

In online chats that are devoid of physical signals from body language or tone of voice, personalizing interactions through messaging content and pace can be of paramount importance for improving interactions between transacting parties. Indeed, Blom identified that a key motivation for using personalization to be the enablement of access to information content [for the customer/client], which can help facilitate interactions and transactions (Blom, 2000). Process personalization can also be personalized: “customized personalization is about personalizing to the consumer’s interactional style and needs in the moment, as well as more stable or longer-term facets such as their demographic profile and/or manifest tastes”, but the customized personalization can have varying effects depending on context and degree (obsequiousness, for instance, can be upsetting) (Churchill, 2013). In service encounters, personalization improves customer impressions (Surprenant and Solomon, 1987) and in a persuasion study, personalized mobile messages successfully helped individuals by significantly reducing daily snacking (Kaptein et al., 2015). But in the context of student-advisor instant messaging interactions (Siqueira and Herring, 2009), the adoption of accommodating temporal patterns has been shown to disrupt one’s own temporal consistency.

For the freelancing context, we define personalization as the way in which a worker caters to the needs of a client by incorporating relevant job specifications into their message text or by client messages, at the expense of their own work schedule or time zone. At the outset, the relationship between standardization and personalization may seem divergent and potentially conflicting, and prior literature has long recognized the tension between information standardization and flexibility (Hanseth et al., 1996). However, we discuss below how these practices might coexist and the potential trade-offs between the two in terms of their effects on outcome success during different stages of a project, as well as over a freelancer’s long term career trajectory.

To conduct this study, we obtained data from a corporate partner (whose specific name will not be disclosed per agreements for data sharing) that is a leading platform in online freelancing. Example categories of work include data entry, software development, design, writing, etc. The dataset we acquired consisted of 2,031,068 messages, from 56,222 projects posted between January 1, 2010 and March 1, 2010, involving 58,397 freelancers and 25,480 clients. For each project we observed their associated project descriptions, bid text, messages, as well as timestamps for these artifacts such as the submission and award dates of bids, the completion and payment times, as well as individual message timestamps. We did not impose limitations based on project category. For each stage of a project, we constructed two separate data frames using this sample. Observations in the first frame consisted of worker-job pairs (or conversations) that incorporated worker-related information such as bid price, bid text, reputation as well as information associated with the job, including project description text, submit date and buyer identification. In a separate freelancer-level frame we included long term worker-related attributes such as average bidding price and bid volumes.

We operationalized standardization and personalization in communication depending on the phase of the project. To more precisely capture standardization in the execution phase, we removed freelancers who multitask and work on more than one project at once – multitaskers represented roughly 12% of those who were awarded projects.

For measuring success at different stages, we gather the job award status to assess the outcome of the bidding stage, project completion status for the execution stage, and overall monthly revenue to account for long-term earnings. Both award and completion status are binary variables where “awarded” or “complete” corresponds with 1 while all other statuses (“rejected”, “incomplete”, or “pending”) are marked as 0. Revenue is a dollar amount calculated on a monthly basis, the final value of of revenue per month is normalized with standard scaling.

Beyond these key explanatory variables, we also include other controlled variables: reputation is measured by whether the worker has received reviews in the past (had prior reviews), bid price is the amount that the freelancer is proposing to charge for their work (this variable is log-transformed to remove skewedness), freelancer message count is the total number of messages the freelancer sends within the project, number of bids won and projects completed account for how many projects the freelancer’s has historically being hired for and completed, respectively, and are also log-transformed. All predictor variables are normalized for analysis via standard scaling. In Table 1 we provide descriptive statistics of both key and control variables for each of our models.

Using separate linear regression models for different stages, we observe the effects of standardization and personalization techniques toward project hire, completion outcomes as well as earnings. When testing the hypotheses about the bidding and execution stages (H1 and H2), we eliminate the possibility that hiring and completion statuses are jointly determined with our explanatory variable by including a project-level fixed effects when running the logistic regression model. This captures time-invariant and job specific properties that might impact the model outcomes, as well as employer-level fixed effects, since there can be only one employer per job.

The models also include observable worker characteristics that may vary across bids such as reputation status and bid price. At the revenue level (R3), we first ran a regression model that used the four aforementioned strategies (measured by our key variables) to predict monthly revenue. Subsequently, we used the two bidding stage measures to predict the total freelancer bidding volume over the three month period to provide further insights for results of the revenue model.

Table 2 shows the bidding stage regression results, where we explore the impacts of standardizing and personalizing first messages on the project award outcome (1 is awarded and 0 if not). The coefficients show that increasing standardization during the bidding stage hurts a freelancer’s hiring probabilities, thereby supporting H1a. Specifically, standardizing first message content by one standard of deviation reduces their winning probabilities by .03%. Meanwhile, personalizing and curating the contents of a bid proposal based on the job posts increases their chances of winning the project (which is in alignment with H1b), but only slightly – personalizing bids by one standard of deviation improves award probability by .036%.

Note that we also controlled for freelancers’ bid prices (which were log transformed after adding one since the log of zero is undefined), reputation – measured via the dummy variable had prior reviews, which represents whether the freelancer has received a rating for their work in the past, and a historical bidding success variable – the number of bids the freelancer won prior to the current project. We intentionally chose to not include actual rating values because the majority of ratings are positive and most jobs do not end up receiving reviews - their inclusion would cause an inflated measure of reputation. As one would expect, having previously won bids and reviews to showcase on the profile is favorable for hiring, whereas bidding at a higher price harms hiring probabilities of a freelancer.

Table 3 shows our results for the execution stage model, where we explore the impacts of standardizing or personalizing responses time on the job completion. In this stage, we observe that in alignment with H1a, being online at regular hours of the day has a small and positive but insignificant effect on a freelancer’s chances of completing a project Meanwhile, being highly responsive to client messages (the personalization technique) significantly hurts completion, which is in agreement with H2b, but the effect is negligible.

Reputation and bid prices have a similar effect as in the bidding stage model. This suggests that reputable freelancers have higher chances of satisfying the demands of a client. Workers who demand higher payments will have a harder time gaining approval from their clients, since more costly payments will likely lead to increased expectations for work quality. Having received ratings for prior work is positively correlated as well. We also controlled for the number of messages that a freelancer sends within the project, since message frequency will have a consequential impact on the variance/regularity of a worker’s messaging schedule, and found that messaging more positively impacts completion probabilities. Meanwhile, having successfully completed projects slightly helps execution of the current one.

Our earnings model uses a freelancer-level instead of a project-level frame to capture revenue from all jobs of a month. Here we measure the effects of the same two pairs of standardization and personalization techniques above to investigate the question posed in R3. The four measurements are aggregated for each freelancer frame by averaging, and fixed effects are added to account for time variance.

Table 4 reveals that only messaging practices in the bidding stage had significant impacts on overall revenue. Specifically, standardization has a positive effect on revenue - increasing content standardization by one standard of deviation results in a .12% growth in monthly revenue, likely because it enables workers to submit more bids. Meanwhile bid personalization no longer offers the same enhancing effects it had at the project level. In fact, personalizing bid content by one standard of deviation can cost workers .15% of their monthly revenue. Reputation continues to impact success in the same ways as before, and bid higher for individual projects naturally increases overall freelancer earnings.

To test our hypothesis that the inverted effects of content standardization is related to how it enables workers to take on larger volumes of work, we ran an additional model using bid volume as the dependent variable. The results show that increasing first message standardization by one standard of deviation during bidding can allow workers to apply to 36.6% more projects, thereby increasing their total earnings in the market.

Given these empirical findings, we propose design recommendations for tools that seek to support gig workers in their various endeavors. Since temporal responsiveness was shown to be harmful toward project completion success, designers might consider mechanisms that help workers stay focused and on task. This may take the form of an application or plugin, which may adopt features akin to those found in focus and productivity apps. Current platforms such as Upwork may also want to reconsider the inclusion of responsiveness ⁴⁴4https://support.upwork.com/hc/en-us/articles/211062968-My-Stats in worker profiles, since a worker’s ability to respond to messages quickly might negatively impact their ability to finish a project.

To make bid personalization easier for workers, tool designers might attempt to use natural-language processing (NLP) methods to extract job requirements from project descriptions and surface them to workers in a more readable fashion. Note that even though current systems do have skill tags that allow clients to clearly define the scope of their project, we can expect many jobs to have unique specifications that cannot be captured by the limited options of a skill tag drop-down. Finally, for workers with relatively low monthly revenues, tools can provide reminders to motivate them to submit more bids and so that they may maximize their hiring probabilities and work volumes.