The Impact of User Interactions on Freemium Game Performance

: Company A’s Project R is a freemium-model game business wherein the company makes money by (a) obtaining a large number of users who play its game essentially for free and (b) converting a small number of them into paying users. In Project R, paying ability boosting items were added to increase profits. Doing this initially increased monthly sales by 20%; however, after two months, the playing time of existing non-paying users declined, and more new non-paying users abandoned the game as well. It seems that the addition of paying ability boosting items could shorten a game’s life span by destroying the balance between (a) and (b) and causing a long-term decline in revenue. This paper runs a simulation to verify this.


Introduction
Successful product development requires the employment of a company's external resources, not just internal (Chesbrough, 2003).Although many studies have focused on users as the drivers of product innovation, in the form of user innovation, opinion leaders, and user communities (e.g., Jeppesen & Frederiksen, 2006;Valente & Davis, 1999;von Hippel, 1976von Hippel, , 1986)), they have all concentrated on a particular user group (e.g., Jensen, Hienerth, & Lettl, 2014;Jeppesen & Laursen, 2009).
The revenue model for the Chinese online game discussed in this paper is known as the freemium model.In this model, revenues are earned by obtaining a large number of users who play the game essentially for free and by converting a small number of them into paying users (players who purchase paying items that can be used in the game) (Tanaka & Yamaguchi, 2015).Therefore, several user groups that are paying or non-paying coexist within the same game content.Filtering the large volumes of data about players' playing records, paying history, and reviews and opinions about the games (Huang, 2018) has shown that too much differentiation can spoil the playing experience for some user groups and cause the product's long-term revenues to decline.This paper uses a computer simulation to verify this.

The Case of Project R
The topic of our research in this paper is Project R from Company A, a Chinese developer of online games. 1 Project R officially launched its service a year and a half after it was commenced and had been in operation for two years as of March 2020.Project R consists of 60-70 core team members as a medium-scale team, and a full-time staff of five runs its dedicated operating department.Project R's product is an Android and iOS game application for Chinese domestic and overseas markets.The game's genre is that of the typical Massively Multiplayer Online Role-Playing Game (MMORPG).It ranks in the top 50 in its genre for China's Android and iOS markets and is making inroads abroad.Project R's users are classified into four user groups according on their paying behavior and playing history.According to the monthly data of August 2019 for the Chinese market, 5% of its users were high paying, 10% were low paying, 50% were existing non-paying, and 35% were new non-paying.Project R's revenues come from sales of paying items, and profitability has been steadily improving since the official release.However, although it has reached its KPIs for the levels of user activity and profitability, other KPIs remain unachieved, such as the amount paid for paying items and the paying user conversion rate.Therefore, it proposed methods to increase paying items when it did a large-scale update on its first anniversary.
In the freemium-model game business, companies make money by (a) obtaining a large number of users who play its games basically for free and (b) converting a small number of said users into paying users.Project R's development team put emphasis on striking a balance between (a) and (b).However, Company A was simultaneously involved in several mobile game projects, and the competition between projects for development resources was extremely fierce.As a result, Project R succumbed to the pressure of achieving its KPIs and finally consented to adding paying ability boosting items in its first anniversary updates and in further updates one month later.
Two weeks after the release of the new updates, the addition of paying items had caught the attention of many non-paying users so that the percentage with new paying items jumped.By increasing new non-paying users and increasing the amount paid by existing non-paying users, monthly sales went up by around 20% from the pre-update level.Two months later, however, the increase in ownership of new paying items had started to level off and the activity of existing non-paying users (activation rate of the app and playing time after activation) was down, while large numbers of new non-paying users were abandoning the game.
When the Project R team compared the playing history of existing non-paying users, whose activity had declined with their playing history prior to the updates, they identified a strong correlation with the total win rate.It seems that if the total win rate went below a certain level, the game experience deteriorated, thereby reducing play frequency and play time and even leading to an abandonment of the game.Additionally, an analysis of new non-paying users' playing history prior to abandonment found a trend wherein players would quickly abandon the game if their total win rate went below a certain level.
Project R held several emergency general meetings and arrived at the following consensus view.As in the established policy, when the effectiveness of paying items was raised again and again, non-paying users were put in a no-win position; consequently, not only would non-paying users abandon the games but paying users would also develop "paying fatigue."In addition, future paying users would have to be acquired through the influx of a certain level of new non-paying users and activity on the part of existing non-paying users.If this was not done, the products' long-term revenue potential could decline.

Outline of the Simulation Model
However, was this consensus view correct?In this study, we developed a multi-agent simulation model to compare long-term performance.Below are the specific agent action rules.
1) The agents are placed randomly in a 50 × 50 two-dimensional grid.
2) Each agent moves in a random direction for each step.
3) A 40 × 40 area in the middle of the grid is the playing area, and at each step, agents in the playing area are randomly paired to battle each other.
4) The agent with the most points wins the game.Each agent gets a total number of points showing (i) the initial points, (ii) the play points automatically generated from each game, and (iii) the operation points automatically generated by the opponent's actions.The values for (i), (ii), and (iii) are uniform random numbers between 0 and 1.Additionally, (i) is automatically generated at the beginning of the simulation, while (ii) and (iii) are automatically generated for each game.
Based on data from Project R, we set the game experience at 300 steps.The above rules are shown in Figure 1.
After a game's official release, it becomes difficult to continue development if revenues fall below a certain level, and the project is threatened with cancellation.Therefore, we set the condition for ending the simulation at the same level as Project R's warning level, so the simulation ended when revenues fell below 40% of the initial value.Therefore, earnings were defined by the following equation.
Revenues = number of high-paying users + number of low-paying users × 0.5

Results
Next, we compared the long-term performance of two multi-agent simulation models: (A) a balanced model, which corresponds to the period before the paying ability boosting items were added, and (B) an unbalanced model, which corresponds to what happened after the addition of these items (Table 1).We set the initial composition of each user group to be the same as in the monthly data for August 2019.In addition, agents would immediately disappear from the model (abandon the game) when their total win rate fell below a

Figure 2. Comparison of the number of users in two models
The impact of user interactions on freemium game performance (B) Unbalanced model (a) Composition of points: Set at a uniform random number between 0 and 1 for (i) initial points, (ii) play points, and (iii) operation points from each match, because paying items do affect the battle.However, bonus points were added to a user's play points, depending on the user's attributes.Specifically, a multiplier was applied to the points in (ii) to enhance the effect of the game's ability boosting items (see Table 1 for details).
(b) Condition for abandonment: The threshold for each user group's total win rate was determined by the users' attributes because of the fact that paying items affect the match.For paying users, the expected value of the game experience would increase with the purchase of paying items.
Each simulation was run 100 times.The mean ending number of steps for the simulations was 5,547 for (A) the balanced model and

Sensitivity Analysis of the Simulation Model
We conducted a sensitivity analysis on the unbalanced model to see how the use of paying items and the criteria for abandonment by paying users would impact the operating period.
Huang 106 For the impact of paying items in the unbalanced model, we tested the change in impact by setting multipliers for a low-paying user's play points at 2 × and for a high-paying user's play points at 3 × and left the other conditions unchanged from their initial level.For this, we ran a simulation that changed the impact rates in increments of 0.5, with low-paying users ranging from 1.5 to 2.5 and high-paying users ranging from 2.5 to 3.5.We then analyzed the sensitivity of paying items' impact in the three cases.The outcomes are shown in Table 2. First, when the impact was minor, the mean number of steps to the end was around 3,200 or about 500 less than the normal value.Moreover, when the impact was major, the mean number of steps to the end was approximately 4,000, or 300 more than the normal value.
In addition, with respect to the impact of criteria for abandonment on product performance, in the model discussed above, we used Project R's user abandonment criteria and set abandonment at a total win rate of 0.6 or less for low-paying users and 0.65 or less for high-paying users.Subsequently, we ran a simulation that changed abandonment criteria in increments of 0.1 in the unbalanced model, with low-paying users' abandonment varying from 0.58 to 0.62 and high-paying users' abandonment varying from 0.63 to 0.67.The outcomes are shown in Table 3.If abandonment criteria are set low, there are 800-1,200 more steps, and when the abandonment conditions are set higher, there are 400-600 fewer steps.This sensitivity analysis confirmed that the unbalanced model is shorter than the balanced model (5,547 steps), so the preceding simulation outcomes are probably robust.

Conclusion
As the case of Project R suggests, adding paying items in the simulation temporarily increased earning short-term revenues but caused them to decrease in the long run.The importance of differentiating customers' experiences has been observed as a method to make the freemium model profitable (Liu, Au, & Choi, 2014;Seufert, 2014).However, this is premised on disparate user experiences in each case or on a weak correlation between Huang 108 users.When the experiences of different users are in conflict or impact each other, excessive differentiation can harm the experience of some user groups and cause long-term performance to deteriorate.

Figure 1 .
Figure 1.The flow chart of a typical agent of points: Set at a uniform random number between 0 and 1 for (i) initial points, (ii) play points, and (iii) operation points from the matches, since paying items have no effect.(b)Condition for abandonment: The threshold for each user group's total win rates was set at 0.45.

3
,682 for (B) the unbalanced model.Changes in new non-paying users and existing non-paying users are shown in Figure2.In (A) the balanced model, both new non-paying users and existing non-paying users slowly decline, and a certain number of both non-paying users and existing non-paying users remain even at the end of the simulation.However, in (B) the unbalanced model, the number of non-paying users drops dramatically so that they disappear during the simulation, and after around step 2,000, the situation becomes one of matches of paying users against themselves.

Table 1 .
Comparison of two models

Table 2 .
The results of sensitivity analysis for changes in effect

Table 3 .
The results of sensitivity analysis for changes in