Segmenting Early Gameplay to Improve Monetization, Retention and Virality

The Situation:

A client was concerned that one of their flagship Facebook games was underperforming across all key metrics. The company had high hopes for the game and had invested heavily in delivering a quality product. Although A/B testing had been built into the development process, the game, which was a quest orientated fantasy game, was failing to achieve these high levels of expectation.

Metrics had charted worrying trends since the game was launched; revenues were lower than expected, virality was not contributing to forecast growth of users and retention rates were low following initial playing sessions. The revenues, retention rates and other measures were by no means disastrous but the game was clearly under-performing and there was no clear solution on how improvements could be made to increase overall profitability.

GamesAnalytics was commissioned to provide the solution by characterizing playing behaviors using analytics, and devising strategies at individual player level to improve the performance of the game across all key metrics.

The Approach:

It was clear from initial analysis that the early part of game play was vital to the overall success of the game and that it was essential to ensure that individual players had the best possible experience during this period so that revenues could be maximized.

To enable the understanding of different playing patterns, a rich analytics dataset was created with variables that described players’ potential to generate revenue, their potential to introduce new players as well as many other descriptive variables which allowed an understanding of the patterns of game play across significant events in the game.

The intention was to use this data to create player segments by grouping together players with similar playing experiences and potential, and then devise different treatments for each segment at an individual level to optimize their game play.

In this way, each solution would be appropriate to particular player segments, whether it be enhanced tutorials, in-game messaging or specific rewards and incentives to increase revenues and virality.

The Solution:

Statistical techniques were used to build segments of players based on all this information. Each segment had particular characteristics as can be seen in the diagram below.

Each segment is displayed by their revenue potential and virality potential and the types of players are markedly different.

 

For example ‘Social Involvers’, as the name suggests, are highly sociable and will respond well to rewards and incentives to introduce significant numbers to the game. ‘Early Enthusiasts’ and ‘Confident Completers’ have the highest revenue potential and so should be targeted with offers to increase their spend.

The ‘Need Guidance’ segment require enhanced tutorials as their game play suggests that they are struggling with some of the concepts in the game and need additional support to improve their playing experience. Interestingly these players have relatively high revenue potential so we are testing if these players are willing to pay for enhanced support.

The two largest groups, ‘Losing Momentum’ and ‘Sporadic Semi-Engaged’ both have low retention rates and some further analysis was undertaken to get a clear understanding of what issues were driving poor retention. We found that extra features were required to keep these segments interested so the game was developed to provide an additional set of attractive features based on our analysis of their  previous playing patterns.

The Benefits:

By using a combination of targeting in-game messaging to appropriate player types, and improving the overall user experience and features within the game for specific segments, GamesAnalytics was able to work with the client partner to implement solutions that improved all aspects of the games play.

Dashboards where then used to measure the positive effect of these strategies and interventions.

Currently, the client is seeing significant improvements in the % Paying, Virality and Retention Rates metrics. The combination of these improvements is on track to deliver a 20% increase in overall game revenue which is a five-fold return on the investment in the analytics environment.

 

Improving Virality thru Predictive Modelling

The Situation:

One key aspect of measurement which is critical to the success of any game is virality – encouraging existing players to invite their friends to play the game. Virality is an effective and inexpensive way of quickly building the game playing base; and so it is a subject that often a key objective of analytics projects – how to improve game virality?

In 2010, GamesAnalytics was commissioned to undertake an analytics exercise to find ways of improving the virality of our client’s Facebook game based on a treasure hunt theme. Our intention was to assess what factors led players to invite others and how we could use this information to increase the overall number of players that send invites so that the overall numbers playing the game was dramatically increased.

Our Approach:

Our first activity was to collect event data by implementing our COLLECT SDK. COLLECT was configured to gather a detailed range of game events at individual player level that would be the basis of our analysis. By spending time playing the game and becoming familiar with the game mechanics, we were able to configure the event collection so that we had access to a rich and complete dataset to support the analysis.

It was clear from our initial analysis that there were two main types of players; those that were socially motivated to play the game i.e. they were highly likely to send invites to friends, and those whose main focus was to complete the game and had naturally low likelihood to invite.

Further we found that social players were divided into two types. Those that invited early to maximise the rewards they received and those that invited after several sessions and therefore were issuing the invites on the basis of a personal recommendation.

Based on this information, we were able to further analyse detailed event patterns and build algorithms using our PREDICT technology that scored the likelihood of each players to invite.

Our Solution:

By building predictive models, we were able to score each player from 0 (very unlikely to invite) to 1 (very likely to invite). The factors that contributed to the model score are shown in the chart below. As you can see, there are a mixture of variables including some ‘social’ events such as creating wall posts, gifting items, accepting gifts and accepting invites. Other important variables which contribute significantly to the model are number of sessions, total stamina, highest level etc. The chart on the right hand side shows model deciles with decile 10 being the 10% of the playing base with the highest scores, decile 9 being the next best 10% etc.

 

The overall percentage of players that make invites is 3.7%. In decile 10 this increase to 14.8% ie the model is being very effective at predicting players who invite. However, and most importantly, 85.2% of players in decile 10 do not currently invite but are predicted to be highly likely to invite.

In decile 9, we have 10.1% of players inviting and 89.9% who are highly likely to invite.

Therefore our target group to encourage to invite are those players in deciles 9 and 10 who do not currently invite but have a high model score so we predict that they are highly likely to invite. This is a significant volume of the playing base and has highlighted a very important segment of the overall playing base that could fundamentally affect the success of the game.

Once we had isolated there players, we were able to set up in-game messages using our ENGAGE SDK to give these players additional incentives and rewards for inviting 10 or more people. These rewards accumulated as players issued more and more successful invites. We also set up test to determine the level of rewards that acted as the best incentivisation and built this into the in-game messaging.

The Benefits:

By using in-depth analytics and predictive modelling in this way we were able to delivery timely and appropriate in-game messages that encouraged increased virality by delivering appropriate reward mechanisms.

Over the course of the test period, we saw virality increase by 27% which in turn expanded the playing base significantly and did so on a very cost effective basis.