How Game Catalogues Are Ranked, and What Personalisation Does to Discovery

Landing at a game catalog without the personalized social feeds of a Steam or Facebook Groups is going to give you a curated popularity heatmap of the top downloaded games. Online re-ranking just culls from that short-list to hone it down to titles that you are statistically more likely to enjoy, based on your unique game habits. The definitions and exclusions that govern this process are based on greyscale summary points, rather than personal identification, such as 'Positive Reviewed (Steam and Metacritic)'.

These two ranking functions, based broadly on popularity and personalized to individual conversion, is a basic social network that can be understood and manipulated. It's worth noting that these 'recommendations' are more closely aligned to a popularity based dashboard.

Beyond these automated recommendations, there's a human level to this as well. Users are able to pick 12 existing tags to exclude from each queue. Whomever the games beat out for a coveted recommendation feel the effect of this, being pushed further down the user's discovery queue.

It's one thing to be on a discovery queue in Google Play or Steam anyway- but to be blocked out of collections you're Pitching for when the user's been putting orders in for years is an altogether different level of exclusion. However, examiners at Google and iGaming believe it's just a way for a recommender to fall back to a popularity-ranked item list, as in the cold start problem.

Google Play's recommendation system uses a two-pronged approach: offline candidate generation and online personalized re-ranking. While Google uses other users and machine learning models, Steam prioritizes games that the user has previously played and matching them via tags.

While Steam users can actively exclude a collection of tags, Google Play's personalized ML model does all the work in a singular swooping pass. Regardless of strategy, the end result is the same: games are prioritized by popularity and then re-ranked by user data.

However, this user-based data causes concentration on a fraction of the catalog, and a narrow mission focus can deplete that space.

While the Steam community has put together, you 'Style/Belong' to this category by actively choosing to exclude the Steam style genre tags. Steam play, however, doesn't let you exit the Steam genre group directly from the catalogue. Google play, meanwhile doesn't invest any development time into notifying users of this.

Google uses a pre-trained model to generate personalized recommendations, which are then refined using candidate generation. This system uses "style" and "fandom" identification. Yet, they don't let users know their recommendations are boxed into these categories, let alone where they can pursue the rights to opt out.

Additionally, one way out of cold start is being a player who shows interest in a broad rank of items. The fastest way out of cold start is to collect explicit preferences at onboarding.

Another common way for the discovery system to suggest new games is to surface recently played and favorites. This can lead to a vicious cycle of the same games being recommended over and over again, as the system relies heavily on user behavior.

Ultimately, the goal of any discovery system is to optimize for what can be measured and surfaced. This means that recently played games, favorites, and items with usable metadata are prioritized, while the long-tail catalog can be unseen.

These cold start challenges mean that it's difficult for smaller developers to break into the recommendations without a massive promotional campaign- which they just paid a hefty tax to announce. Meanwhile, major publishers waste their efforts on the duplicate marketing tactics that have already saturated the market.

So what does that mean for you? As a buyer, it's important to explore the catalog beyond the recommendations and search for games that interest you. As a developer, it's crucial to optimize your game's metadata and collect explicit user preferences to increase discoverability.