Question
About 20,000 games are released on Steam every year, and most of them are barely seen. What choices can an independent studio without a big publisher make before launch to improve its chances of being noticed?
Data
- 92,492 games released between 2018 and 2025: store data, tags, languages, price and monthly review history.
- I separated out games backed by major publishers; the question is about the organic success of indie games.
- I used first-year reviews as the measure of success, and counted a game with more than 500 as a “hit”.
- Data quality turned out to be a finding too: player tags before 2021 are selective, with successful games carrying twice as many tags. I recalculated tag-based results on games from 2021 onwards.
Method
The study follows the Google Data Analytics framework: ask, prepare, process, analyze, share, act.
- Pre-registration. I wrote down and dated five hypotheses before looking at the data. Analyses that were not pre-registered are labelled “exploratory”.
- Methods that get closer to causality. Propensity-score matching, regression discontinuity with placebo thresholds, and repeated language snapshots to check whether the relationship runs the other way.
- Machine learning. A LightGBM model split by time (training up to 2023, testing on 2024–2025), SHAP contributions, and text models that extract signal from the store description.
- Map. A UMAP embedding that places similar games next to each other; on the site, every star is a game.
Findings
- Most games go unseen. 54% of indie games get fewer than 10 reviews in their first year; 4.8% pass 500. Yearly releases grew 2.8 times, yet the hit rate stayed between 4.0% and 5.4% every year.
- Early Access spreads interest over time. Games whose interest carries into a second year have twice the hit rate of those that flare up and fade.
- Chinese support is linked to success, but partly in reverse. Games with Simplified Chinese get 57–75% more reviews. But the most successful games add Chinese after launch 13 times more often; success also brings the language.
- Familiar wins. The more unusual a game’s tag combination, the less attention it gets.
- Copying a hit doesn’t work. 15.5% of games that closely resemble big hits become hits themselves; so do 16.7% of the games released into the same region before the hit.
- There is no “50-review threshold”. Across one million monthly slices, there is no jump at 50 reviews.
- Store text carries information. Adding the description raises the model’s ability to tell hits apart (PR-AUC) from 0.27 to 0.31. Long, concrete descriptions go with success; words like “unique” or “epic” do not.
What I don’t know
- These are associations. The quasi-causal methods rule out some alternative explanations, but they are not an experiment.
- My own “inverted U” hypothesis, that moderately unusual games would get the most attention, was not supported.
- The text model’s result is an upper bound; even after removing award and sales sentences, I can’t be sure all leakage is gone.
What I learned
Writing the hypotheses down in advance made it easier to report them honestly, even when the result was not what I expected. It also made clear to the reader which analyses were planned and which came afterwards.