A soft launch exists to answer one question: if we spend money acquiring players for this game, do we get more back than we put in? Everything else is a proxy for that. The mistake we see most often is teams collecting twenty metrics, reading them in the wrong order, and talking themselves into a global launch that the data never supported.
The order matters because the metrics form a chain. A game with weak retention cannot be rescued by clever monetisation, and a game nobody installs cannot be rescued by either. So read them in the order the player experiences them.
1. Retention, before anything else
Retention is the first gate because it is the hardest thing to fix later. Monetisation can be tuned in a week. Ad creative can be replaced in a day. A core loop that bores people on day two is a rebuild.
The numbers below are the rules of thumb we work to. They move with genre, market and year, so treat them as a starting point for your own benchmarks rather than gospel:
| Genre | D1 | D7 | D30 |
|---|---|---|---|
| Hyper-casual | 35 to 45% | 8 to 15% | 2 to 5% |
| Casual puzzle | 40 to 50% | 15 to 22% | 6 to 10% |
| Midcore / strategy | 35 to 45% | 15 to 25% | 8 to 15% |
Two practical warnings. First, compare like with like. D1 measured on organic installs from friends and family is not D1 measured on paid traffic from a broad audience, and the gap is usually large enough to change a decision. Always read retention on paid, non-incentivised installs.
Second, watch the shape, not just the points. A curve that drops hard between D1 and D3 and then flattens tells you the onboarding is losing people who would otherwise have stayed. A curve that decays steadily all the way out tells you the mid-game runs out of content. Those are completely different problems with completely different fixes.
2. Session depth, which explains the retention
Retention tells you something is wrong. Session metrics tell you what. The three we read together:
- Average session length. Short sessions in a puzzle game usually mean the difficulty curve spikes early.
- Sessions per day. Low count with healthy length means players enjoy it but have no reason to come back. That is a live ops and notification problem, not a gameplay one.
- Tutorial and level completion funnels. The single most useful chart in a soft launch. If 30% of players never finish the tutorial, nothing downstream matters.
Instrument the funnel before the soft launch, not after. Retrofitting analytics events into a live build costs you a release cycle and the first cohort of data is the most valuable you will ever get.
3. Monetisation, once retention holds
Only once retention clears your genre threshold does monetisation data mean anything. Optimising ARPDAU on a game that loses 90% of players by day seven is polishing something that is already gone.
The numbers worth tracking:
- ARPDAU (average revenue per daily active user), split by ads and IAP so you can see which lever is actually working.
- Rewarded video opt-in rate. In a well-designed casual game this is often above half. A low rate usually means the reward is not worth the interruption, or players cannot find it.
- Conversion to first purchase, and how many days it takes. A long delay is not necessarily bad, but it changes how much UA you can afford to front-load.
4. Acquisition cost, last
CPI is read last because it is the most volatile and the most fixable. Creative changes it. Geography changes it. Seasonality changes it. A high CPI on a game with strong retention is a marketing problem you can iterate on. A low CPI on a game with weak retention is a trap that lets you buy your way to a bigger failure.
The number that actually decides the launch is the relationship between what a player costs and what they return:
If estimated LTV over your payback window is comfortably above CPI in your target markets, you have a business. If it is close, you have a project that needs more work. If it is below, no amount of ad spend will fix it.
Be conservative with the payback window. Modelling LTV over 365 days is how teams convince themselves a marginal game works. Most studios cannot finance a year of negative cash flow, so model over a window you could actually survive, often 60 or 90 days, and check whether the game still clears.
Where to soft launch
The classic test markets are chosen because they are English-speaking, cheap to buy traffic in, and behave enough like larger markets to be predictive. Canada, Australia, New Zealand, Ireland and the Philippines all get used for this.
Two rules. Do not soft launch in a market you intend to make real money in, because early negative reviews on a rough build follow the listing. And do not read a market whose spending behaviour is nothing like your target; cheap installs from a low-ARPU country will make your retention look reasonable and your revenue look broken.
How long to run it
Long enough to see D7 on a cohort that was acquired the way real players will be, and ideally D30 on the first cohort. In practice that is usually four to eight weeks. Any shorter and you are reading noise. Much longer and you are burning runway to confirm something you already knew.
Run changes as proper iterations. Change one significant thing, wait for a fresh cohort, read it. Teams that change five things at once learn nothing except that the number moved.
When to stop
This is the part nobody writes about, so plainly: if after two or three genuine iterations the retention curve has not moved toward your threshold, the concept is probably not the problem you can fix with another sprint. Killing a build at the end of a soft launch is a successful outcome for a soft launch. It cost you a few months instead of a few years.
The studios that survive are not the ones that never make a game that fails. They are the ones that find out early and cheaply.
A short checklist
- Analytics and funnel events instrumented before the first test install
- Retention read on paid, non-incentivised traffic only
- Genre-specific thresholds agreed with the team before you see the data
- One significant change per iteration, fresh cohort each time
- LTV modelled over a payback window you could actually finance
- An agreed point at which you stop
