What to Measure in Your Game, and What to Skip
Measure only what could change a decision about your game: where players stop, lose, retry or go quiet. Start with about five events, each one tied to a question you really have.
Why start with a question and not with the data?
Because a number nobody acts on is decoration. Moss Harbour is an invented fishing game, and every number here is made up for this example. Its maker had one worry: do people get through the first day? So she wrote five questions, and one small record, called an event, for each. A session is one sitting, from opening the game to closing it.
| Question | Event | Result in 80 sessions |
|---|---|---|
| Do people start at all? | game-start | 80 |
| Do they catch a fish? | first-catch | 66 |
| Do they sail out? | island-reached | 52 |
| Do they find the map? | map-opened | 9 of those 52 |
| Do they finish the day? | day-end | 31 |
So 31 of 80 finish the first day, which is 39%. Is that bad? Only compared with a plan, so she wrote one first: about two thirds should finish. With only 80 sessions the true share is very likely between 29% and 50%. Even the top of that range misses her target, so this result can be trusted.
Now find where they go. From start to first catch, 83% carry on. From catch to island, 79%. From island to the end of the day, only 60%. The last step loses the most people. And just 9 of the 52 on the island opened the map, though the island quest needs it.
That gives her something to try: open the map by itself the first time a player lands. It is a guess the numbers point to, not a proven cause, so she will measure again afterwards.
She also counted fish caught in total: 1,204. A big, cheerful number that changes nothing. She deleted it.
Which numbers make a game better?
- Where people stop. See where players quit.
- Where they lose and try again. See the difficulty curve.
- What they use and ignore. See weapon and item usage.
- How money moves. See sources and sinks.
- How long things take. See pacing.
Skip anything about revenue, ad clicks or customer value. A business may want those, but they will not tell you whether level 3 is fun.
Keep it small. Game developer Chris Pruett began with just two events, deaths and level completions, and that alone showed him where his levels failed. He also noted the limit: a player stuck on a puzzle who quits without dying leaves no death to count.
Keep it anonymous too: no accounts, no device ids, no IP addresses. A random label for each sitting is enough, and our event log recipe works that way. The price is that one person playing three times looks like three sessions. Numbers also only tell you where to look. Why it happens is what you find out by watching someone play.
Paste this to your AI
I want to know what to measure in my game, and nothing beyond that.
My biggest worry: [for example: people stop before the first boss].
1. Read the game and draw what a player is supposed to do in the first [10] minutes as a Mermaid state diagram. Show it to me and wait until I say it is right.
2. From that drawing, suggest at most five events. For each, name the question it answers and what I would change in the game after a bad result. Drop any event where you cannot name a change.
3. Write down the numbers you expect if the game works as drawn, such as "two thirds finish day one". Those are my targets, set before any data exists.
4. Add nothing about accounts, device ids, IP addresses, revenue or marketing.
Once real data arrives, compare it with the drawing and the targets, and show me the counts behind each conclusion. If fewer than about 30 sessions reached a point, tell me the data is too thin there. Do not guess.
Or skip this and connect the MCP: your AI then finds all of this by itself.Connect the MCP
How to check it worked
- Finish the sentence. For each event, say "if this is low, I will change ...". If you cannot, delete the event.
- Count your events. More than five? Cut until it hurts.
- Look for the targets. Each number should have one, written before you saw any data.
- Play once and read the list. Everything you did on purpose should be there, once.
- Try to spot a person. Nothing in the list should tell you who played.