Measuring fun

Core principle

Fun is a feeling, but it has causes you can test. Measure the causes in a simulation, check understanding with personas, measure the feeling through what real people do, and keep only the sim measures that predicted what people did. No single source signs off fun: a balanced sim is a precondition, a persona's "I'd come back" is a hypothesis, a retention chart is engagement.

What causes fun, and how sure we are

Name the evidence tier whenever a claim rests on one of these.

Engagement is not fun. Near misses raise the urge to keep playing while players rate them as unpleasant (Clark et al. 2009). Variable rewards, daily spins and loss-framed streaks raise opens, not enjoyment.

Steps

1. Write the fun promise and the guardrails before measuring. The promise: who plays, which feeling (mastery, discovery, rivalry, people), at which moments. The guardrails: what the game must never push people to do, such as skip sleep, play while driving, pay, or share where they live. Selling a way out of a loss (a paid streak freeze, a paid retry) is a guardrail failure. Write down each measure's range and the decision a miss triggers, before you look at any result.
2. Run the sim scorecard (table below) on real behaviour data, with a ladder of bots: a random bot, a greedy beginner who plays the best of what it got, a planner who goes for a goal, a bold player who takes calculated risks, and one restricted bot per honest play style ("only plays safe", "only explores").
   • Bend the replayed behaviour. Nobody in last month's data played the game. When a bot goes for something, it changes the activity within that person's own range (amount, time, place) with a set chance of success. That chance is a guess: write it down and test it with people.
   • Run the bad cases: one very active player, a player who stops, the cheapest exploit.
   • In team games, measure the team and the people in it. Add "carry": the share of a team's points made by its most active member. Most points from one person means the others are watching, not playing.
3. Use personas for understanding only. Use cil:ux-review for whether people find the game, understand the rules and read the copy right. A persona's feeling ("it never wins") is a hypothesis for step 2 and step 4, never evidence of fun.
4. Test with people, over the time the game really spans.
   • Compare two versions instead of rating one. Ask "which was more fun, A or B?"; ratings drift and bunch up (Yannakakis and Martínez 2015).
   • Take the decision away once. Each player plays one round where the app decides for them and one where they decide, in random order. When a round lasts a week, split the testers into two groups instead. If behaviour doesn't drop without the decision, the fun is elsewhere (the rival, the activity itself), and a simpler game would do.
   • Ask for a moment, not a score. "Tell me about that game" a few days later. A specific moment ("I called it and made it by two minutes") is a peak; "it was nice" is not.
   • Read the behaviour signals (table below) against a holdout, in week 3, not on day 1.
5. Calibrate the sim against people. For every scorecard measure, check whether it moved with the human results. Does "open into the last round" predict a rematch? Keep the measures that predicted, retire the rest, adjust the ranges, and write the calibration next to the scorecard.
6. Report: the scorecard with pass or fail and the feeling each miss stands for, what only people can settle, and one question.

Fun is not a launch gate. Block a launch only when a guardrail fails or the scorecard shows a broken game (choices are fake, volume decides, one piece wins everything). Otherwise ship it as a test: part of the users, a holdout, no streaks or spins, and the kill criteria from step 1. The fun verdict comes from week 3.

Sim scorecard

A "round" is whatever unit scores: a round, a day, a week. A "piece" is whatever players pick: a card, an option, a strategy, a map zone, a quest. In a solo game the goal is the rival: "win" means making it, the bot ladder compares going for it with not trying (if ordinary behaviour already makes most goals, the goal is fake), and "volume doesn't decide" compares the time to a reward for a three-times-a-week player and a daily one. Skip rows that don't exist in the game, such as collecting. Starting ranges come from one game and must be recalibrated in step 5.

A fair average can hide a missing peak: when a style wins half its games but never gets the big moment, the dominant-piece and peak measures show it, not the win rate.

Behaviour signals

Common mistakes