Which numbers measure difficulty?

Fail rate, attempts per completion and the share of players who stop at the level. Fail rate is fails divided by all finished attempts. Attempts per completion is the average number of tries before a win. Stop rate is the share of players who started the level and never started the next one.

The first two measure how hard the level is. The third measures whether players accept that. A hard level that players push through is fine. A level where players stop is a problem, even if its fail rate looks normal.

Difficulty numbers per level and how to compute them.
NumberFormulaWhat it tells you
Fail rateFails / (completes + fails)How often an attempt ends in a fail
Attempts per completionAttempts / completesHow many tries a win costs
Stop ratePlayers who never start the next level / players who started this oneWhether players give up here
Deaths per startDeaths / level startsChallenge pace inside a level

Sources: Manalens game metrics documentation; Game Developer, Closing the loop (2 February 2014).

How do fail rate and attempts relate?

If each attempt ends in a win or a fail, attempts per completion is 1 / (1 minus the fail rate). So 50% means 2 attempts, 75% means 4 and 80% means 5 (our arithmetic). SayGames gives the same two examples for 50% and 80%.

The average hides a long tail. SayGames notes that where the average is two starts, some players need 10 or 20 attempts. In Lily's Garden, a study of the first 500 levels found about 3.2 attempts per completion after the early levels, with hard levels above 7 on average and some players needing 30 or more.

So look at the share of players who need many attempts, not only the mean. The players in the tail are often the most loyal ones.

Fail rate and the average attempts per completion. Our arithmetic.
Fail rateAttempts per completion
20%1.25
50%2
67%About 3
75%4
80%5
90%10

Sources: SayGames (30 July 2026); Kristensen and Burelli (2024).

How do I find a difficulty spike?

Chart attempts per completion and stop rate by level, side by side. A spike is a level where both jump above the levels around it. A sudden fail rate jump with a normal stop rate is a hard level. A stop rate jump is a wall.

Shadow Blade, a platformer, computed deaths per level start for every level and expected a saw shape: rising within each chapter, then dropping a little at the start of the next. The real chart did not follow it. The team then counted unique players per level and found levels with high drop rates. Level 14 had a sudden high difficulty ratio and blocked progress, so they made it easier.

Sources: Game Developer, Closing the loop (2 February 2014).

What events should I send?

Send level_start, level_complete and level_fail, each with the level and the attempt number. Add score, time in the attempt and the cause of the fail as properties. Add the difficulty setting if your game has one.

In Manalens these names are reserved game events. The Game KPIs dashboard reads them and shows a levels table with players, starts, completes, fails and the fail rate, and a level funnel that you can edit for your own milestones.

Unity

Manalens.Track("level_fail", new Dictionary<string, object> { ["level"] = 7, ["attempt"] = 3, ["score"] = 420, ["cause"] = "timer", ["difficulty"] = "normal" });

Sources: Manalens game events and metrics documentation.

Is a player who stops at a hard level gone?

Not always. SayGames found that what looks like churn can be a pause: players stay on one level for several sessions or days before they beat it. So read stop rate after a wait, such as 7 days, and check if the player still opens the game.

Split the players at a hard level into three groups: passed, still trying and gone. Only the last group is lost. If most players at the spike are still trying, a hint or a checkpoint can be enough. If most are gone, tune the level.

Sources: SayGames (30 July 2026).

How does this work in a premium PC game without levels?

Use checkpoints, rooms or bosses as your levels. Send the same start, complete and fail events with the checkpoint as the level property. Deaths per checkpoint start then plays the role of the fail rate.

Keep optional hard content apart from the main path. In Celeste, 34.6% of players had climbed the mountain on 18 November 2026, while 8.9% had the achievement for the Farewell chapter, the hard post-game chapter. A low number there is by design. Tune the main path for most players, and let the optional content stay hard for the players who want it.

Sources: Steam Web API, Celeste global achievement percentages; Manalens game events documentation.

How do I test a tuning change?

Change one thing, ship it in a new build, and compare the same level across builds. Keep the build in each event, then filter by build. Compare attempts per completion, stop rate and D7 retention for players who reached the level in each build.

Make small changes. Socialpoint's match-3 work and SayGames both tune by adjusting existing levels rather than redesigning them. A smaller change is easier to read, and it keeps the level the designer meant.

Check that the fix does not just move the wall. If the next level now has the spike, the curve still has a wall.

  1. Pick the level with the highest stop rate.
  2. Read its fail causes and its death map.
  3. Change one element and raise the build number.
  4. Wait until enough players reach the level in the new build.
  5. Compare attempts, stop rate and D7 between builds.

Sources: SayGames (30 July 2026); Socialpoint Analytics (13 January 2025); Manalens app events documentation.

Where can I see fail rates by level and build?

Manalens is the analytics and growth engine built for games. Send level_start, level_complete and level_fail from the Unity or Godot SDK. The Game KPIs dashboard shows the levels table and the level funnel, and every report filters by build, so you can compare a tuning change in one click.

Sources: Manalens product documentation.

Frequently asked questions

What is a good fail rate for a level?

There is no single good number. A tutorial level should have a low fail rate, and a boss can have a high one. Compare each level with the levels around it. A level that is far above its neighbours, with a high stop rate too, is the one to tune first.

Should I add an easy mode instead of tuning levels?

An easy mode or an assist option helps players who want it, and it keeps the intended challenge for others. Track the setting as a property, so you can see how many players pick it and whether they finish. Still tune levels where players on normal settings stop.

How many players do I need before I tune a level?

Enough that one stubborn player does not drive the average. Count players, not attempts. For a premium game, a playtest of a few dozen players can show a wall. For smaller changes, wait until a few hundred players reach the level in each build before you compare.

Does making a game easier always improve retention?

No. SayGames notes that in some games a stronger spike creates a moment of challenge that makes play more exciting. Only a test tells you which way your game works. Compare D7 retention and completion between builds, not only the fail rate of the tuned level.

Sources and implementation