The A/B Testing Machine Behind Candy Crush

King tests every Candy Crush level the way a performance marketer tests ad creatives: ship variants, measure conversions, cut the losers.

The data optimization behind it is why the game still prints money 13 years later.

Every Level Is an A/B Test

At King, a level doesn't ship because a designer thinks it feels right. It ships because the data says it performs.

Like a restaurant that doesn't keep a dish on the menu because the chef likes it. It stays because the table turns and margin say so.

Multiple versions of the same level go out to different player groups. King measures how many attempts it takes to clear, where players quit, and when they open their wallets. If a level is too easy, nobody pays. If it's too hard, players leave. The version that hits the sweet spot between engagement and spending survives. The rest get pruned.

This isn't how most games are designed. Most studios playtest with a few hundred people, tweak based on feel, and ship. King runs what amounts to a continuous multivariate test across millions of players. The levels that exist in your game right now earned their spot through performance metrics, not creative instinct.

It's closer to how a growth team optimizes a landing page than how a game studio builds content. Booking.com runs a similar playbook at $170B+ scale. The same way a bar owner notices which drink special empties the rail by 10 PM and runs it again next Friday. Except King does it across 200 million players at once.

The Behavioral Models Running Behind Every Move

Testing individual levels produces useful signals. But King doesn't stop at the level.

Player behavior data from millions of sessions feeds into machine learning models that track patterns across the entire game. These models learn how reward timing affects retention, how failure frequency influences spending, and how progression pacing determines whether someone plays for 2 weeks or 2 years.

It's the difference between a gym that tracks which classes retain members past January and one that just posts a schedule and hopes people show up.

The Guardian reported that King operates more like a behavioral science system than a traditional game studio. AP News documented the scale of experimentation: player behavior data, machine learning models, and large-scale testing all running simultaneously.

The result is a game that adapts based on what players actually do, not what designers assume they'll do. Tech metrics reshape behavior in ways designers can't predict. If players in a certain cohort start quitting at level 47, the system learns that and adjusts. Not next quarter. Continuously. The psychology behind lasting behavior change is baked into the system itself.

The Whale Problem Nobody Talks About

Once the system understands behavior at scale, it starts sorting players into categories.

Only 2-4% of Candy Crush players ever spend money. A much smaller group within that 2-4% accounts for most of the revenue. The industry calls them whales.

This creates a design problem that most people don't think about. The game has to serve two completely different audiences at the same time: the 96% who will never pay but keep the game alive through social sharing and engagement metrics, and the 4% (really the top fraction of that 4%) who generate almost all the revenue.

Every nightclub works like this. Hundreds of people on the dance floor create the energy. The 6 people ordering bottle service pay the rent.

Casual players get low friction. Easy levels, frequent rewards, a sense of progress. They're the retention layer. High-spend players encounter calibrated resistance: difficulty spikes placed at moments where the data shows they're most likely to pay rather than quit.

The same game. Two different experiences. Determined by how you behave.

Like a coffee shop that gives regulars a free pastry and charges the tourist $7 for the same one. Same counter, different transaction.

Monetization Is a Precision Instrument

Monetization in Candy Crush isn't random. It's not even primarily about the items in the store.

It's about timing.

The system knows which players are likely to spend. It knows which moments produce the highest conversion rates: after a string of failures, when a timer blocks progress, when a limited-time event creates urgency. Difficulty spikes at these exact points aren't design oversights. They're the product.

Think of it as friction engineering. The frustration a player feels at level 65 after failing 8 times in a row is a calculated output of a system that tested thousands of difficulty curves and identified this specific pattern as the one that converts frustration into revenue without causing enough pain to make the player quit. Slot machines use the same psychology.

The difference between Candy Crush and a game that just makes hard levels is precision. King knows who will pay, when they'll pay, and how much pressure to apply before they break.

The Game That Never Ships a Final Version

Every purchase, every quit, every skipped level becomes a data point that re-enters the optimization pipeline.

This is the feedback loop that makes everything compound. New levels ship continuously through what the industry calls Live Ops: ongoing content updates, rotating events, limited-time mechanics. Each release generates fresh behavioral data. That data feeds the ML models. The models update segmentation. Segmentation refines monetization timing. And the cycle repeats.

There is no final version of Candy Crush. The game you play today is measurably different from the one that existed 6 months ago. Not because King redesigned it. Because billions of player interactions taught the system what to change. Like a subway system that reroutes trains based on real-time ridership data instead of running the same schedule it printed 10 years ago.

This is what separates Candy Crush from games that had one good idea and slowly declined. The system doesn't just run. It learns. Every dollar spent and every player lost makes the next iteration slightly better at extracting revenue and preventing churn.

That's why a puzzle game about matching candy generates over $1 billion a year, more than a decade after launch. Not because it found a perfect formula once. Because it keeps finding a better one.

Top Tweets of the day

1/

Source: @tedescau

Love it. Smart ones are automating their whole job. Some are just quietly raking in money while their bosses don't know that their job has become just quality control. It simply means 10 good people or 1 expert can do the job of 100.

2/

Source: @jean__gatt

TIL TikTok blocked direct links long time ago. Whenever a hack appears, exploit it as much as you can and don't talk about it until you've run it dry.

Even people who shared AI Jailbreaks have stopped sharing them publicly.

3/

Source: @clairevo

Perfect app for female beauty influencers. Massive TAM.

Whenever a new image model launches, it creates multiple little applications/use-cases like this.

Rabbit Holes

1. B2B Cold Outbound: 5 Example Outbound Plays by Pain, Not Persona

2. Google Ads AI Agent: Complete 2025 Guide

3. AI Marketing Examples: 13 Times AI Actually Delivered

What’d ya think of today’s newsletter? Hit ‘reply’ and let me know.

Do me a favor and share it in your company's Slack #marketing channel.

First time? Subscribe.

Follow me on X.

More Startup Spells 🪄

1. ESPN's Last Dance Marketing Playbook During COVID (LINK)

2. Stealing Reddit Content That Goes Viral on X (Twitter) (LINK)

3. SlapMac Viral Marketing: How a Prank App Made $5,000 in One Week (LINK)

4. Neurogum's Cross-Channel Growth Loop: How TikTok Feeds Amazon, Google, and Retail (LINK)

5. Showtime's Churn Prevention Secret Tactic: 60-40-20 Discount Loophole (LINK)