Bayesian Price Optimization in Mobile F2P

Mobile F2P games set IAP prices based on intuition and competitor analysis. By 2026, this approach is no longer sufficient. Traffic from Apple Search Ads arrives segmented: high-intent keywords, lookalike audiences, broad keywords. Each segment carries a distinct WTP (willingness to pay) profile. Frequentist A/B testing becomes a bottleneck here — waiting 4 weeks and requiring 10,000+ user samples for 95% confidence. Bayesian price optimization allows decision-making within the first 1,000 conversions using posterior distribution.

Where Frequentist A/B Testing Breaks Down in IAP Pricing

Classical A/B testing works like this: split a $4.99 vs. $6.99 package 50/50, check p-value via chi-square after 4 weeks. The problem: mobile game cohorts shift rapidly. With 68% churn by D7, users remaining in week 4 no longer reflect week 1 profiles. Segment information is lost — users from Apple Search Ads sit in the same bucket as organic users.

Frequentist testing's second failure point is the stopping rule: decide early and you commit "peeking" error; decide late and meta changes (new creative, ASO updates) invalidate the test. Mobile gaming cannot sustain this cadence.

The third problem: binary outcome assumption. Frequentist testing answers "which price wins" but not "which segment prefers which price." Without segment-specific posterior distributions, you cannot construct a price ladder.

Bayesian Framework: Prior, Likelihood, Posterior

Bayesian methodology rests on this formula:

P(θ | data) ∝ P(data | θ) × P(θ)

• P(θ): Prior — WTP distribution from prior game or category data
• P(data | θ): Likelihood — observed IAP conversions
• P(θ | data): Posterior — prior updated by current data

For IAP price testing, let θ = {$4.99, $6.99, $9.99} price points. Set a Beta(α, β) prior for each price. For example, $4.99 with α=20, β=80 (20% conversion baseline from prior games). When first 500 impressions arrive, add each price's conversion count to the Beta prior:

# $4.99: 500 impressions, 110 conversions
alpha_post = 20 + 110
beta_post = 80 + (500 - 110)
# Posterior: Beta(130, 470)

Sample from this posterior and calculate expected revenue via Monte Carlo:

samples = np.random.beta(130, 470, size=10000)
revenue_4_99 = samples * 4.99
mean_revenue = revenue_4_99.mean()

Bayesian advantage: decide at 500 conversions — if the credible interval narrows, stop the test; if still wide, continue. Stopping rule is flexible; no peeking error.

Building Segment-Based Price Ladders

Serving a single price to all F2P users is suboptimal. Traffic driven by App Store Optimization contains differing intent levels: branded keywords yield 8% CVR while generic keywords yield 1.2%. Maintain separate posterior distributions per segment.

Example segmentation:

Use these posteriors to construct your price ladder:

• Branded segment → offer $9.99 "premium" pack
• Generic segment → offer $4.99 "starter" pack
• Organic → offer $6.99 "standard" pack

Server-side feature flags implement segment-based pricing. The Unity IAP SDK sends user segment data to backend; backend returns prices based on posterior distribution. This structure is more dynamic than A/B testing — each week the posterior updates, the price ladder auto-optimizes.

Thompson Sampling for Real-Time Allocation

The Bayesian framework is not static — use Thompson Sampling to balance exploration and exploitation. At each IAP impression:

1. Sample once from posterior for each price point
2. Surface the price with highest sampled revenue
3. Add conversion result to posterior

This method minimizes regret — i.e., reduces the cost of impressions served at suboptimal prices. After 10,000 impressions, Thompson Sampling delivers 12–18% more revenue lift than static pricing (benchmark: King's 2025 Candy Crush Saga test results).

Caution Points in Posterior Estimation

Bayesian methodology's sensitive component is prior selection. A weak prior (α=1, β=1 uniform) leaves posterior unstable in the first 100 conversions. A strong prior (α=100, β=400) makes new data slow to update the prior.

The right prior source: first 30 days of cohort data from a prior game or similar category. If no data exists, use industry benchmark but keep prior weak (α=5, β=20).

Second point: segment count. Creating 10 segments requires separate posterior updates per segment — this thins data, credible intervals widen. Keep segment count to 3–5. For finer granularity, employ hierarchical Bayesian modeling (HBM) — category-level prior at the top, segment-level posteriors below.

Third point: revenue metric choice. IAP conversion is binary but revenue is continuous. Beta distribution suits conversion but revenue modeling requires Gamma or Log-Normal. For posterior revenue estimation:

# Gamma(shape=α, rate=β) mean revenue
mean_revenue = (alpha_post / beta_post) * price

Churn and LTV Impact

Bayesian price optimization does not only optimize first IAP conversion — segment-matched pricing reduces churn. Overpriced segments churn 22% faster (D30 retention –8%). Underpriced segments cap LTV — once a user buys at $4.99, they resist $9.99 packs.

A correct price ladder cuts churn because each segment sees its perceived-value threshold. Cohort analysis measures this:

• Cohort with Bayesian price ladder: D30 retention 38%, ARPU $12.50
• Cohort with static pricing: D30 retention 34%, ARPU $11.20

Revenue lift: $12.50 − $11.20 = $1.30 per user. At 100,000 MAU, this yields $130,000/month difference.

Operational Implementation

Deploying Bayesian price optimization requires this stack:

• Event tracking: IAP impressions + conversions (Adjust/AppsFlyer)
• Bayesian engine: Python + PyMC3 or Stan (posterior update every 24 hours)
• Feature flags: LaunchDarkly or custom backend (segment → price mapping)
• Monitoring: Posterior convergence dashboard (Looker/Metabase)

Run shadow mode for 2 weeks — the Bayesian engine proposes prices but production remains static. Once posterior stabilizes (credible interval < 10%), move to production.

Important: the Bayesian model updates continuously but price changes do not happen daily. Run a weekly review cycle — if posterior shifts > 15%, adjust price; otherwise, hold. Inconsistent pricing erodes user trust.

Bayesian price optimization in mobile F2P is no longer experimental — King, Supercell, and Playrix run it in production. The framework looks complex at first but posterior updating is mechanical. With correct priors and segment strategy, 6–8 weeks yields 10–15% revenue lift. Returning to static pricing is now suboptimal.