What are Saturation Curves?
Saturation curves model a fundamental marketing truth: the first dollar you spend usually works harder than the millionth dollar.
Core Concept: As you increase marketing spend in a channel, each additional dollar typically generates less incremental return. This is called "diminishing marginal returns" or "saturation."
Real-world example: Your first $10,000 in Google Ads might reach eager buyers searching for your product. Your next $90,000 reaches people less interested, requiring more impressions to convert. The curve captures this relationship.
The Intuition: Quick initial gains that slow down rapidly. Like picking low-hanging fruit first.
Marginal Return (ROI per Dollar)
📊 Real Example - Search Ads:
A SaaS company's Google Ads performance:
- First $10K: 200 conversions ($50 CAC)
- Next $10K: 141 conversions ($71 CAC) - 30% less efficient
- Next $10K: 115 conversions ($87 CAC) - further decay
Why? The first budget captures high-intent "software solution" searches. Additional budget goes to broader, less-intent keywords.
✅ When to Use Root:
- Channels with limited high-quality inventory (Search, Shopping)
- Niche B2B markets with small TAM
- When you see rapid early saturation in tests
- Retargeting campaigns (audience exhausts quickly)
The Intuition: Needs minimum spend to be effective, then grows rapidly, then saturates. Like needing critical mass for awareness.
Marginal Return (ROI per Dollar)
📺 Real Example - TV Campaign:
A CPG brand's national TV campaign:
- $0-500K: Minimal impact (too few people see it)
- $500K-2M: Rapid growth (achieving frequency, word-of-mouth kicks in)
- $2M-5M: Slowing growth (reaching saturation)
- $5M+: Plateau (everyone who'd buy has seen it)
Why? TV needs frequency (3-7 exposures) to work. Below threshold spend = wasted money.
✅ When to Use Hill:
- Brand awareness campaigns requiring critical mass
- TV, Radio, OOH (traditional media)
- New product launches needing minimum visibility
- Markets where network effects matter
Parameter |
Low Value Effect |
High Value Effect |
k (half-saturation) |
Saturates at lower spend |
Can absorb more budget |
s (shape) |
Gradual transition |
Sharp S-curve |
The Intuition: Similar to Hill but with more symmetric growth and saturation phases. Models natural adoption curves.
Marginal Return (ROI per Dollar)
📱 Real Example - Social Media (TikTok):
Fashion brand's TikTok advertising:
- Month 1 ($50K): Slow start - algorithm learning, finding audience
- Month 2-3 ($150K): Explosive growth - viral content, algorithm optimized
- Month 4+ ($300K+): Plateau - audience saturated, creative fatigue
The symmetric S-curve captures both the learning phase and saturation phase equally well.
✅ When to Use Logistic:
- Social media with algorithmic learning
- Influencer campaigns (building then saturating audience)
- Content marketing (slow build, rapid growth, plateau)
- When both ramp-up and saturation are gradual
The Intuition: Similar to logistic but can handle negative values and is bounded between -1 and 1. Often rescaled for practical use.
Marginal Return (ROI per Dollar)
🎯 Real Example - Programmatic Display:
E-commerce programmatic campaign:
- Low spend: Poor performance (not enough data for optimization)
- Medium spend: Rapid improvement (algorithm learns, finds pockets)
- High spend: Diminishing returns (good inventory exhausted)
Tanh's smooth transitions model the algorithmic learning well.
✅ When to Use Tanh:
- Programmatic advertising with ML optimization
- When you need smooth, differentiable curves (for optimization)
- Channels with gradual learning and saturation
- When modeling symmetric growth patterns
The Intuition: Borrowed from enzyme kinetics in biochemistry. Shows rapid initial response that gradually approaches a maximum.
Marginal Return (ROI per Dollar)
📧 Real Example - Email Marketing:
B2B company's email campaign to 100K list:
- 1 email/month: 5% click rate
- 4 emails/month: 15% total clicks (not 20%)
- 8 emails/month: 18% total clicks (heavy saturation)
- 12+ emails/month: 19% clicks + unsubscribes spike
Classic Michaelis-Menten: approaches maximum engagement asymptotically.
✅ When to Use Michaelis-Menten:
- Email marketing (list fatigue)
- Affiliate/partnership channels (partner capacity limits)
- Direct mail (household saturation)
- Any channel with biological-like response limits
⚠️ Statistical Note: Michaelis-Menten is mathematically equivalent to Hill with s=1, but parameterized differently. Some teams prefer it for interpretability (Km directly tells you spend at 50% max).
Quick Comparison Guide
Root (Square Root)
- Simple, few parameters
- Works well for search/shopping
- Can't model threshold effects
- Always starts at zero
Use when: Immediate response channels
Hill
- Models threshold effects
- Flexible S-curve shape
- More parameters to estimate
- Can be unstable with little data
Use when: Brand/awareness campaigns
Logistic
- Symmetric S-curve
- Well-understood properties
- Less flexible than Hill
- Fixed symmetry assumption
Use when: Natural adoption curves
Tanh
- Smooth, differentiable
- Bounded output
- Requires rescaling
- Less interpretable params
Use when: Need smooth optimization
Michaelis-Menten
- Interpretable parameters
- Based on proven theory
- Only concave (no S-shape)
- Assumes specific mechanism
Use when: Capacity-limited channels
Practical Implementation Tips
For Beginners:
- Start with Root or Michaelis-Menten for most digital channels
- Use Hill only if you see threshold effects in your incrementality tests
- Let the data decide: Run models with different curves, validate with holdouts
- Don't overthink it: Wrong curve shape is usually less problematic than wrong spend levels
For Statisticians:
- Identifiability issues: With limited spend variation, complex curves (Hill) may not be identifiable
- Priors matter: In Bayesian frameworks, use informative priors on saturation parameters
- Consider transformations: Log-spend or spend^0.5 as input can improve convergence
- Test specification: Use cross-validation or information criteria to select curves
- Interaction effects: Saturation can vary by creative, season, or audience
⚠️ Common Pitfall: Don't use S-curves (Hill/Logistic) just because they're flexible. If you've never spent below the supposed "threshold," you can't estimate it reliably. The model might invent a threshold that doesn't exist.