Knowledge Graphs as Control Planes: Strategic Research Report

Document: KNOWLEDGE_GRAPHS_AS_CONTROL_PLANES.md Version: 1.0.0 Date: January 30, 2026 Author: MIZ OKI Research Team Status: ✅ VALIDATED

Executive Summary

This report synthesizes enterprise case studies demonstrating how knowledge and data graphs serve as control planes that power cross-channel activation and drive measurable ROI lifts. Analysis of Criteo, Salesforce, and Zeta Global implementations reveals common architectural patterns that MIZ OKI has already implemented—with key differentiators that create competitive advantage.

Key Finding

"The competitive advantage is not the graph alone, but the graph as an active control plane—reading signals, reasoning over them, and writing back optimized decisions."

1. Enterprise Case Study Analysis

1.1 Criteo: Shopper Graph + Ghost-Bid Incrementality

Technical Architecture:

┌─────────────────────────────────────────────────────────────┐
│  CRITEO SHOPPER GRAPH                                       │
├─────────────────────────────────────────────────────────────┤
│  Identity Layer                                             │
│  ├─ Cross-device identity stitching                         │
│  ├─ Deterministic + probabilistic matching                  │
│  └─ Privacy-safe hashing (SHA256)                           │
├─────────────────────────────────────────────────────────────┤
│  Commerce Signal Layer                                      │
│  ├─ Product views, cart events, purchases                   │
│  ├─ Category affinity scoring                               │
│  └─ Recency-frequency-monetary (RFM) signals                │
├─────────────────────────────────────────────────────────────┤
│  Bidding Control Plane                                      │
│  ├─ Real-time bid computation per impression                │
│  ├─ Ghost-bid holdout (control group sees no ad)            │
│  └─ Incrementality = Treatment conversion - Ghost conversion│
└─────────────────────────────────────────────────────────────┘

Ghost-Bid Methodology: - Random 2-5% of eligible impressions receive "ghost bids" (no actual ad shown) - Control group tracked for organic conversions - Treatment effect = P(convert|ad) - P(convert|no ad) - Enables true causal ROAS measurement vs. platform-reported attribution

Why It Works: - Graph enables identity-centric targeting (same user across devices) - Ghost-bids provide unbiased baseline for incrementality - Real-time bidding optimizes based on measured causal lift

1.2 Salesforce: Data 360 + Agentforce

Technical Architecture:

┌─────────────────────────────────────────────────────────────┐
│  SALESFORCE DATA 360                                        │
├─────────────────────────────────────────────────────────────┤
│  Profile Unification Layer                                  │
│  ├─ 266M fragmented profiles → unified                      │
│  ├─ Identity resolution across CRM, marketing, commerce     │
│  └─ Real-time profile updates (streaming ingestion)         │
├─────────────────────────────────────────────────────────────┤
│  Intelligence Layer                                         │
│  ├─ Segmentation & audience building                        │
│  ├─ Propensity scoring (churn, upsell, conversion)          │
│  └─ Journey orchestration triggers                          │
├─────────────────────────────────────────────────────────────┤
│  Agentforce Control Plane                                   │
│  ├─ AI agents read from unified graph                       │
│  ├─ Contextual decision-making per interaction              │
│  ├─ Workflow automation (email, chat, sales handoff)        │
│  └─ ROI tracking back to graph (closed loop)                │
└─────────────────────────────────────────────────────────────┘

Key Innovation: Graph as Read/Write Memory - Agents read context from unified profile - Agents write interaction outcomes back to graph - Creates compounding intelligence over time

Why It Works: - Single source of truth eliminates data silos - Real-time profile enables contextual personalization - Agentic layer automates high-volume decisions

1.3 Zeta Global: Data Cloud + United Airlines

Technical Architecture:

┌─────────────────────────────────────────────────────────────┐
│  ZETA DATA CLOUD                                            │
├─────────────────────────────────────────────────────────────┤
│  Identity Graph Layer                                       │
│  ├─ 240M+ persistent identity profiles                      │
│  ├─ Cross-device, cross-channel identity resolution         │
│  └─ Privacy-compliant data partnerships                     │
├─────────────────────────────────────────────────────────────┤
│  Behavioral Signal Layer                                    │
│  ├─ Real-time intent signals                                │
│  ├─ Purchase propensity scoring                             │
│  └─ Engagement history (email, web, app)                    │
├─────────────────────────────────────────────────────────────┤
│  Orchestration Control Plane                                │
│  ├─ Athena AI: audience insights + recommendations          │
│  ├─ Advisor AI: campaign optimization                       │
│  ├─ Multi-channel execution (email, display, mobile)        │
│  └─ ROI attribution back to identity graph                  │
└─────────────────────────────────────────────────────────────┘

United Airlines Use Case: - MileagePlus member data enriched with Zeta behavioral signals - AI-driven offer personalization based on travel patterns - Cross-channel activation (email + display + in-app) - 10× ROI from targeted vs. broad campaigns

Why It Works: - Rich behavioral graph enables intent prediction - AI layer automates offer selection at scale - Cross-channel activation prevents siloed optimization

2. Common Architectural Patterns

2.1 Three-Layer Control Plane Architecture

All three enterprises share a common architectural pattern:

┌─────────────────────────────────────────────────────────────┐
│  LAYER 1: UNIFIED DATA GRAPH (Foundation)                   │
│  ├─ Identity resolution (privacy-safe)                      │
│  ├─ Entity relationships (user → product → campaign)        │
│  ├─ Behavioral signals (views, clicks, conversions)         │
│  └─ Temporal context (recency, frequency, seasonality)      │
├─────────────────────────────────────────────────────────────┤
│  LAYER 2: INTELLIGENCE LAYER (Reasoning)                    │
│  ├─ Propensity/affinity scoring                             │
│  ├─ Segmentation & audience building                        │
│  ├─ Causal inference (CATE, uplift modeling)                │
│  └─ Optimization (budget allocation, bid adjustments)       │
├─────────────────────────────────────────────────────────────┤
│  LAYER 3: ACTIVATION LAYER (Execution)                      │
│  ├─ Real-time bidding / ad serving                          │
│  ├─ Email/SMS/push orchestration                            │
│  ├─ Agentic workflow automation                             │
│  └─ Measurement & feedback loop                             │
└─────────────────────────────────────────────────────────────┘

2.2 Key Success Factors

2.3 Graph-as-Control-Plane Pattern

The defining characteristic of these implementations is bidirectional graph interaction:

       ┌──────────────────┐
       │   DATA GRAPH     │
       │  (Source of      │
       │   Truth)         │
       └────────┬─────────┘
                │
    ┌───────────┴───────────┐
    │                       │
    ▼                       ▲
┌───────────┐         ┌───────────┐
│   READ    │         │   WRITE   │
│ (Context) │         │ (Outcome) │
└─────┬─────┘         └─────┬─────┘
      │                     │
      ▼                     │
┌─────────────────────────────────┐
│     INTELLIGENCE LAYER          │
│  (Reasoning + Optimization)     │
└─────────────────────────────────┘
      │
      ▼
┌─────────────────────────────────┐
│     ACTIVATION LAYER            │
│  (Execution + Measurement)      │
└─────────────────────────────────┘

Critical Insight: The graph is not just storage—it's an active participant in the decision loop, receiving outcomes and informing future decisions.

3. MIZ OKI Alignment & Differentiation

3.1 Capability Mapping

3.2 MIZ OKI Differentiators

3.3 Gap Analysis

4. Ghost-Bid Incrementality: Technical Deep-Dive

4.1 Why Ghost-Bids Matter

Problem with Platform-Reported ROAS: - Platforms count all conversions after ad exposure - Ignores users who would have converted anyway ("sure things") - Overstates true advertising impact by 30-50%

Ghost-Bid Solution: - Randomly select 2-5% of eligible impressions - Don't show ad (ghost bid) - Track organic conversions in ghost group - True lift = Treatment conversions - Ghost conversions

4.2 Mathematical Foundation

Incremental Lift (τ) = E[Y|T=1] - E[Y|T=0]

Where:
  Y = conversion outcome (0 or 1)
  T = treatment assignment (1 = ad shown, 0 = ghost/no ad)
  E[Y|T=1] = conversion rate when ad shown
  E[Y|T=0] = conversion rate when no ad (organic baseline)

Incremental ROAS (iROAS) = (τ × value_per_conversion × N_treated) / ad_spend

Example:
  - Treatment CVR: 2.5%
  - Ghost CVR: 1.8%
  - Incremental lift: 0.7% (= 2.5% - 1.8%)
  - Platform-reported lift: 2.5% (overstated by 3.6×!)

4.3 Statistical Considerations

4.4 Integration with Existing Architecture

MIZ OKI already has the foundations:

1. Experiment Framework (experimentation_incrementality_integration.py) - Experiment creation and lifecycle management - Treatment assignment via memoized hash - Outcome tracking and analysis

2. Causal Estimation (uplift_pacing_integration.py) - CUPED variance reduction - Geo holdout support - Lift gate validation

3. Budget Reallocation (autonomous_budget_reallocation_mvp.py) - ReLU gating for confidence-based decisions - DSP client interfaces - Idempotent execution

Gap: No explicit ghost-bid experiment type or dual-budget tracking.

5. Strategic Recommendations

5.1 Immediate Actions (V6.14.0)

5.2 Medium-Term (Q2 2026)

5.3 Long-Term (2026+)

6. ROI Projections

Based on enterprise benchmarks:

Expected Impact (Conservative): - 5-12% ROAS improvement (from better targeting) - 8-20% CPA reduction (from waste elimination) - 2-6% wasted spend reduction (from ghost-bid insights)

7. Conclusion

The enterprise case studies validate MIZ OKI's architectural direction. The key differentiator is measured incrementality—knowing which ad dollars actually drive conversions vs. which are wasted on users who would convert anyway.

Strategic Priority: Implement ghost-bid holdout testing to achieve Criteo-level incrementality measurement while maintaining MIZ OKI's unique advantages in neuro-symbolic reasoning and multi-model orchestration.

Appendix A: Source Materials

Appendix B: Related MIZ OKI Documentation

Document generated by MIZ OKI Research Team • January 30, 2026