Document Type: Framework
Status: Active
Version: v1.3
Authority: MWMS HeadOffice
Applies To: Ads Brain and Affiliate Brain experimentation systems
Parent: Ads Brain Canon
Last Reviewed: 2026-04-18
Purpose
This document defines the controlled experimentation framework for Google Ads within MWMS.
Experiments exist to:
• protect capital during optimisation
• test changes against a control campaign
• produce statistically reliable results
• prevent emotional campaign editing
• improve scaling confidence
• allow controlled interaction with platform automation
• ensure behavioural optimisation systems remain interpretable
No structural campaign change should be made without testing unless authorised by HeadOffice.
Modern advertising platforms optimise using behavioural and probabilistic signals.
Experiments ensure those optimisation processes remain governed and measurable.
Scope
This framework applies to:
• Google Ads experiments inside MWMS
• control-versus-variant campaign testing
• ad variation experiments
• audience expansion experiments
• bidding strategy experiments
• landing page experiments
• campaign architecture experiments
• automation interaction experiments
• risk-controlled optimisation decisions
• Google Ads experimentation inside Affiliate Brain Phase 4 Structured Testing
This document governs how Google Ads experiments should be structured and interpreted.
It does not govern:
• offer viability approval
• capital allocation approval
• final survivability authority
• Finance Brain override
• HeadOffice override authority
• general Ads Brain campaign execution outside the experiment framework
Those remain governed by Affiliate Brain, Finance Brain, HeadOffice, and related MWMS systems.
Definition / Rules
Linked Canon
• Affiliate Brain Canon
Experiment Principle
Every optimisation must answer one question:
Does the change outperform the control?
Control Campaign
↓
Experiment Variant
↓
Traffic Split
↓
Statistical Comparison
↓
Winner Applied
The control campaign must always be preserved until a winner is proven.
Platform behaviour must not be interpreted as proof without structured comparison.
Experiment Types
Google Ads supports multiple experiment structures.
Each is used for different optimisation layers.
Type 1 – Ad Variation Experiments
Used for testing ad-copy changes across multiple campaigns.
Typical tests include:
• headline variations
• CTA changes
• messaging shifts
• offer framing
• value positioning emphasis
Example:
“Start Free Trial”
vs
“Start Free Today”
Traffic is automatically split and performance compared.
Type 2 – Custom Experiments
Full campaign A/B testing.
Google duplicates the campaign.
Original Campaign
Experiment Campaign
Traffic is split between both versions.
Used for testing:
• landing pages
• targeting structures
• campaign architecture
• keyword grouping
• match-type logic
• bidding strategy structures
Custom experiments allow safe testing of structural changes without risking full campaign exposure.
Type 3 – Bidding Experiments
Used when testing automated bidding strategies.
Examples:
• Manual CPC vs Target CPA
• Target CPA vs Maximise Conversions
• Maximise Conversions vs Maximise Conversion Value
• Impression share bidding for brand defence
Bidding experiments help prevent sudden performance collapse caused by learning phases.
Guardrails should not be applied prematurely before baseline data is understood.
Type 4 – Audience Experiments
Used for testing audience signals.
Examples:
• campaign without audiences vs campaign with in-market audiences
• observation audiences vs targeted audiences
• audience-filtered broad match tests
Audience signals increasingly influence platform optimisation behaviour.
Audience experiments allow controlled evaluation of behavioural targeting signals.
Type 5 – Performance Max Experiments
Used to compare Performance Max against other campaign types.
Examples:
• Performance Max vs Shopping
• Performance Max vs Search
Experiments validate whether automation improves results or introduces signal distortion.
Audience Observation to Targeting Progression
Audience expansion must follow a structured progression.
Stage 1 — Observation
Apply audiences in observation mode.
Observation mode allows performance learning without restricting reach.
Observation does not change delivery behaviour but produces insight.
Stage 2 — Signal Evaluation
Identify audience segments with:
strong conversion performance
strong engagement performance
consistent behavioural alignment
Stage 3 — Targeted Experiment
Run experiment targeting selected audiences.
May include:
broad match testing
budget prioritisation
creative alignment
landing page relevance alignment
Audience filtering may stabilise expansion risk when match looseness increases.
MWMS Ad Testing Hierarchy
Testing should move from low-risk changes to high-impact changes.
Level 1 – Creative Tests
Ad Variation Experiments
Level 2 – Targeting Tests
Audience Experiments
Level 3 – Strategy Tests
Bidding Experiments
Level 4 – Campaign Architecture Tests
Custom Experiments
Level 5 – Channel Tests
Performance Max Experiments
Higher-risk changes should only occur once lower-level signal clarity is achieved.
Experiment Traffic Split Logic
Traffic splits must balance:
learning speed
risk exposure
statistical reliability
Recommended structures:
50 / 50 split
70 / 30 split
Traffic allocation methods may include:
search-based split
cookie-based split
Search-based split:
faster learning
possible user overlap
Cookie-based split:
cleaner separation
slower learning speed
Choice depends on:
traffic volume
statistical tolerance
interpretation requirements
Variable Isolation Rule
Test only one major variable at a time.
Examples of major variables:
bidding strategy
audience targeting logic
landing page structure
creative messaging
match-type expansion
Multiple simultaneous major changes destroy experiment interpretability.
Learning Phase Awareness
Platform optimisation includes learning phases.
Short-term performance volatility does not necessarily indicate structural performance differences.
Experiments must run long enough to:
stabilise platform learning
collect sufficient behavioural signals
avoid premature conclusions
Typical duration:
2-4 weeks
or until sufficient data volume exists.
Experiment Interpretation Rule
Experiment outcomes must be evaluated using multiple signals:
conversion rate
cost per conversion
return on ad spend
conversion volume
signal stability
behaviour consistency
Short-term performance spikes or drops may reflect:
learning phase instability
traffic mix shifts
audience variation
creative fatigue transitions
Interpretation must consider signal reliability.
Experiment Application Rule
Once experiment results demonstrate improvement:
changes may be:
applied to original campaign
launched as new campaign
Applying to original campaign preserves historical learning signals.
Launching new campaign provides cleaner change history.
Account notes should record experiment application to preserve traceability.
Role of Experiments in MWMS
Experiments operate inside:
Affiliate Brain – Phase 4 Structured Testing
Experiments provide:
• risk-controlled optimisation
• data-driven scaling decisions
• capital protection
• structured interaction with platform automation
Experiment outcomes feed back into Affiliate Brain evaluation systems.
Final Rule
No structural campaign change should bypass controlled experimentation unless explicitly authorised by HeadOffice.
Platform automation must remain subordinate to structured experimentation discipline.
Drift Protection
The system must prevent:
emotional editing of campaigns without a control condition
multiple major variables changing simultaneously
control campaigns being paused before a winner is proven
overlapping experiments contaminating the same variable
broad match expansion without query governance
audience targeting without observation stage
automation assumptions replacing experiment evidence
platform volatility being mistaken for structural improvement
Experiments must protect signal integrity before pursuing optimisation.
Architectural Intent
Ads Brain Google Ads Experimentation Framework exists to ensure Google Ads optimisation inside MWMS remains controlled, interpretable, and capital-aware.
Its role is to ensure campaign changes are tested against preserved controls, sequenced by risk level, and used to generate reliable evidence before structural adoption.
Change Log
Version: v1.3
Date: 2026-04-18
Author: MWMS HeadOffice
Change:
Added structured handling for modern platform behaviour including:
audience observation → targeting progression
experiment guardrail logic for behavioural optimisation systems
traffic split interpretation logic
learning phase awareness
broad match expansion control discipline
Preserved original control-vs-variant logic, experiment hierarchy, and MWMS structured testing alignment.
CHANGE IMPACT
Pages Created: None
Pages Updated:
Ads Brain Google Ads Experimentation Framework
Pages Deprecated: None
Registries Requiring Update:
MWMS Architecture Registry
Canon Version Update Required: No
Change Log Entry Required: Yes