Optimization — Dossier

Objective

Improve campaign performance against plan KPIs using automated rules and Optimization Agent proposals — within guardrails, with human approval when thresholds are exceeded.

Inputs

Live execution_manifest slice + active plan_version per track (track_id)
Master plan caps (cross-track budget envelope)
GA4 performance data (primary)
Platform metrics (secondary)
CRM downstream metrics (lagged conversions)
Guardrail configuration per tenant

Outputs

Applied optimization changes (budget, bid, audience, creative rotation)
Updated execution_manifest (live state per platform entity)
Append to optimization_log / change_set_id (linked to plan_version, never mutates approved plan JSON)
Pending approval items when over threshold
Optional: plan.revise.recommended when drift thresholds met
Events: optimization.applied, optimization.pending_approval, plan.revise.recommended

Optimization types

Type	Examples	Auto if within guardrail
Budget shift	Move 10% Search → Meta	Yes, if < cap
Bid strategy	Target CPA adjustment	Yes, within bounds
Audience	Exclude low-intent geo	Yes
Creative	Pause ad with CTR < floor	Yes
Structural	New campaign type	No — requires replan
Spend increase	+25% daily budget	No — human approve

Decision loop

After optimization: plan & manifest sync

Layer	Updated when	Purpose
`execution_manifest`	Every applied optimization	Source of truth for live campaigns
`optimization_log`	Every applied / approved change	Audit trail + input to plan revise
`MediaPlan` vN	Only on plan revise / replan approval	Client-approved contract for spend & strategy
Reports	Next scheduled run	Show changelog + drift vs `vN` until revise approved

Optimizations never silently rewrite the approved plan document. When live state diverges materially, the system recommends or drafts vN+1 so client reports and invoices reconcile to an approved baseline.

When optimization triggers plan revise

See Plan update — triggers. Optimization Agent computes drift score after each apply:

drift_score = max channel |manifest_budget[channel] - plan_budget[channel]| / plan_budget[channel]

If drift_score > tenant threshold → emit plan.revise.recommended with structured diff preview.

Frequency

Rule engine: hourly for spend pacing
Agent deep review: daily (configurable per tenant)
Intraday: emergency pause on tracking failure or spend anomaly

KPI examples

Vertical	Primary optimization metric
Health	GA4 `book_appointment` CPA
Schools	GA4 `sign_up` CPA
Tourism	Assisted conversions / ROAS proxy
Ecommerce	GA4 `purchase` ROAS

Safeguards

Never optimize on platform-reported conversions alone if GA4 diverges > tolerance
Freeze increases when CRM upload match rate drops
Respect plan budget flight — no overspend without approval

GA4-absent degraded mode

GA4 is the source of truth, but onboarding v1 makes the GA4 link optional (client must grant Admin). When GA4 is not connected for a tenant, optimization runs in an explicit degraded mode, never silently:

Condition	Behavior
GA4 connected	Full optimization on GA4 SoT (default)
GA4 not connected	Optimize on platform-reported + CRM signals only; flag `degraded_attribution` on the tenant; cap automated spend increases (tighter guardrails) until GA4 connected; surface a "Connect GA4 for full optimization" banner

For pilots, GA4 Admin grant is a soft gate — required to meet the Phase 1 "GA4 SoT reporting" exit criterion (see execution gameplan B5).

Human touchpoints

Proposals exceeding guardrail thresholds
Conflicting signals (GA4 up, CRM down)
Policy-sensitive creative performance issues