Skip to module content
🧠AI Academy
0/91
Sign in
Module 12 Β· ~8 min

Paid Media Automation: Read, Write, Retract

The three-layer model + the autonomy ladder to sell by rung.

Reading progress
0/8 Β· 0%

The big idea

πŸ’‘Key idea
Paid media automation has three layers, and most agencies only build one. READ: MCP-connected data pulls that always include prior-period comparisons. Agents can only reason about changes β€” not snapshots. A metric without its context is just a number. WRITE: Mutations executed through narrow, intent-named tools, each with a dry-run mode, a human confirmation step, and an audit-log append. The diff IS the approval UI. RETRACT: Conversion-quality feedback flowing back to the platforms. This means retraction uploads for junk leads, enhanced and offline conversions for real revenue events, and value rules that teach the algorithm what a good conversion actually looks like. The retract layer is doubly critical wherever broad-targeting modes run β€” PMax, Advantage+, Max-Conversions β€” because those systems actively expand toward whatever the worst-scoring 'conversions' look like. Without retraction, you train the algorithm on your own junk.
Quick check
1 question Β· instant feedback
0/1
  1. The three layers of paid-media automation are:

Numbers that matter

4
Rungs on the autonomy ladder β€” reporter, advisor, guarded operator, autonomous within policy.
6+
Standing loops worth productizing (wasted-spend, search-term mining, pacing, geo, creative fatigue, lead-quality).
Quick check
1 question Β· instant feedback
0/1
  1. Prior-period comparisons on pulls matter because:

Deep dive

3/3 open

Agents don't reason about numbers in isolation β€” they reason about change. A CPA of €42 means nothing without knowing whether it was €31 last week or €58.

The standard cadence is a daily shallow pull that checks for anomalies and posts a Slack brief. Weekly, a deeper pull generates a diagnostic narrative β€” what changed, at what granularity (keyword, ad group, geo, device), and why it probably happened. Monthly, a full audit.

Designing the pull skills to output period-over-period deltas, keyword-level granularity, geo splits, and Quality Score components is right because it gives the model the raw material it needs to reason β€” not just report.

Every mutation class gets its own narrowly-named tool β€” 'Wasted Spend Finder,' 'Search Term Mining,' 'the Audit' β€” each with an explicit spec of what upstream analysis it consumes before it's allowed to run.

Before any mutation executes, dry-run mode shows exactly what would change: which keywords would be added as negatives, which budgets would shift and by how much. The human confirms. The change executes. The audit log appends.

Every mutation class is separately permissible, separately loggable, and separately revocable. This is the template to follow for any write MCP β€” not just paid media.

Six standing loops are worth building as scheduled skills, each a separate retainer line item.

Wasted-spend (weekly): search terms β†’ LLM classification of irrelevant/duplicate/cannibalizing β†’ proposed negatives with cost attached β†’ dry-run β†’ one-click push.

Search-term mining: continuous expansion of keyword lists from actual search queries, with intent classification.

Pacing sentinel (daily): compares actual spend rate to target and flags under- or over-delivery before it becomes a monthly-budget problem.

Geo rebalancer (monthly): redistributes budget across geographies based on conversion efficiency data.

Creative fatigue watch (Meta-specific): monitors frequency and CTR decay per creative; when thresholds are crossed, auto-drafts a creative refresh brief for human review.

Lead-quality reconciler (weekly): pulls CRM dispositions, generates retraction uploads for junk leads, and produces a summary of what the algorithm now believes a good conversion looks like. This last one is a genuine differentiator β€” almost no agency ships it.

Quick check
1 question Β· instant feedback
0/1
  1. Rung 4 (autonomous within policy) requires:

How to run it

  1. Prior-period comparisons on every pull
    Agents need deltas to reason about changes.
  2. Dry-run + confirm + audit on every write
    The diff IS the approval UI.
  3. Retract layer for junk conversions
    Especially with broad targeting (PMax, Advantage+, Max-Conv).
  4. Autonomy-ladder placement in the SOW
    Sell by rung β€” most clients belong at 2–3.
Quick check
1 question Β· instant feedback
0/1
  1. The retract layer matters especially with:

In the field

πŸ”¬Worked example
Autonomy ladder β€” sell by rung: (1) Reporter (read+narrate; human reads brief). (2) Advisor (read+recommend with reasoning + projected impact; human approves in Slack). (3) Guarded operator (execute pre-approved CLASSES of change within budget/percentage caps β€” e.g., add negatives from mined list; Β±20% budget moves; human reviews audit log). (4) Autonomous within policy (Measureβ†’Predictβ†’Validateβ†’Optimizeβ†’Learn cycle; human sets policy, audits weekly). Most clients live at rung 2–3.
🚫When not to reach for it
Never let a rung-4 autonomous agent reallocate cross-channel budget on platform-self-attributed numbers. Rung 4 requires the measurement layer (Module 5's caveat); without it, the agent optimizes toward whichever platform lies loudest.
Quick check
1 question Β· instant feedback
0/1
  1. The mutation-tool template pattern is:

Pitfalls & takeaways

Failure modes

  • Rung-4 sold before the measurement layer exists.
  • Skipping the retract layer β€” broad targeting eats junk-lead lookalikes.
  • Snapshot reporting without prior-period comparison β€” agents can't reason about it.
  • Mutation tools without dry-run β€” the diff IS the approval UI.

Durable takeaways

  • Agents reason about changes, not snapshots β€” pull with prior-period comparison.
  • The mutation-tool template is dry-run + confirm + append-only log.
  • The retract layer is what stops broad targeting from eating junk.
  • Sell by rung on the autonomy ladder β€” most clients live at 2–3.

Do the work

πŸ‹οΈProve you learned it

Pick one standing loop (wasted-spend weekly). Implement: search-terms pull β†’ LLM classification of irrelevant/duplicate/cannibalizing queries with cost attached β†’ proposed negatives list β†’ dry-run diff β†’ one-click push β†’ audit log entry. Ship at rung 3 with human approve in Slack.

0 chars
πŸ“¦Artifact to produce
Autonomy-ladder placement in the SOW + standing loops (wasted-spend, search-term mining, pacing, geo, creative fatigue, lead-quality) as scheduled skills.

Sources

  • Β· GMS Field Manual Β§12 (Paid Media Automation)