---
name: migrate-from-manual-to-smart-bidding
description: >-
  Moves a campaign off Manual CPC or Maximize Clicks onto automated conversion/value bidding
  without losing the performance the manual setup earned. The core of the skill is risk control: a
  documented pre-switch baseline, a migration method matched to data strength (direct switch at 50+
  conversions/month, a 50/50 experiment for high-stakes proof, or a staged path through untargeted
  Maximize Conversions), pre-agreed rollback criteria, and a hands-off learning window before any
  judgment. Use it whenever a manually-bid campaign has accumulated enough conversions to automate.
  Not for choosing the destination strategy (select-a-bidding-strategy), fresh campaigns starting
  on Smart Bidding (set-up-conversion-based-bidding / set-up-value-based-bidding), the target math
  (calculate-bid-targets), or post-migration growth (scale-bids-and-budgets).
---
# Migrate from Manual CPC to Smart Bidding

## Purpose
The dangerous part of a bidding migration is not the click that changes the strategy, it is the
weeks around it: judging volatile learning data as failure, missing a genuine breakdown because no
rollback line was drawn, or being unable to prove anything because nobody saved the "before"
numbers. This skill wraps the switch in a baseline, a method choice, rollback criteria, and a
disciplined evaluation so the transition is measurable and reversible.

## When to run
- A Manual CPC / Maximize Clicks campaign has 15+ conversions/month (50+ makes it comfortable) and
  the decision to automate is made.
- The account's data-gathering exit trigger (documented in select-a-bidding-strategy) has fired.
- Manual bid maintenance cost has outgrown its value and conversion signal now supports automation.

## When NOT to run
- Destination strategy undecided -> select-a-bidding-strategy.
- Building a brand-new campaign directly on Smart Bidding -> set-up-conversion-based-bidding or
  set-up-value-based-bidding.
- No validated CPA/ROAS number -> calculate-bid-targets.
- Several thin campaigns would each migrate weakly -> set-up-portfolio-bid-strategies to pool
  first, then migrate the pool.
- The campaign is unprofitable at breakeven, automation optimizes toward the stated target, it
  does not repair broken economics; fix efficiency first.

## Prerequisites
- 4 weeks of clean performance history (no anomalous events inside the window).
- Verified, stable conversion tracking; known conversion lag.
- Destination strategy + initial target (from select-a-bidding-strategy / calculate-bid-targets).
- Stakeholder aligned on the plan and timeline (learning volatility is coming).

## Procedure
1. **Session guardrail.** Call `gads_policy_guardrail` before the session's first write.
2. **Freeze the baseline.** Pull the last 4 weeks via `gads_get_account_performance` and
   `gads_run_gaql_query`: conversions (must be 15+/month; 50+ is the comfortable bar), average
   CPA, average ROAS, conversion value, average CPC, CTR, impression share, IS lost to rank and to
   budget. Also read the conversion-lag figure. Save the snapshot; it is the yardstick for step 9.
3. **Pick the method** (Decision rules): direct switch, 50/50 experiment, or staged. Confirm with
   the user.
4. **Set the initial target generously.** tCPA: trailing average CPA plus a 5-10% buffer. tROAS:
   trailing average ROAS minus a 5-10% buffer. The first weeks are for learning, not for hitting
   the dream number, tighten later in 10-15% steps.
5. **Write down rollback criteria BEFORE switching.** Defaults: (a) daily cost above ~2x the
   historical daily average for 3+ consecutive days -> reduce the budget via `gads_update_budget`
   (do not revert the strategy for a cost spike); (b) post-learning CPA sustained above ~2x target
   -> tighten target or revert; (c) conversions collapse to ~zero -> check tracking first, revert
   only if tracking is intact. Reverting mid-learning for ordinary volatility is explicitly out of
   bounds.
6. **Execute, direct switch (write).** `gads_set_campaign_bidding_strategy` to the destination
   strategy with the step 4 target. Preview (validate_only=true default) showing old strategy, new
   strategy, target, and rollback criteria; apply only after user approval. Verify via
   `gads_get_campaign` that the new strategy is live.
7. **Execute, experiment (write).** `gads_create_experiment` with bid strategy as the only
   changed variable, 50/50 split; `gads_schedule_experiment` for 30+ days. Preview and get
   approval for each write. Freeze both arms, any other change contaminates the comparison.
8. **Execute, staged (write).** First switch to Maximize Conversions with no target
   (`gads_set_campaign_bidding_strategy`, previewed + approved). After 2-4 weeks, read the achieved
   average CPA via `gads_run_gaql_query` and add the target at that level (second previewed +
   approved write). This path suits 15-50 conversions/month or low risk tolerance.
9. **HUMAN STEP (outside VigilDog).** Send the stakeholder brief: what changed and why, two
   conversion cycles of expected volatility, the evaluation date (learning end + one conversion
   cycle), and the explicit ask for no panic-reverts during learning.
10. **Learning window (read-only).** Daily anomaly scan only: `gads_get_campaign` for spend and
    status, `gads_run_gaql_query` to confirm conversions still record. No target changes, no
    budget moves >10%, no keyword edits.
11. **Evaluate.** After learning (two conversion cycles) plus one more cycle, pull performance
    excluding the learning window and trailing conversion-lag days. Compare to the step 2
    baseline: CPA and volume within ~20% (or better) means a healthy transition. Slightly below ->
    wait one more cycle. Substantially below -> run the monitor-and-maintain-bid-strategy-health
    diagnostic before touching anything.
12. **Close out.** Experiment path: winner by conversions (volume strategies) or conversion value
    (value strategies), learning + lag excluded, significance ~80%+ -> `gads_promote_experiment`
    if the new strategy won, `gads_end_experiment` if the original held (each previewed +
    approved). All paths: once stable, begin tightening toward the calculated target in 10-15%
    increments, one conversion cycle apart, every change previewed and approved.

## Decision rules
- **Method by data strength:** 50+ conversions/month, stable history, familiar queries -> direct
  switch. High-stakes campaign or proof required -> 50/50 experiment (30+ days). 15-50
  conversions/month or extra caution -> staged (Maximize Conversions first, add target from
  achieved averages).
- **Initial target buffer:** tCPA = average + 5-10%; tROAS = average - 5-10%. Never launch on the
  calculated ideal, that is the destination of the tightening schedule, not the starting point.
- **Learning etiquette:** ~two conversion cycles hands-off; a mid-learning change resets the
  clock. The only sanctioned interventions are tracking fixes and the budget-cap response in the
  rollback criteria.
- **Evaluation window:** minimum one learning period + two conversion cycles (typically 3-6 weeks
  total) before declaring success or failure; always exclude learning + conversion-lag days.
- **Success bar:** post-switch CPA/ROAS and volume within ~20% of baseline is acceptable (the
  payoff compounds after tightening); consistent improvement is a bonus, not the bar.
- **One variable.** The migration changes bidding only. Concurrent keyword, budget, or creative
  changes make the result unattributable.

## Common failure modes
- **Panic reverts during learning.** Volatile week-one CPAs are the algorithm exploring, not
  failing; the step 9 brief and the pre-agreed rollback lines are the antidote.
- **Launching at the ideal target.** A calculated-target launch strangles delivery and gets
  automation falsely convicted; buffer first, tighten after.
- **No baseline.** Without step 2 there is no honest before/after, only vibes.
- **Judging with lagged data.** The most recent conversion-lag days always undercount; include
  them and every migration "fails".
- **Bundled changes.** Restructuring keywords "while we're at it" destroys the experiment's
  interpretability and the direct switch's accountability.

## Related skills
- Before: select-a-bidding-strategy (destination), calculate-bid-targets (target),
  set-up-google-ads-conversion-tracking (signal integrity).
- Alternative: set-up-portfolio-bid-strategies (pool thin campaigns before migrating).
- After: monitor-and-maintain-bid-strategy-health (post-migration checkups),
  scale-bids-and-budgets (growth once stable), optimize-bid-modifiers (strip stale manual-era
  modifiers, Smart Bidding ignores all but device -100%).