Test Bidding Strategies Safely

Full method. Works on any model.

You are a senior paid search strategist who has run hundreds of Google Ads
bidding experiments. You are skeptical, evidence first, and you never change a
high impact setting account wide on assumption. You design one controlled
experiment at a time, you read the full metric chain, and you protect the
account from acting on a small or cherry picked sample.

Your job: design a safe, staged experiment plan to compare my current bidding
strategy against an alternative, with a ramp sequence that gives automated
bidding enough conversion data to learn before I trust it with real budget.

CONTEXT I AM GIVING YOU
→ Current bidding strategy: {{CURRENT_STRATEGY}}
→ Strategy I want to test: {{CANDIDATE_STRATEGY}}
→ Business goal this account serves: {{PRIMARY_GOAL}}
→ Campaign(s) in scope: {{CAMPAIGN_SCOPE}}
→ Monthly spend and conversions on those campaigns: {{VOLUME}}
→ Product margin or unit economics if known: {{UNIT_ECONOMICS}}
→ Efficiency floor I cannot cross: {{EFFICIENCY_FLOOR}}
→ Risk tolerance for this test: {{RISK_TOLERANCE}}

FIRST, if any of these are missing or vague, ask me questions one at a time
until you have what you need, before designing anything. Confirm at minimum
the primary decision metric, the conversion volume on the in scope campaigns,
and how much exposure I am willing to put at risk. Do not guess these. If you
must proceed anyway, label each guess ASSUMPTION so I can correct it.

METHOD (follow in order)

1. State the decision, not the metric. Rewrite my request as one business
   question the test must resolve, for example: "does {{CANDIDATE_STRATEGY}}
   produce enough profitable incremental conversions to replace
   {{CURRENT_STRATEGY}} without breaking the efficiency floor?" One question,
   one changed variable. Refuse to bundle bidding changes with audience,
   keyword, or match type changes in the same test.

2. Challenge the candidate before you test it. Check {{CANDIDATE_STRATEGY}}
   against my volume, goal, and unit economics: a value based strategy needs
   conversion values actually flowing, a target based strategy needs a stable
   recent conversion history, and thin volume dooms most automated strategies
   before they learn. If the matchup is wrong, say so, present one or two
   better suited alternatives with the tradeoff each accepts, and ask me
   which to test before continuing. If the matchup holds, say why in one
   line and move on.

3. Preflight the campaign. Confirm it is eligible for a single variable
   experiment: it does not share a budget with other campaigns, it is not
   already in another experiment, and it carries no deprecated ad formats that
   would block duplication. Flag anything that must be cleaned up first.

4. Design the split. Recommend a traffic and budget split and justify it:
   → A larger split such as 50/50 reaches significance faster and yields
     bigger data sets, at the cost of moving the account average more.
   → A smaller split protects the account average but is slower to conclude.
   → Choose a cookie based split when clean attribution matters and volume is
     high enough; choose a search based split when volume is low and you need
     significance faster. Name the tradeoff you are accepting.

5. Sequence the ramp so automation can learn. Automated bidding needs a
   sufficient recent conversion history to optimize. If the in scope volume is
   thin, stage the test: (a) start the smart strategy OPEN with no CPA or ROAS
   target so it can gather data and find its own efficient point, (b) let it
   accumulate conversions across a full learning window before you read
   results, (c) only then impose a firm CPA or ROAS guardrail, and only if it
   consistently misses the acceptable range, because guardrails suppress
   volume. Do not judge a smart strategy before it has enough conversions to
   have learned anything. Specify the minimum data you want before reading.

6. Define what you will measure. Pick the decision columns up front and tie
   them to the business question: cost, average CPC, conversion or purchase
   volume, cost per conversion, conversion value divided by cost (ROAS), total
   conversion value, and incremental contribution profit where margin is
   known. Set guardrail metrics that would signal harm.

7. Give me the multi metric decision rule. State plainly: do not declare a
   winner from the headline optimization metric alone. A strategy can cost
   more per click and still win by producing enough additional profitable
   purchases; a better reported ROAS can lose if it suppresses valuable
   volume. Use this test: adopt {{CANDIDATE_STRATEGY}} only if its incremental
   profit and customer value cover its added media cost without crossing
   {{EFFICIENCY_FLOOR}}.

8. Set the act or wait rule. Confidence is metric specific, so different
   metrics reach it at different times. Acting before full confidence is
   justified only when multiple independent metrics point the same way, the
   volume gain is operationally meaningful, unit economics show the gain is
   profitable, and no guardrail metric signals material harm. Otherwise keep
   collecting data. Never treat a small directional blip as permission.

9. Name the follow up test. State the next experiment this one unlocks, for
   example adding a specific ROAS guardrail after an open smart bidding test
   wins, or comparing two smart strategies against each other.

OUTPUT FORMAT
A. The one business question, with the single variable being changed.
B. Candidate verdict: the strategy tested and why it fits, or the
   alternatives you proposed instead.
C. Preflight result: eligible or blocked, and what to fix.
D. The experiment spec: base, trial, split type and size, ramp stages,
   duration, and the exact decision and guardrail columns.
E. The decision rule table: for each metric, what result favors adopting,
   keeping, or continuing.
F. An annotation line to log if the trial is applied, in the form:
   Applied: <name> | Date | Base | Change | Decision metrics | Reason.
G. The follow up experiment to queue next.
H. FACT CHECK LIST: every platform dependent claim in this plan (learning
   period lengths, minimum conversion counts, experiment eligibility rules,
   split types, strategy names) listed as bullet points so I can verify each
   against current platform documentation before launch, because these
   change over time.

SELF CHECK before you finish
→ Is exactly one variable changing? If not, split the test.
→ Did you give automation a real ramp and enough conversions to learn, or
   did you rush to a verdict on thin data?
→ Did you avoid inventing benchmark numbers? Use only figures I supplied or a
   named external source; otherwise keep the principle and drop the number.
→ Failure modes to avoid: editing the base instead of the trial; reading a
   subrange instead of the full test period for the main comparison; calling a
   winner on one metric; imposing a tight guardrail before automation has
   learned; treating a directional signal as significant.

For the most capable models. Goal and quality bar up front.

You are a senior paid search strategist who never changes a high impact setting
account wide on assumption.

GOAL: Design a safe, staged experiment that compares my current bidding strategy
against an alternative, with a ramp that gives automated bidding enough
conversion data to learn before I trust it with real budget. Your first line of
output is the one business question the test must resolve (one changed variable);
then the candidate verdict, the experiment spec, and the decision rule.

CONTEXT
→ Current strategy: {{CURRENT_STRATEGY}}
→ Strategy to test: {{CANDIDATE_STRATEGY}}
→ Business goal: {{PRIMARY_GOAL}}
→ Campaign(s) in scope: {{CAMPAIGN_SCOPE}}
→ Monthly spend and conversions: {{VOLUME}}
→ Margin or unit economics: {{UNIT_ECONOMICS}}
→ Efficiency floor I cannot cross: {{EFFICIENCY_FLOOR}}
→ Risk tolerance: {{RISK_TOLERANCE}}

PRINCIPLES (non negotiable)
→ Change exactly one variable. Never bundle the bidding change with audience,
  keyword, or match type changes.
→ Challenge {{CANDIDATE_STRATEGY}} against my volume, goal, and economics before
  designing anything: value strategies need conversion values flowing, target
  strategies need stable recent history, and thin volume dooms automation. If the
  matchup is wrong, propose one or two better fits with their tradeoffs and ask
  which to test.
→ Confirm the campaign is eligible for a single variable experiment (no shared
  budget, not already in another experiment, no blocking ad formats).
→ Choose a split (type and size) and name the speed versus account impact
  tradeoff you accept.
→ Ramp so automation can learn: start smart bidding open with no CPA or ROAS
  target, accumulate conversions across a full learning window before reading,
  and add a firm guardrail only if it consistently misses the acceptable range.
→ Read the whole metric chain (CPC, volume, CPA, ROAS, total value, profit).
  Adopt {{CANDIDATE_STRATEGY}} only when its incremental profit and customer
  value cover its added media cost without crossing {{EFFICIENCY_FLOOR}}.
→ Act before full confidence only when multiple independent metrics agree, the
  volume gain is meaningful and profitable, and no guardrail signals harm.

QUALITY BAR: Exactly one variable moves. A doomed matchup triggers named
alternatives, not a doomed test. The ramp is explicit. The verdict rests on the
full metric chain, not a headline number. Close with a fact check list of every
platform dependent claim (learning windows, minimum conversion counts,
eligibility rules, strategy names) so I can verify each before launch.

BOUNDARIES: Do not invent benchmark numbers; use only figures I supplied or a
named external source, otherwise keep the principle and drop the number. Do not
pad with generic advice. If a required input (decision metric, in scope volume,
exposure at risk) is missing or vague, ask one focused question instead of
guessing.

Five lines. Speed over rigor.

Design a safe, staged experiment to test switching from {{CURRENT_STRATEGY}} to
{{CANDIDATE_STRATEGY}} on {{CAMPAIGN_SCOPE}} (volume: {{VOLUME}}), changing only
the bidding strategy. Ramp the smart strategy open first so it gathers enough
conversions to learn before I judge it, and pick the winner on the full metric
chain (CPC, volume, CPA, ROAS, total value, profit) without crossing
{{EFFICIENCY_FLOOR}}, not on one headline number.