Keep an advertising change and experiment ledger to explain overlapping effects

Link intent, state and timing

Use one row for each meaningful intervention or external event. Fields include client and object, hypothesis, intended mechanism, planned start, verified effective time, affected population, measurement window, primary outcome, concurrent events and evidence references. Add status values such as planned, applied, verified, abandoned and inconclusive. A submitted proposal is not an applied intervention; a provider acceptance is not necessarily proof that the expected behavior has begun.

In a teaching example, Monday brings a landing-page rewrite, Tuesday a creative refresh and Wednesday an additional sales representative. Qualified-lead CPA falls over the following week. The ledger marks three interventions with overlapping windows and flags attribution as unresolved. The team can still describe the business outcome, but cannot assign all improvement to the creative. A future learning plan might isolate one factor or use a suitable comparison group; the ledger does not retroactively create that control.

AdAce Ads audit history and external account histories can provide evidence about recorded changes. The ledger adds the hypothesis and business context that a technical event often lacks. Include off-platform changes such as stock availability, prices and qualification rules. When those facts are unknown, mark them unknown rather than assume the account was the only changing system. A clean-looking campaign history can coexist with a large operational change elsewhere in the funnel.

Expose collisions between interventions

• Preserve both intent and observed state. If a requested budget changed twice before the measurement window, the experiment did not run exactly as designed. Record the deviation instead of editing the original hypothesis to match the outcome.
• Assign a collision flag to overlapping interventions on the same outcome. The flag does not ban necessary operational work; it tells the reviewer that separating effects will require additional evidence or a weaker conclusion.
• Maintain a distinct evidence field for every status transition. Planned, executed and verified are different claims. A named reviewer should be able to locate the supporting record without reading an entire team chat.

Maintain the learning record

1. Create the hypothesis row before the intervention. State the expected direction, the relevant business outcome and what would make the result uninterpretable. Avoid choosing the primary metric only after seeing which number improved.
2. Record a baseline extraction with dates, metric definitions and known incomplete data. Attach existing evidence rather than exposing raw personal records. Document which parts of the funnel are measured and which remain outside the account reports.
3. When an authorized change occurs, add its actual effective time and affected scope. Link to the execution or state evidence where available. If execution fails or is only partially verified, keep that status visible and adjust the learning window.
4. Review concurrent events before evaluating results. Check both account history and the team's business-event log. Mark overlaps, late conversion updates and definition changes, then decide whether the original comparison remains meaningful.
5. Close the row with an evidence-based conclusion: supports further testing, contradicts the mechanism, inconclusive or operationally stopped. Keep the original hypothesis and deviations. A useful experiment record remains informative even when the test does not produce a winner.

Documentation is not causal identification

Sources and further reading

• AdAce Ads audit log
• AdAce Ads Google change history
• AdAce Ads Meta activity history