Scaling plans usually describe what happens when a campaign appears successful. Fewer plans define what happens when the evidence weakens, the operating system breaks or the test cannot answer its original question. Without a stop-rule, teams can keep spending because stopping feels like admitting failure—or stop too early because normal volatility feels alarming.
A stop-rule is a pre-agreed condition that pauses, limits or ends an action. It is written before the result is known, when the team is less tempted to reinterpret the standard around a preferred outcome.
Four types of stop-rule
1. Safety stop
This protects customers, data, brand or compliance. Examples include broken consent behavior, exposed personal data, a misleading claim, an invalid checkout or a landing page that sends inquiries to no owner. A safety stop should operate immediately and should not depend on statistical significance.
2. Measurement stop
This activates when the test can no longer be interpreted. Examples include a primary conversion failure, a material definition change, mixed test traffic, an integration outage or concurrent changes to the control.
The correct action may be to pause and repair—not to label the campaign a loser.
3. Economic stop
This limits financial exposure. It can use cumulative spend, cost per validated outcome, contribution margin or another business-relevant condition. Account for normal delay: a long sales cycle should not be judged as if revenue appears on the day of the click.
4. Learning stop
This ends a test that cannot answer a useful question. Examples include insufficient eligible volume, repeated changes to the hypothesis, a result that does not distinguish the options, or a market condition that makes the original question irrelevant.
Write conditions, not feelings
“Pause if performance is bad” is not a rule. A useful stop-rule includes:
- the metric or event;
- the data source;
- the threshold or observable condition;
- the minimum observation requirement, if applicable;
- the action;
- the decision owner;
- the restart requirement;
- the exceptions that can override the rule.
For example:
Pause new spend if the validated inquiry event fails two consecutive controlled tests or if more than one production inquiry is confirmed missing from the CRM. The RevOps owner verifies the failure; the media owner pauses affected campaigns. Restart requires a passed end-to-end test and documented incident resolution.
This is stronger than a vague instruction to “watch tracking closely.”
Use a rule card
| Field | Example |
|---|---|
| Decision protected | Whether to increase weekly search budget |
| Primary outcome | Sales-accepted inquiries |
| Safety stop | Inquiry route or consent behavior fails |
| Measurement stop | Source linkage coverage falls below agreed level |
| Economic limit | Maximum approved test exposure before review |
| Learning stop | Test design is changed materially or volume remains inadequate |
| Pause owner | Named role with access to act |
| Restart evidence | Passed QA, reconciled data and renewed approval |
Do not misuse statistical rules
Not every stop requires an experiment winner. Safety and measurement failures can invalidate the process immediately. Performance decisions need more care. Repeatedly checking a noisy metric and stopping when it first crosses a line can bias the result.
For formal experiments, define the hypothesis, success measures and analysis plan before launch. Google Ads recommends testing one variable at a time and selecting the success metrics before the test begins. Platform experiment results can still show no clear winner when there is not enough evidence.
If the team lacks statistical expertise, use conservative fixed review points and distinguish “no evidence of a win” from “evidence of a loss.”
Build escalation into the rule
Not every threshold should trigger the same action:
- Observe: add a warning and inspect the cause.
- Limit: prevent further scale while the current test continues.
- Pause: stop the affected campaign or route.
- Terminate: close the test and preserve the result.
- Incident: activate a broader response for safety or system failures.
This ladder avoids the false choice between ignoring a warning and shutting down everything.
Review stops as evidence
After a rule triggers, record:
- the exact evidence and timestamp;
- whether the trigger was valid;
- the action taken and delay;
- the root cause;
- the exposure prevented or learning preserved;
- whether the threshold should change;
- the restart decision.
A high false-trigger rate creates fatigue. A rule that never triggers despite visible damage is too weak, poorly instrumented or not operationally owned.
Pre-scale stop-rule checklist
- Is the business decision explicit?
- Are safety, measurement, economic and learning stops considered?
- Does each rule name a source and observable condition?
- Can the responsible person actually pause the system?
- Are maturation delay and normal volatility handled?
- Is there a limit on financial exposure before review?
- Does the rule distinguish pause from final failure?
- Is restart evidence defined?
- Will the trigger and response be logged?
A stop-rule does not make growth work timid. It makes risk bounded and learning intentional. The team can scale with more confidence because it has already agreed what evidence would make continued action irresponsible.
