# Evaluate and decide

Use this for running or finished tests, a decision checkpoint, or a performance
question. The result is a supported next action, not a metric dump or an automatic
declaration that a channel works. Reports can use this method without launching
or changing anything.

## Establish what happened

Read the original question, offer/materials, audience, spending limits and decision
criteria. Gather relevant channel and product evidence for comparable periods;
retain sources, dates and gaps in existing results. No previous results file is
required to investigate a campaign. Separate delivery, response and business
outcome; show counts behind rates and spend behind efficiency claims.
Use the actual run and its plan/terms, not just the provider campaign ID. When
counting new customers or purchases, deduplicate people and receipts; distinguish
first purchases from renewals, test payments, refunds and other non-target events.

Distinguish cumulative snapshots from separate-period totals: do not sum
overlapping counts or spend, or treat repeated snapshots of one experiment as
independent tests.

Check attribution gaps, conversion delay, timezones, unlike definitions and
unavailable data. Do not combine incomparable measurements or turn missing data
into zero. Aggregated results cannot identify a winner at an unobserved finer
level: campaign totals do not establish which ad worked. Match the signal to
the chosen outcome; engagement is not a substitute for visits or purchases.
Review, account readiness and delivery failures are not evidence of absent demand.

Exclude known-invalid measurements from conclusions and forecast baselines.
A tracking repair does not validate earlier faulty data; distinguish the verified
measurement period after the repair.

## Interpret the result

Use the relevant channel method to locate the observed bottleneck. Distinguish
too little exposure, irrelevant response, a broken visitor path and an offer
people can use but do not choose. A bottleneck is observed; its cause may remain
a hypothesis. Investigate proportionately before proposing the remedy.

Compare against the agreed purpose, not a new success threshold chosen after
seeing results. Consider sample size, buying cycle and changes to destination,
offer, audience or season. Small or confounded results can justify another bounded
test without proving a reusable strategy. Competitor outcomes and historical
analogs provide context, not this campaign's missing evidence.

Explain economics at the relevant level: measured spend and useful outcomes,
qualified leads and their buying delay, or customer contribution where known.
Do not substitute assumed lifetime value for evidence. Zero sales alone does
not show a country or product is wrong. Before proposing another geography,
check relevant exposure, the working path and the cost of adapting the offer.
A campaign conclusion is not automatically a verdict on the whole product.

## Recommend the next action

Recommend continuing to an agreed checkpoint, fixing a demonstrated problem,
testing an explained change, stopping, or cautiously expanding. State why, the
expected benefit, cost/risk and the evidence needed for the next decision.
Inconclusive can be the right answer, but not indefinite observation without a
decision point. Neither weak delivery nor a noisy good result warrants an
automatic bid or budget increase. Stopping spend and finishing evaluation are
separate: allow relevant conversion or buying delays to mature after a pause,
without treating the later observation as permission to restart delivery.

When the next action depends on required operating safeguards, use
[observation and continuation checks](run-test.md#arrange-observation) to establish
what can proceed; a performance assessment does not replace that verification.

Before expanding, consider consequential evidence about delivered value, quality,
refunds or repeat use through [product evaluation](../product/evaluate.md).
Sales alone do not establish useful outcomes. Use relevant available evidence,
not a mandatory full audit or retention waiting period for every product.

Make changes interpretable: retain what changed and when; separate hypotheses
where feasible. Do not erase a test's original criteria or relabel its historical
results to fit the new approach. Keep the conclusion, useful lessons and their
evidence limits in the existing test results. Distinguish proposed next work
from accepted decisions; update the plan and observation arrangement after
agreement, then use the channel skill for authorized operations. Keep a one-test
adjustment in its plan; update the playbook only when an ongoing method change
is adopted, not because one result looks successful.