An advertiser sees twelve sales in a campaign report. The tempting conclusion is that the campaign created twelve customers. That may be true, partly true or wrong. The report first needs to tell us what it counted.
Three different questions often get compressed into “performance”: did an action happen, did the attribution rule assign it to this campaign, and did the advertising cause an outcome that would otherwise not have happened?
Start with the receipt
A click event is evidence that a defined interaction was recorded. A purchase event is evidence that the purchase instrumentation fired. Both require basic checks: duplicate events, cancelled orders, bots, missing acknowledgements and inconsistent identifiers can change the totals.
Before debating a sophisticated attribution model, reconcile a small set of permitted test events across the publisher, destination and reporting system. Make sure “purchase” means the same thing in each place. Keep test traffic out of the campaign result.
Attribution is a rule
Imagine a fictional customer already comparing two running shoes. They see a sponsored card, visit the product page and buy later that week. A seven-day click window might assign the sale to that ad. A one-day window might not.
The underlying purchase has not changed. The reporting rule has.
Agree the window, event definitions and treatment of repeat visits before the campaign. Keep click-attributed and view-attributed results separate if both are reported. Do not add platform totals together as if each system had an exclusive claim on the customer.
| Question | What would help answer it? |
|---|---|
| Did the action happen? | Reconciled event and order records |
| Does our rule credit this campaign? | A documented window and matching method |
| Did the ad change the outcome? | A credible comparison with what happened without it |
Incrementality asks a harder question
Suppose an illustrative randomized test has 1,000 eligible people in each of two groups. The group offered an ad records 30 purchases; the control group records 25. The observed difference is five purchases, or half a percentage point. It is not automatically proof of lift: sampling uncertainty, assignment quality, exposure and contamination still matter.
The example also shows why thirty attributed purchases would be a different claim from five additional purchases. A customer can genuinely click an ad and still have purchased without it.
For a small pilot, purchase counts may be too sparse to make a reliable causal claim. You can still validate delivery, investigate relevance and observe downstream actions. Describe those findings for what they are rather than presenting a tiny test as settled return on investment.
Give the report a decision to support
If the question is whether to fix a destination, arrival and completion events may be enough to locate a problem. If the question is whether to move a substantial budget from another channel, stronger evidence is warranted.
Ask for the raw counts, the definition behind each metric and the comparison used. A single return-on-ad-spend figure conceals too much when the buying model, attribution rule and sales cycle are still being established.
For a first Jolts campaign, agree the counting rules and the business outcome before launch. That gives both teams a shared basis for deciding what to repeat, what to repair and what remains unknown.