A report assigns SEK 50,000 in purchases to a ChatGPT campaign. The statement that the campaign had SEK 50,000 of attributed purchase value can be supported by that report, provided its definitions are clear. The statement that the campaign created SEK 50,000 of sales that would otherwise not have happened requires different evidence.
The distinction matters when deciding whether to expand, retain or remove advertising. Attribution connects observed outcomes to eligible interactions under reporting rules. Incrementality asks how outcomes differ from a credible estimate of what would have happened without the advertising intervention.
Write the claim before choosing the chart
Ask what the stakeholder needs to know. Campaign monitoring might require the number of reported purchases after eligible ad interactions. A finance decision might require the additional contribution generated by spending more. A channel comparison might require whether moving budget improves total business results.
These questions can use some of the same data but do not have the same evidence requirements. If the question concerns additional sales, a polished attribution chart does not become a causal estimate simply because it is presented to finance.
OpenAI reporting documentation defines attributed outcomes and reporting choices. Use those definitions when describing the platform figures. The experimental reasoning in this guide is an internal decision method and does not imply that the platform provides a native lift-testing tool for every account.
Understand the missing comparison
For each exposed customer, the business observes what happened under the actual journey. It does not also observe that same customer’s simultaneous journey without the advertisement. Causal analysis needs a defensible comparison that approximates this missing alternative.
In a hypothetical example, twenty reported purchases follow eligible ChatGPT ad interactions. Some customers might have purchased anyway, some may have purchased sooner because of the ad, and some may represent genuinely additional demand. The attribution total alone does not identify the sizes of those groups.
Equally, a channel might influence later activity that receives credit elsewhere. That possibility is not proof of hidden value. Treat both cannibalization and assistance as hypotheses to investigate rather than stories selected to defend a preferred budget.
Place common evidence in the right category
A before-and-after rise in total sales establishes that sales changed during the period. It does not isolate the campaign when promotions, seasonality, distribution or other channels changed too. A correlation between daily spend and orders can also reflect the team increasing spend on days it already expects stronger demand.
A customer survey may reveal remembered touchpoints, but memory and sample selection limit what it can establish about causal contribution. A source parameter can improve journey tracing, but does not reveal the no-ad outcome. Each piece can inform the investigation without being promoted beyond its scope.
Use cross-channel overlap analysis when several reports claim the same order. Removing duplicate credit from a spreadsheet creates a cleaner allocation model, not necessarily an incremental-sales estimate.
Which conclusion does the evidence support?
- Reported link
20 purchases were attributed under the stated reporting rules.
- Business movement
Total store sales rose at the same time; the cause remains unresolved.
- Incremental effect
A credible comparison estimates the difference versus no advertising.
Assess whether a credible intervention is feasible
A well-designed comparison can vary exposure or spending while protecting the interpretation of outcomes. The design must fit available controls, measurement, business constraints and sufficient information. Do not assume random user-level withholding is possible just because it would be statistically attractive.
The holdout feasibility guide starts with those practical constraints. Where a randomized design is feasible, preserve the assigned comparison and analyze the intended populations. Where only an observational design is possible, make its identifying assumptions and plausible alternative explanations explicit.
The outcome should match the business question. If the decision is whether advertising produces more new customers, measuring only attributed platform conversions builds the attribution mechanism into the outcome. Consider an appropriate independent business measure and the delay needed for it to mature.
Carry uncertainty into the decision
An estimated incremental effect is still an estimate. Report its uncertainty, the tested population and the period or spending level to which it applies. A positive point estimate with a wide range may support further testing while remaining insufficient for a large budget increase.
Convert the effect into economics carefully. Additional revenue is not additional profit, and the cost of generating it includes the relevant advertising and operating costs. State which costs are included rather than letting the word return imply a complete financial calculation.
Use a test hypothesis register to define the decision before results arrive. This reduces the temptation to change the business claim after seeing whichever metric looks favorable.
For ordinary reporting, a precise sentence is often enough: the campaign received a stated amount of attributed purchase value under named windows and time basis; incremental impact has not been established by that report. This preserves the information the report genuinely provides and identifies the next evidence needed. The advertiser can then make a proportionate decision using attribution for monitoring and a suitable comparison for claims about additional business.
Sources and scope
Match advertising claims to the evidence available and distinguish rule-based attribution from estimated causal incrementality.
Working methods and examples are editorial suggestions. Check current platform requirements and available features before implementation.