An automated recommendation moves budget towards a campaign with a higher reported return on ad spend. The dashboard improves. The business still needs to ask whether the additional spend generated additional demand, and whether the resulting orders were commercially worthwhile.
Return on ad spend, or ROAS, divides attributed revenue by advertising spend. It is useful for describing performance under a specified attribution system. It does not, by itself, identify what would have happened without the advertising.
Two profitable-looking campaigns can have different jobs
A branded campaign reaches people already searching for the company. A category campaign reaches people comparing unfamiliar options. Either may produce additional sales, but their reported returns do not answer the same strategic question. Shifting budget between them solely on attributed ROAS can change the mix of demand being captured.
Google's Conversion Lift documentation explicitly distinguishes attributed conversions from incremental conversions. The latter concerns conversions caused by advertising rather than those assigned by reporting rules. That distinction should survive any AI-generated account summary.
Read the economics of the actual order
In a hypothetical example, a campaign records A$20,000 in attributed revenue from A$5,000 in spend: a reported ROAS of four. That ratio says nothing about product cost, discounts, fulfilment, payment charges or returns. If the campaign shifts the order mix towards a less profitable product, the revenue ratio can improve while contribution deteriorates.
Use the organisation's agreed contribution definition and make the included costs visible. Avoid importing an industry-wide target that ignores the product range or repeat-purchase pattern. A subscription offer and a single-purchase accessory should not inherit the same economics merely because both run through the same ad account.
These are budgeting calculations for a declared scenario, not a forecast that a particular return will be achieved. The value lies in making the assumptions inspectable.
Before asking AI for the next budget recommendation, give it the commercial context.
Provide the system with campaign purpose, attribution settings, product margin assumptions and relevant commercial constraints. Ask it to separate an observed metric change from its proposed explanation. A recommendation based only on attributed revenue should say so.
The system can flag patterns worth investigating. It should not silently fill missing margin data or assume every returning customer would otherwise have disappeared.
Now consider a second hypothetical campaign with lower attributed ROAS but more first-time customers. It is tempting to declare that campaign strategically superior. That would also exceed the available evidence. New customers can have poor economics, and returning customers can still be influenced by advertising.
The comparison needs the question made explicit: are we trying to improve immediate contribution, reach genuinely additional customers or learn whether a particular audience can be acquired profitably? Each objective changes the evidence required. An AI recommendation that combines them into a single ranking may conceal the trade-off the operator needs to decide.
Testing the next increment of spend
If the question is whether extra media creates additional sales, an incrementality design may be appropriate where scale and platform capabilities permit. Google's geo-experiment guidance distinguishes several experiment objectives. Select the design around the actual budget decision and document spillovers, timing and other campaigns that complicate interpretation.
When a credible experiment is not feasible, report that limitation and use a combination of commercial data and cautious monitoring. A weak causal claim does not become stronger when produced automatically.
Frontier's paid media work brings campaign diagnosis, creative and offer decisions together. The companion article on planning around uncertain assumptions explains why committing budget in stages can preserve useful options while the evidence develops.