Platform-reported ROAS vs incremental ROAS: why the dashboard grades its own homework

ROAS is the conversion value attributed to ads divided by ad spend. In the platform, "attributed" means credited by the platform's own rules. Incremental ROAS (iROAS) is incremental revenue, meaning revenue caused by the ads, divided by ad spend. It rests on incrementality: the sales you would not have had without the ads.

Credit is a rule for dividing sales between touchpoints. It is not proof that an ad caused one. That is the whole difference, and it is why we write "platform-reported" with a hyphen: the number has an owner.

Three reasons, none of them sinister.

• The platform writes the rules. Attribution models are rules for assigning credit. A platform's rule can only see its own ads, so it credits them.
• Credit is not cause. Under last-click attribution, the last channel clicked gets everything. View-through conversions go further: they credit an ad that someone saw but did not click, within a set window.
• Some sales were coming anyway. A vendor explainer on platform versus incremental ROAS says the platform number can overstate channels such as brand search and retargeting¹. Those reach people who already know you.

None of this makes the platform useless. Dashboards are good at speed and at comparing ads inside one channel. They are poor at answering "what would have happened without this spend?".

Yes. And that is the part most "your ROAS is lying" posts skip.

The same vendor explainer says platform ROAS can understate other channels, such as YouTube, and sales that never involve a click or happen in physical shops¹. A vendor report on its own Meta experiments goes the same way. Haus, which sells incrementality tests, reports that across 640 Meta tests run on its platform, Meta's seven-day click reporting under-stated incrementality by about 15% on average for direct-to-consumer brands². Read that with care. It is vendor data, from brands that chose to run tests, and they spend far more on Meta than many brands do. The same report says Advantage+ campaigns over-reported relative to manual ones². Even inside one platform, the error changes direction.

So the direction of the error depends on the channel and the settings. It is not "platforms always over-report". Some channels look better in their own reports than they are, some look worse. That is why one correction factor for everything does not work, and why we check channel by channel.

An agency's published channel benchmark makes a similar point. It lists a different incrementality factor for each channel, derived from its own geo-holdout experiments and media mix models, and applies it as iROAS = platform ROAS × factor³. The incrementality factor is the ratio of incremental conversions in a test to the conversions the platform reported. The agency sells incrementality work and the data is its own. Do not copy its numbers into your plan.

For ZFD: a platform ROAS is a hypothesis about your channel. We write it down as one.

Platforms offer their own lift studies. That is better than nothing. But the platform designs, runs and reports the result for its own product.

AdExchanger, a trade publication, ran a piece headlined "Beware the platform incrementality trap". It quotes industry people who warn against letting a platform decide whether sales are incremental or customers are new⁴. A customer the platform has never converted for you may already be in your own customer records.

An agency podcast about testing a Meta attribution setting is a useful example of the limit. The hosts say themselves that all the revenue in the test is platform-reported⁵. That does not make the result wrong. It means nobody outside the platform has checked it.

So a lift test run and reported by the same platform deserves an independent check. What counts as independent?

• A holdout test you run yourself, where part of the audience or some regions do not see the ads.
• A reconciliation of the platform's claimed sales against your own order data.
• A look at total revenue against total spend over time, before and after a big change.

You do not always need a test. Start with what you already have, and name the bias of each method.

1. Put reports next to the shop. Add up the sales every platform claims for one month. Compare with real orders. If the claims exceed the orders, some sales are claimed twice.
2. Watch total efficiency. MER (marketing efficiency ratio) is total revenue divided by total marketing spend, without splitting credit between channels. It cannot isolate one channel, but it does not depend on any platform's counting rule.
3. Switch something off in a few regions. A regional pause is a rough geo test. It is noisy, so say so in the write-up.
4. Test only where the decision is open. When a test is not worth its cost, when not to run an incrementality test explains what to do instead.

A short test on little volume is mostly noise. Run it for too short a window and you may read chance as an answer. We would rather you write "can't tell" than a confident wrong number.

Do not borrow a correction factor either. Sellforte, a measurement vendor, writes that a universal correction factor is rarely defensible¹. Haus says benchmarks exist but will not tell you what you need to know, because what works for one business often does not work for another⁶.

Decide on incremental margin, not attributed revenue. Revenue is the start. Money is the end.

Example (illustrative): Two channels, €50,000 of spend each. Channel A reports a ROAS of 6.0, but a test finds only a third of its credited revenue is incremental: iROAS 2.0. Channel B reports 2.0, but its ads also drive sales the platform cannot see: iROAS 3.0. With a contribution margin of 40% of revenue, A returns €40,000 of margin on €50,000 of spend. B returns €60,000. The dashboard ranks A first. The bank ranks B first. The numbers are round and invented; the logic is not.

Our convention: contribution margin means revenue minus the variable costs of each order, before marketing. That is our view, not a market standard.

Which of your channels would you defend to your CFO with something other than the platform's own number? Start there. If the list is short, that is your measurement agenda.