
Measuring the true effect of your ads with Geo Testing
Figuring out an ad's real effect is tricky. Clicks don't tell the whole story and attribution models fall short. The Solution: Geo Testing.

Incremental CPA is spend divided by the conversions the advertising actually caused, rather than by the conversions attributed to it. Because attributed conversions include ones that would have happened anyway, reported CPA is always the more flattering number.
iCPAReported CPA divides spend by the conversions a platform attributes to itself. Incremental CPA divides the same spend by the conversions that would not have happened otherwise. Since attributed conversions include people who were already going to buy, incremental CPA is always higher — often by a multiple rather than a margin, and most on the channels whose reported numbers look best.
It is the cost-side counterpart to iROAS, and the two are used for different decisions rather than being interchangeable. iROAS is a revenue ratio and drives budget allocation between channels. Incremental CPA is a cost per unit and drives bidding and target setting, which is why it is the figure that matters for anyone operating a bid strategy day to day. A team optimising towards a reported CPA target is optimising towards a number that includes conversions the advertising did not produce.
The practical consequence of that is systematic overbidding on the channels that capture existing demand. If a target CPA of £30 is set against reported conversions and the true incremental CPA is £95, every bid decision is being made on a figure three times too generous. The bidding algorithm dutifully buys more of exactly the traffic that was already converting, and the reported number stays healthy while the incremental one deteriorates.
Establishing it requires the same machinery as any incrementality question: withhold the advertising from someone and count the difference. A geo experiment, a conversion lift study or ghost ads all work. What does not work is any amount of attribution modelling, because attribution allocates credit among observed touchpoints and never sees what would have happened without them.
For a business case the figure to compare against is contribution margin per conversion rather than revenue per conversion. An incremental CPA of £95 against a £120 order value looks acceptable and is not, if the gross margin on that order is £40. Setting the threshold on margin is the same correction that applies to iROAS break-even, and it is missed about as often.
One ratio, its relationship to the reported figure, and the threshold it should be judged against.
iCPA = ad spend / incremental conversionsIncremental conversions come from an experiment, not from the platform's conversion column.
iCPA = reported CPA / incrementality rateAt 30% incrementality the true cost is 3.3× the reported figure.
iCPA = average order value / iROASTwo views of the same experiment — one cost-side for bidding, one revenue-side for allocation.
iCPA < contribution margin per conversionMargin, not revenue — see the one-proportion z-test calculator for the lift measurement.
A retailer runs paid social with a reported CPA target of £28. A geo holdout measures how many of those conversions were incremental. Average order value is £86 and gross margin is 41%, so contribution per conversion is £35.30.
The true cost per conversion is £74 against a £35.30 contribution — the channel is losing roughly £39 on every conversion it genuinely creates.
The reported CPA of £28 sits comfortably under the £35.30 contribution and suggests a profitable channel worth scaling. The incremental figure inverts that completely: only 38% of attributed conversions were caused by the advertising, so the real cost is £74 and each one destroys about £39 of value. Note what the bidding system has been doing throughout — optimising towards the £28 target means buying more of the traffic that converts most readily, which is disproportionately the people who would have converted anyway, so the incrementality rate degrades as the algorithm improves against the wrong objective. The interval matters here too: at the optimistic end of 3,680 incremental conversions the iCPA is £58, still well above contribution. The conclusion survives the uncertainty, which is what makes it actionable rather than merely suggestive.

Figuring out an ad's real effect is tricky. Clicks don't tell the whole story and attribution models fall short. The Solution: Geo Testing.


In today’s privacy-focused era, the different attribution models create many blind spots for marketing analysts and decision makers. However MMM & Geo Tests can help.

Applying it to a live measurement problem is the part that goes wrong. If you are designing an experiment, reading a result you do not trust, or trying to work out what your marketing actually caused, that is the work we do.