Incrementality: testing whether Google Ads really adds revenue

A test is only worth running if it can overturn a budget decision

Before choosing a test method, you need to settle two things: which block of budget is up for decision, and from which result you would actually cut it. Without them, the test costs you revenue and delivers an interesting number with no consequences. With them in place, you can calculate before launch whether the test can separate the answers at all.

Why test at all instead of trusting the account? Attribution distributes revenue; it does not measure what would have been lost without the ad. In eBay's field experiment, simple regressions showed an ROI of over 1,400% for paid search, while the experiment showed minus 63%. The 95% interval ran from minus 124% to minus 3%. The scale does not carry over to a less well-known shop, but you should still take the direction of the error seriously.

Your margin raises the hurdle above iROAS 1

iROAS is additional revenue divided by additional cost. Google assigns the status "Significant Positive iROAS" when the net new revenue was greater than the amount spent. For a retailer, that is not enough, because only the contribution margin of that additional revenue stays with you. The rule is therefore:

Break-even iROAS = 1 / contribution margin before ad spend

The margin must refer to the same basis as the conversion value, so to gross revenue if you report gross values, and it must include returns, shipping and payment costs. How to derive it is explained in the article on break-even ROAS.

Google's own example from its help pages shows what this means: iROAS 2.2 with a range of 1.3 to 3.5. Depending on the assumed margin, the same result leads to entirely different decisions:

At a 30% margin, a result can therefore be significantly positive and still make a loss; this interval would only give a clear yes from a margin of around 77%. The calculation for the third column: 2.2 times margin minus 1. It has one limitation: it counts short-term contribution margin. If the block mainly brings in new customers with strong repeat purchasing, the hurdle may be lower, but you then have to justify that with customer data.

Which test is realistic for a mid-sized retailer

Sorted by what I would check first (requirements according to Google, as of November 2026):

According to Google, geo tests are noisier than user-based lift and often need more budget and a longer duration; on the other hand, GeoX works with aggregated regional data without user-level attribution.

When setting up, you should allow for three consequences. Larger holdouts of up to 50% increase the sample but, according to Google, also carry a higher opportunity cost in missed conversions, and that is real money. Switching from country to city or postcode targeting reduces overall traffic, according to Google, so even before the test you are no longer measuring exactly your normal account. And with budget-constrained campaigns, the system shifts freed-up budget to the control regions, inflating the baseline there. According to Google, you then have to reduce the daily budget manually and switch to individual campaign budgets.

For Austria and Switzerland, I consider the number of regions the hardest bottleneck: you will usually only reach 50 or more usable regions there with postcode groups, and revenue per region becomes thin. In Switzerland, calculate in CHF and with the Swiss margin.

The calculation before launch: can the test separate your hurdle at all?

This rough check does not replace statistical planning, but it filters things out in ten minutes. All you need is shop data.

Worked example with invented figures: a shop generates €400,000 in monthly revenue, half of it in the test regions. The block in question costs €10,000 a month; in the test, it is switched off in those regions for four weeks, which means around €5,000 less spend. At a 40% margin, the hurdle is 2.5. For the block to pay for itself, €12,500 of revenue would have to disappear there, a little over 6% of €200,000.

Now the comparison: divide the revenue of the test regions by that of the control regions, for every four-week window in the past. If this ratio fluctuates by 8% anyway, your 6% effect disappears in the noise, and the test will very probably decide nothing. If it fluctuates by 2%, the effort is worthwhile.

Test where the doubt is greatest and the share of the budget is noticeable, so brand or existing customers rather than the whole account. At eBay, brand ads had no measurable short-term benefit: 99.5% of the lost click traffic continued via organic search, while non-brand worked for new and infrequent buyers but not for frequent ones. This need not apply to a lesser-known brand or with aggressive competitors bidding on your name, which is why your own calculation for brand is worth doing.

What to do if no formal test is possible

Then do not pretend you have measured. I consider before-and-after comparisons following a pause weak, because seasonality, prices and competition all move at the same time. Sophisticated models do not reliably rescue this either: in 663 Facebook experiments, observational methods with over 5,000 user characteristics estimated a median lift of 24% to 64% for lower-funnel outcomes, while the experiments showed 5%. That is Facebook, not Google, but it still shows the limits of models without a control group.

What stays honest is an open commercial bet with a rule set in advance. I would write it down like this:

• Block and step: for example, cut brand spend by a quarter, rather than restructuring the account.
• Period: outside promotions and seasonal peaks, long enough for your conversion lag.
• Metric: total revenue and contribution margin from the shop, not the account report, supplemented by the overall MER calculation.
• Reversal rule: if shop revenue falls by more than the threshold set in advance, the budget goes back.

That proves nothing, but it limits the possible damage. It works poorly where the block is small relative to total revenue, because the effect then does not even become roughly visible. In that case, the decision remains an assumption, and you should call it one.

How to read the result once you have it

Read the interval against your hurdle first, and only then the status. If the whole range is above the hurdle, keep or increase. If it is below, cut. If the hurdle lies within the range, you have no answer. The point estimate on its own is misleading in that case.

On certainty level, Google recommends 90% for Conversion Lift, with 50% to 90% counting as directional results. I would scale this by reversibility: a directional result is enough for a small step you can undo next month. It is not enough for a large, permanent cut.

According to Google, "No significant lift detected" means that there was no statistically significant difference between the groups. You must not conclude from this that the campaign has no effect, especially if the test was small. Check the range again here: if it includes the hurdle, you know nothing new.

A test also only applies to the budget, period and target ROAS under which it ran. According to the help page, Google's iROAS describes the effect at your current media spend. If you then double the budget, that is a new question the old figure does not answer. How to set up the tests in practice is shown in the articles on Conversion Lift for Search and PMax and on the geo test for retailers in Germany, Austria and Switzerland.