Incrementality: how to find out what advertising actually brought

Every ad platform can answer the question “how many conversions should I credit to this campaign”. None can answer “how many of them would not have happened without it”. Those are not two versions of the same thing. It is the difference between accounting and causation.

At small budgets you can live with attribution. At large ones, and in channels that reach people who already know you — brand, remarketing, first-party audiences — the gap is so wide that attribution cannot support a decision at all.

Why platform ROAS overstates

An ad system sees only what happened after a click on, or an impression of, its own ad. It does not see the person who would have bought anyway and merely clicked the ad on the way instead of the organic result. It credits itself with that conversion — and it is right that the click happened. It is not right that the click caused it.

The effect is strongest where advertising reaches people with existing purchase intent:

• Brand search. In a public dataset of 225 geo tests of DTC brands, the median incremental ROAS was 0.70×, below the line of profitability.
• Remarketing. People who were in the cart often return without an ad.
• First-party audiences. Existing customers are the cheapest conversions and also the ones least likely to be new.

Geo holdout: the most accessible method

The principle is simple. Split your territory into two groups. In one you switch the channel off (or cut it sharply), in the other you leave it unchanged. Afterwards you compare total revenue from those areas — not the campaign’s own numbers.

The procedure:

1. Pick the question. One. “Does the brand campaign bring revenue that would not otherwise arrive?” Not “how is marketing doing”.
2. Split the territory. Regions or postcode areas, ideally so both groups have similar historical revenue and a similar trend over time.
3. Work out what you can detect. Start from the smallest difference that would change your decision, then derive the length and volume required. A test that cannot detect a difference smaller than thirty per cent will not answer “is this worth it”.
4. Let it run the whole time. Typically two to six weeks. Do not look every day and do not evaluate in flight.
5. Evaluate against the counterfactual. That is, against what the control regions predict would have happened in the test regions without the intervention — not against last month.

What the platforms themselves offer

Lift measurement reached the interfaces in 2026:

• Experiment Center in Google Ads consolidated testing tools in January 2026. It includes lift studies, where a control group sees no ads at all and is compared with the exposed group.
• Conversion Lift opened in self-serve form for Search and Performance Max on 7 October 2026 according to PPC Land; lift metrics became available in the API in August 2026.
• Meridian GeoX, a geo experimentation layer feeding a marketing mix model, was previewed in May 2026 with testing to begin later that year. It is not a generally available tool for every advertiser.
• Uplift Experiments for Demand Gen were announced in May 2026.

Platform tools have one systemic weakness: they measure lift within their own ecosystem. A geo test measured on total company revenue is independent of that — and for exactly that reason more uncomfortable.

What to do when you lack the volume

Not every company has the revenue to build a geo test on. The fallbacks, best first:

1. A simple on/off test over time on one channel, alternating weeks and watching seasonality. Weaker, but better than nothing.
2. Watch the total, not the channel. If you switch brand off and organic clicks rise by most of that volume, you have learned something substantial without statistics.
3. Ask in the order. “How did you hear about us?” is imprecise, but for small companies it is often the only available reading of reality.
4. Compare with the period before the channel launched, adjusted for seasonality. The weakest method; use it knowing it will not separate the influence of everything else.

What to do with the result

A test result is not a verdict of “on or off”. It is a new number for the decision:

• Incremental ROAS below break-even (that is, below 1 ÷ margin) means the channel in its current form reduces profit. Before switching it off, try reducing its scope — for instance brand only on queries where competitors bid against you.
• Incremental ROAS well above platform ROAS happens with upper-funnel channels (video, content). There attribution understates instead.
• A result with no demonstrable difference means “we do not know”, not “it does not work”. Either size the test better, or accept that at this spend it is not worth finding out.

And above all: the result holds for that period, market and setup. A year later it may have changed. Incrementality is not a one-off certificate.

What to take away

Incrementality is an uncomfortable discipline, because it usually lowers the numbers a company was proud of. It is also the only way to know what you are buying at larger budgets. The recommendation we follow: test brand once a year, remarketing once a year, and before any large budget increase verify that the channel you are sending money into really does bring additional business.