What is Geo lift test?

A geo lift test measures incrementality by treating geography as the thing that is randomised. Advertising runs in some regions and not in others, and the difference in outcomes between them is what the advertising caused.

It is the older of the two main incrementality designs and the more broadly applicable. An audience holdout needs a platform that can suppress delivery for chosen individuals. A geo test needs only the ability to turn advertising on in one place and off in another, which is true of every channel including the ones that have no targeting at all.

That makes it the only causal instrument available for television, radio, print, out-of-home, sponsorship and any other activity that reaches people without knowing who they are.

The comparison is only meaningful if the regions would have behaved the same way without the advertising, so the matching decides whether the test is worth running.

The requirement is historical co-movement: treated and control regions whose outcomes tracked each other over a meaningful period before the test. If two regions rose and fell together for six months, a divergence during the test is attributable to the treatment.

What does not work is matching on characteristics. Two cities with similar populations, incomes and demographics can have quite different seasonal patterns, competitive landscapes and purchase behaviour, and a match built on the description rather than on the behaviour will produce regions that were never going to track.

Practical requirements that follow:

Enough regions. One treated region against one control is not a test; it is an anecdote with two data points. Several on each side lets idiosyncratic local events average out.

A real pre-period. Long enough to establish the co-movement and to notice regions that do not.

Balanced volume. Regions contributing wildly different conversion counts make the comparison dominated by the largest, which reintroduces the single-region problem.

Geo tests fail in specific and recognisable ways, and most of them are visible in advance.

Spillover. Advertising that crosses the boundary — a metropolitan signal reaching a neighbouring region, online targeting that resolves imprecisely, people who commute — puts treatment into the control.

Local events. A store opening, a regional promotion, a competitor's campaign, a public holiday that falls differently, a weather event. Each affects one side and not the other.

Migration. People travel, and a person who sees advertising in a treated region and buys at home in a control region has moved the effect across the boundary.

Other marketing. Email, organic activity and existing customers do not stop at regional lines, and a national promotion running during the test affects both sides unevenly if the customer bases differ.

The pre-period catches a useful share of these, because regions that are already diverging will continue to. The rest is a matter of choosing boundaries with real separation and knowing what else the business has planned during the window.

Size it before starting. Decide the smallest lift worth acting on and calculate the conversions per side required to detect it. Regional data is noisier than audience-level data, so geo tests generally need more volume or more time than an equivalent holdout.

Cover the purchase cycle. Conversions arriving after the window are conversions the treated side does not get credit for.

Hold everything still. No creative refresh, no budget change, no structural edits, and ideally no other regional marketing.

Measure at the region, not the ad account. The outcome is the business's own — orders, revenue, customers — read by region from the system of record. Reading the ad platform's attributed figures defeats the purpose, since those are the numbers the test exists to check.

Plan the switch-back. Control regions that have gone without advertising for weeks may take time to recover, and the recovery period is not clean data either.

The comparison is the change in the treated regions against the change in the control regions over the same period, which removes anything that affected both.

That difference, scaled to the treated regions' size, is the incremental outcome. Divided into the spend, it gives an incremental cost per acquisition — the figure to compare against what the business can afford.

Two honest caveats belong in the report. Regional tests have wider uncertainty than audience-level ones, so the answer is a range rather than a point. And the result describes the period it ran in: a test through a quiet season does not describe a peak.

Untargetable channels, where nothing else is available.

Whole-channel questions — does this channel add anything — which are hard to answer at the audience level because the channel reaches people through several campaigns.

Spillover effects, where word of mouth or household sharing would contaminate an audience holdout but stays inside a treated region.

Arguments about brand spend, which is the case where the stakes are usually highest and the attributed evidence weakest.

Where a business cannot bear to withhold advertising from any region for long, an alternative design alternates instead of splitting.

Advertising runs everywhere, switching on and off on a schedule — a fortnight on, a fortnight off, repeated — and the comparison is between the on and off periods within each region rather than between regions.

The advantage is that every region gets advertising, so no market is dark for the whole test and the seasonal differences between regions cancel out because each region is its own control.

The costs are real. It needs more time, because each region contributes several alternating periods rather than one continuous one. It is vulnerable to carryover, since advertising in one period affects the next. And it suits businesses with short purchase cycles, because a long cycle blurs the boundary between periods entirely.

Where those conditions hold, it is the design that meets the commercial objection most people raise about incrementality testing.

Do the regions co-move historically, over a period long enough to be convincing?

Are there enough of them on each side that no single region dominates?

Is anything else regional happening in the window — promotions, openings, competitor activity?

Can the outcome be read by region from the system of record rather than from the ad platform?

Has the test been sized, and can the account supply that volume in the time available?