Method 06 · Proof
Cause
testing.
Ad platforms report the sales they think they caused, and they mark their own homework. Cause Testing compares people or places that saw your ads with similar ones that didn't. The difference is what your ads actually added, often called incrementality.
Written by Dan Picken, founder. Last checked 27 October 2026.Example shape. The lift is what the ads earned. The gap above it is credit they took.
The platform
grades itself.
When someone sees an ad and then buys, the platform counts the sale. But some of those people were going to buy anyway: they already knew you, searched your name, or were coming back from an email.
Google's own Conversion Lift compares a group that sees the ads with a group that doesn't, because the platform's normal reporting can't separate the two.
Google's researchers described the geographic version back in 2011: split regions at random into test and control, run ads in one, and compare.
You don't need a big-company budget to run one. Meta publishes GeoLift, an open-source tool for measuring lift by region, which builds a stand-in control group from similar areas when you can't split regions perfectly.
Why tests need
to be big.
Ad effects are small next to everyday noise. In 25 large field experiments, Lewis and Rao found sales so variable that an informative test could easily need more than 10 million person-weeks.
That's the honest reason most accounts can't prove small differences. It's also why the questions worth testing are big ones: does this channel add anything, or does brand search earn its budget?
The other half is reading the result honestly. A lift with a range that crosses zero isn't a lift. It's 'we can't tell yet'. The reader below shows that range.
How it
works.
- 1
Pick the question
One clear question, like 'what do my brand ads actually add?' or 'is Meta driving sales or just claiming them?'
- 2
Design the test
Hold ads back from some people or regions, and keep everything else the same.
- 3
Run it long enough
Weeks, not days, with enough sales in both groups to tell a real difference from noise.
- 4
Act on the gap
The lift is what the ads earned. Budget follows the lift, not the platform's report.
Try it
Read a
test result.
Ran a holdout? Enter the conversions and the size of each group (people, or a baseline like last year's sales in those regions). You'll get the lift, and a range that shows how sure you can be. The starting numbers are an example.
Saw the ads
Held back
Lift from the ads
-
Rough 95% range, treating conversions as random counts. Real tests also have uneven groups and seasons, so leave a margin before moving big budgets.
Three results,
three decisions.
The same test can come back three ways. What to do with each:
| Result | What it means | What to do |
|---|---|---|
| Lift, and the whole range is above zero | The ads caused extra sales | Keep the spend, and test whether more of it keeps paying |
| The range crosses zero | Can't tell from this test | Run longer, use bigger groups, or test a bigger change |
| The ads group did worse | Usually a problem with the groups | Check the groups were comparable before acting |
The most common mistake is reading the middle row as 'the ads don't work'. It means the test couldn't tell.
In my experience
Brand search is where I most often see the gap between what the platform claims and what the ads actually add. Plenty of those people were already looking for you and would have clicked the result right underneath. That doesn't always mean switch brand ads off. It means test it before calling it your best campaign.
When it's
not worth it.
- Accounts with only a few sales a week often can't detect realistic effects. Lewis and Rao's work shows why.
- A test measures the period you ran it in. Re-test when things change.
- Groups must be comparable. A holdout region with a big local event, or a new competitor, breaks the test.
What you get
- A test plan: question, groups, length, what counts
- Set-up in Google Ads, Meta or by region
- A plain-English read-out of the result
- Budget changes based on what the ads actually added
The technical bit
- Incrementality
- Sales that would not have happened without the ads.
- Holdout
- A group deliberately kept from seeing the ads, for comparison.
- Geo test
- Running ads in some regions and not others, then comparing.
- Attribution
- The platform's way of giving a sale credit. Useful, but it isn't proof.
Where it's used
Questions
people ask.
Is my account big enough for a lift test?
It needs enough sales in both groups to see a real difference. Smaller accounts can still run simpler before-and-after or regional tests, as long as the result is read honestly.
Will a test lose me sales?
Possibly a few, in the holdout group, for a few weeks. That is the cost of finding out whether the rest of your budget is earning its keep.
What does statistically significant mean?
That the difference is big enough, given the noise, that it's unlikely to be chance. In plain terms: the range of likely results doesn't cross zero.
What is a confidence interval?
The range the true result probably sits in. A lift of 20%, likely between 5% and 35%, is a real lift. A lift of 20%, likely between -10% and 50%, isn't proven.
How big should the holdout group be?
Big enough to measure the effect you care about, and small enough that you don't lose much while testing. Often 20% to 30% of regions or audience. The test planner on the Cause testing service page gives a rough guide from your numbers.
What is a geo test?
Ads run in some regions and not in similar ones, then total sales are compared. Google's researchers described the approach in 2011.
The other methods
Want the whole market mapped, not just the top 3 fixes? See the Market Blueprint.
Want your market
mapped?
Takes a minute. Audits come back within 1 to 2 business days. For anything else, I'll get back to you as soon as possible.
What happens next
- 1I'll ask for read-only access, or just your website for a search audit.
- 2I go through it and write up the 3 biggest fixes.
- 3A 20-minute call to walk you through it. No pitch unless you ask.