Attribution vs. incrementality: which number should you trust?

Attribution answers, "Which touchpoints get credit for this sale?" Incrementality answers, "Would this sale have happened without the ad?" A report can answer the first question perfectly and tell you nothing about the second.

Picture someone who has already decided to buy from you. They search your brand name, click your ad, and buy. Attribution gives that ad the credit, because it was there. Incrementality may give it very little, because the sale was already coming. Retargeting can raise the same question, since it often reaches people who are already close to buying.

So when the numbers disagree, one isn't simply wrong. They're measuring different things. Trouble starts when you use a credit report to answer a question about cause.

Every attribution model works from the interactions it can see. Last-click attribution favors whatever came just before the sale, which often means brand search and retargeting. Data-driven attribution spreads credit more carefully, but it's still sharing credit among observed interactions. In Google Ads, it looks at interactions with Google's own ads, so it can't see your email, your sales calls, or a podcast mention.

Each platform also reports on its own performance. If Google and Meta both touched a buyer, each may claim the sale in its own report. Add up the conversions from your platforms, compare the total with your CRM, and you'll often see the gap for yourself.

None of this makes attribution useless. It's fast, granular, and available every day, which suits many decisions. It simply can't show what would have happened if the ad had never run.

Incrementality needs a comparison. You show ads to one group, hold them back from a similar group, and measure the difference. That difference is the lift, and it's the closest thing marketing has to a measure of cause.

Some ad platforms run these tests for you. Google Ads Conversion Lift compares a group that sees your ads with a control group, using either user-based or geography-based splits. It isn't available to every account, and Google notes that the geography option supports offline data. Meta's Conversion Lift tests also split people into a test group and a control group.

You can run a geo experiment on your own as well. Choose similar regions, change spend in some and hold it steady in others, then compare results using your CRM. This can suit lead generation well, because qualified leads and revenue live in your own system. Expect it to take planning, and to need enough volume in each region.

Pick the method by the size of the decision. Small, frequent choices need quick data, and a directional answer is usually good enough. Large budget choices can justify a slower, cleaner answer.

The two work best as a pair. A test gives you an accurate reading at one moment, and attribution gives you a rough reading every day. Combine them, and you can adjust the daily number using what the test found.

Divide the incremental conversions from your test by the conversions attribution reported for the same campaigns and period. That gives you an adjustment factor, and spend divided by incremental conversions gives you your real cost.

Now apply that factor to your weekly attribution numbers for that channel until the next test. Keep running your dashboards as usual, with one adjusted line for what the spend really produced. Retest when your spend, targeting, or market changes enough to make the old factor doubtful.

A test is only worth its cost if the result will change what you do. Start with the channel where you have the most spend and the most doubt about what it's really producing. Then write down the decision before the test begins.

1. 1. State the decision, such as, "If lift is below our break-even point, we'll move part of this budget."
2. 2. Pick a metric from your CRM, such as qualified leads or revenue, instead of platform conversions.
3. 3. Run the test long enough to cover the normal delay between click and outcome.
4. 4. Estimate the required sample size first. The Sample Size Calculator can show whether your volume can detect the change you care about.
5. 5. Hold other changes steady in both groups while the test runs.
6. 6. Report the result as a range. A confidence interval that includes zero means you couldn't detect lift, which isn't proof that the ads did nothing.

Trust each number for the job it can do. Attribution is reliable for comparing similar things inside a channel, as long as your tracking is sound. Incrementality is your best evidence for what your spend caused, as long as the test was designed well and is still recent.

If you only have attribution today, treat its return figures as an upper estimate for channels that catch existing demand, then plan one test. Results from that test can also feed your budget work. The Marketing Funnel Forecaster is a simple way to see what an adjusted number means for your pipeline. For a more complete program, see Attribution & Incrementality.