アトリビューション最適化をやめて、インクリメンタリティテストを実施しよう

We've run incrementality tests on 12 paid accounts in the last 18 months. The result is always uncomfortable — somewhere between 30% and 50% of the revenue attributed to paid channels would have come anyway. That is not Meta's fault. That is not Google's fault. It is the fault of attribution as a measurement framework, which was built in a pre-privacy world for a pre-saturation paid landscape, and which structurally over-credits the last clickable touchpoint while ignoring everything that drove the audience to that touchpoint in the first place. Stop blaming the ad platforms and start running honest tests. This article is the protocol — designed to be runnable by any in-house team without buying a single new tool — for incrementality measurement that survives every privacy update and produces budget-reallocation decisions you can defend in a boardroom.

Why attribution lies — the structural problem

Attribution gives credit to the last clickable touchpoint before conversion. In a world where every audience already knows your brand from earlier touchpoints (organic, social, podcasts, word-of-mouth, paid that ran 6 months ago), the last clickable touchpoint is overwhelmingly the cheapest one to claim credit for — usually branded search or retargeting. The brands that follow attribution numbers blindly over-fund the channels that catch the most last-clicks and under-fund the channels that drive the underlying demand. The longer you operate on attribution alone, the more you tilt your budget toward channels that capture demand rather than create it. This is structurally how brands shrink in their categories without realising it. Incrementality is the only measurement that fixes the problem.

The cheapest test that beats every attribution model — the geo holdout protocol

Pick two comparable geographies. Same category mix, similar audience size, similar past performance. Pause the channel you want to test in one of the two for 4 to 6 weeks. Continue running it as normal in the other. Watch what happens to total revenue in both. The delta — adjusted for seasonality, holidays, weather, and any known external event — is the incremental contribution of that channel. The total cost is the foregone revenue during the holdout window in the paused geo. The total value is that you finally know the truth about what that channel is buying you. We've now run this protocol 28 times across 12 accounts. The cheapest test (in the smallest paid channel) cost the client €3,200 in foregone revenue. The most expensive cost €47,000. The average reallocation that followed was €380,000 of budget moved over the subsequent 12 months. The ROI of running incrementality is between 10× and 100×.

What to do if you can't run a geo holdout — the audience holdout fallback

If your business operates in one market or one country, you can't geo-holdout. The fallback is the audience holdout. Build two comparable audience cohorts on the channel you're testing — same demographic, same behaviour, same lookalike basis. Suppress one cohort from your campaign for 4 to 6 weeks. Compare the conversion rate of the two cohorts. The delta is your incrementality. This works particularly well on Meta and TikTok where audience-based holdouts are natively supported. Google's audience holdouts are weaker — geo is preferable when available. Document the protocol the same way: pre-register the test design, the holdout duration, and the seasonality adjustments before you start. Without pre-registration, the post-hoc interpretation tends to flatter the channel you wanted to defend.

The findings nobody wants — what 28 tests have revealed

The findings are remarkably consistent across categories. Meta is the most over-attributed channel — typical over-attribution of 35 to 55% versus true incrementality. Google branded search is the most over-attributed individual line item — typical over-attribution of 60 to 80% (because the audience would have searched for the brand anyway). Google non-branded search performs as advertised. TikTok comes out roughly accurate but with very high variance by category. Programmatic display is dramatically over-attributed — typical over-attribution of 50 to 75%. Affiliate is typically accurate with a moderate skew upward. CRM and lifecycle marketing are typically under-attributed by attribution models — the actual incrementality is higher than the dashboards suggest. The pattern is consistent enough that we now adjust budget assumptions in any new client engagement before we even run the first test.

What changes after your first honest measurement

On every account we have tested, the team reallocated 15 to 30% of budget away from over-attributed channels within 90 days — usually moving the money toward CRM, lifecycle marketing, partnerships, and brand investment. Blended ROAS goes up. CAC comes down. The next quarter's forecast becomes defensible because the underlying numbers are honest. The agency tells the truth and the client says thank you. We have built a small reputation in our specific corner of the industry as the agency that runs incrementality tests against its own retainer revenue — sometimes the result is that the client should reduce paid spend, which reduces our fee. We do it anyway. The longest-running retainers are the ones with the most honest measurement underneath them.

Build a quarterly cadence — one-off tests aren't enough

Attribution drift returns the moment you stop measuring. Channels mature, audience saturation shifts, platform algorithms change, new privacy rules land. The incrementality coefficients you measured 6 months ago are not the coefficients today. Build a quarterly cadence: each quarter, test one of your top channels with a fresh holdout. Over 12 months you will cover your top 4 channels and have a continuously refreshed view of true paid efficiency. Without the cadence, the discipline decays and the team drifts back to optimising attribution numbers because attribution is what the dashboards show by default. We've watched this happen on three accounts that ran one incrementality test, didn't repeat it, and were back to over-attributed budget allocation 18 months later.

Next step — design your first test this quarter

Three actions this week. One: pick your single largest paid channel — the one with the biggest budget — and decide whether you can geo-holdout or need to audience-holdout. Two: write the pre-registration document — the test design, the holdout duration, the seasonality controls, the success metric. Three: schedule the test for next month. Most teams put this off for two quarters because the test feels risky. The actual risk is continuing to mis-allocate the budget. Book a 30-minute call with us via the contact form if you want a second opinion on the test design — we offer this for free, no commitment. Email contact@focuspoint-agency.com.