What's the Difference Between Incrementality and Attribution?

Attribution assigns credit. Incrementality builds the counterfactual. Attribution asks which touchpoints came before a sale. Incrementality asks whether the sale would have happened without the ads at all. One tells you who claimed the conversion. The other estimates who caused it. You need both, but they should not have the same job.

Most marketing teams are not short on numbers. They are drowning in them. Meta has a number. Google has a number. Shopify and GA4 have their own.

Most of those numbers answer the same question: who should get credit for the sale? None of them answer the one that matters in a budget meeting. What would we lose if we stopped spending?

That is why platform ROAS can look great while the bank account tells a different story. Attribution and incrementality are not two versions of the same measurement. One is credit assignment. The other is causal evidence.

Here is the fast version, then the opinionated one.

That is the whole argument in one table. The rest explains why it holds, when each tool wins, and which number to trust when they fight.

‍

Why isn't attribution proof that your ads worked?

Because attribution only shows an ad was present before a sale, not that it caused one. Last-click hands full credit to whatever the buyer touched last, even when they were already going to purchase. It records correlation and reports it as performance. A buyer who would have converted anyway still lands in the report as a win.

Picture a customer who already loves your brand. They open your retargeting ad out of habit, click, and buy the thing that has been sitting in their cart for a week.

Last-click credits that sale to the ad. The ad did almost nothing. The purchase was already coming.

Now multiply that across every branded search click, every retargeting impression, and every email-then-ad sequence. You get a report that looks like genius and a P&L that does not agree.

This is not a small effect. When researchers ran 15 large field experiments at Facebook and compared real experimental lift against what observational attribution reported, the attribution methods often failed to recover the true effect, overstating it in half the experiments by threefold or more (Gordon et al., Marketing Science, 2019).

Attribution is not lying. It was built to assign credit, not to test cause. Those are different jobs, and most teams ask it to do the one it cannot.

‍

When should you use attribution vs incrementality?

Use attribution for daily pacing, creative reads, and campaign hygiene. It gives fast, channel-level signal. Use incrementality for budget decisions, scaling calls, and anything where being wrong costs real money. Attribution tells you what happened. Incrementality estimates what you caused. One runs your day. The other should govern your budget.

Attribution is fine for the day-to-day. Which creative is getting clicks. Which campaign is pacing behind. Whether a landing page tanked overnight. Fast signal, low stakes. Keep using it for that.

The trouble starts when attribution gets promoted into the budget meeting. The moment someone says "Meta is at 4x, let's double it," a credit-assignment number is steering six or seven figures.

That decision needs a different tool. Incrementality builds a counterfactual, a credible estimate of what would have happened without the spend, and measures the difference.

The rule is simple. Attribution for diagnostics. Incrementality and MMM for money.

‍

What happens when attribution and incrementality disagree?

They disagree constantly, and the causal read should carry more weight, assuming the test was powered, cleanly run, and scoped to the decision. A platform reports 5x because it counts every conversion after an ad. A holdout shows 2x because it counts only the revenue that vanished when ads stopped. Those are not two equal opinions. One has a counterfactual behind it.

Here is the conflict buyers actually feel.

The platform number says scale. Finance says margin is tight. The agency says performance is strong. Then the holdout says half the revenue would have happened anyway.

Put numbers on it. A brand spends 50,000 dollars on a channel. The platform reports 250,000 dollars in revenue, a clean 5x. The agency says scale. The dashboard agrees. Everyone feels good.

Then the brand runs a geo holdout. The channel is paused in matched markets, actual revenue is compared against the expected baseline, and the incremental return comes back closer to 2x.

That is a completely different budget decision. At 5x, you scale aggressively. At 2x, you check margin, payback, and saturation before adding a dollar. Same spend, same channel, same month, different standard of evidence.

This is the distinction Meta's own incremental attribution model now makes explicit, by estimating which conversions would not have happened without the ad. The platform with the most to gain from the old way of counting is building a feature around the new one.

When the numbers fight, do not average them. Ask which one has the stronger counterfactual.

The math below is not a measurement. It shows how fast a budget decision flips when reported ROAS is adjusted by an assumed reported-to-incremental ratio.

‍

How do you know an incrementality result is trustworthy?

Demand the receipt. A real result shows its test design, holdout logic, confidence interval, and iROAS, not just a lift percentage. It carries a date, because incrementality decays as creative, spend, and competition shift. And it points to the budget decision it supports. No method, no decision, no trust.

Not every "incremental" claim deserves trust. The word has become marketing paint.

A result you can act on shows its work. Specifically, a real incrementality result should show:

• what was held out
• the treatment and control definition
• the revenue source of truth
• the pre-period fit, so you know the model could predict the baseline before the test started
• the test dates
• the spend level during the test
• the lift estimate and iROAS
• the confidence interval
• the power, or minimum detectable effect
• the budget action that followed

If a vendor or a dashboard hands you a lift number without those, you do not have a measurement. You have a claim. A number with no error bars and no method behind it is a guess in a nicer font.

The benchmarks give you a sanity check, with a caveat. These are not universal numbers. They come from DTC-heavy, measurement-sophisticated brands, mostly Shopify, measuring DTC revenue (Stella's 2025 DTC benchmarks). Across those 225 tests, the median iROAS was 2.31x, with an interquartile range of 1.36x to 3.24x. If your platform-reported ROAS sits far above that range, the next question is whether attribution is claiming demand your ads did not create.

The spread between channels makes the same point. Here is what the median incremental return looked like across those tests.

One more thing the dashboards skip. A result is not permanent. An iROAS measured in February can be stale by April once creative, promo cadence, spend level, or auction pressure moves. Treat every number as a reading with an expiry date.

‍

Where does media mix modeling fit?

Incrementality proves cause for a specific decision. Media mix modeling shows the whole picture across every channel, including the ones you cannot cleanly hold out, like TV, podcasts, and retail media. Attribution is the day-to-day read. Incrementality is the causal proof. MMM is the cross-channel map. You want all three working together, not one.

No single method answers every question. That is the part most "incrementality vs attribution" explainers miss. They treat it as a cage match. It is not.

Attribution gives fast, campaign-level signal. Incrementality gives causal evidence for the decisions that matter. MMM gives cross-channel allocation that no single platform can see, because no platform wants to tell you a different channel deserves the credit.

Use one and you are guessing with a nicer spreadsheet. Use all three and you are making decisions.

But MMM should not be trusted just because it has curves. A model earns trust through out-of-sample validation, reasonable priors, collinearity checks, and calibration against real experiments wherever possible. Incrementality is what keeps the map honest.

‍

Frequently asked questions

Is incrementality the same as attribution?

No. Attribution credits the touchpoints that came before a conversion. Incrementality estimates whether the conversion would have happened without the ad at all. One is bookkeeping. The other is evidence.

Should incrementality replace attribution?

No. Attribution still earns its place for pacing, diagnostics, and creative reads. It just should not be your source of truth for budget allocation. Keep both. Let incrementality govern the money.

Is a platform lift study the same as independent incrementality?

No. Platform lift studies like Meta Conversion Lift are single-channel, and the platform selling the media grades its own work. They beat click attribution, but they cannot see cross-channel overlap or what your other channels contributed. That is the job of independent testing and media mix modeling.

How long does an incrementality result stay valid?

Not forever. A result from February can be stale by April if creative, spend level, promo cadence, or auction pressure shifts. Treat it as a reading with an expiry date, and retest when the inputs change. For a continuous read instead of one-off tests, always-on incrementality keeps the measurement current.

Do you always need to run an incrementality test?

No. A test is worth running when the decision is large, uncertain, and changeable enough to justify it. If the budget at stake is small, or the answer would not change what you do, skip the test and save it for a decision that matters.

‍

Which number should run your budget?

Last-click is not measurement. It is credit assignment dressed up as proof.

Attribution can stay for what it is good at. But the dollar decisions, the scale calls, the cut-or-keep questions, those need a counterfactual behind them.

Here is the standard worth holding every number to. Any number used to move budget should show its counterfactual, its uncertainty, and its expiry date. Most of the numbers on your dashboard cannot.

If you want to see what your channels look like under an actual test, that is what Stella is built for.