Most growth teams still run on attribution: a model that divides the credit for a conversion across the touchpoints that preceded it (last-click, first-click, or some weighted blend in between). Attribution is comfortable because it always produces an answer, and the answer always adds up to 100%. The problem is that it answers the wrong question.
Attribution asks which touchpoint was present when the outcome happened. The question that actually governs a budget is what would have happened if we hadn’t acted at all? That is a question about incrementality (the additional impact of a decision beyond what would have occurred organically), and it is a causal question, not a bookkeeping one.
The trap of crediting the inevitable
Consider a customer who was already going to buy. They search for your brand, click a paid ad that sits on top of the organic result, and convert. Last-click attribution hands the paid ad full credit. But the incremental value of that click may be close to zero. The sale was coming either way. Scale that across a channel and you get the classic failure mode: a channel that looks efficient in the attribution report while adding little real lift, quietly absorbing budget that would work harder elsewhere.
This is not a hypothetical edge case. It is the central reason performance-focused AI vendors have moved their language from attribution toward incrementality and causal lift: measuring true additional impact rather than last-click vanity metrics.
What measuring incrementality actually involves
You cannot observe the counterfactual directly. A given customer either saw the intervention or didn’t. So incrementality is estimated, with methods chosen to fit the decision:
- Randomised experiments (A/B). The cleanest tool when you can split users or requests randomly: the control group is your counterfactual. Best for on-site decisions, offers, and product changes.
- Geo and market-level tests. When you can’t split cleanly at the user level (channel spend, brand campaigns, market-wide effects), you compare treated and untreated regions over time.
- Quasi-experimental / causal-inference methods. Where even geo splits aren’t available, you model the counterfactual explicitly (for example, structural time-series approaches to estimating causal impact). These carry more assumptions and deserve more scrutiny, but they extend measurement to decisions a pure A/B test can’t reach.
The common thread is discipline: define the intervention, define the baseline, and measure the difference, not the total.
Why this matters more once AI is making the decisions
When a model is choosing offers, prices, or recommendations thousands of times a day, an attribution-shaped feedback signal will train it toward whatever correlates with conversion, including the customers who never needed persuading. Feed the same system an incrementality signal and it learns to spend effort where the effort actually changes the outcome. The measurement you choose doesn’t just report on the decision loop; it shapes it.
That is why adBrain treats causal measurement as a first-class part of a decision system, not a quarterly analytics exercise. Prediction estimates what is likely; the decision acts on it; and incrementality proves what changed: the predict, decide, prove loop. Without the third step, you are optimising a number that feels precise and means little.
A practical starting point
You don’t need to rebuild everything to escape attribution’s blind spot. Three moves get you most of the way:
- Pick one high-spend decision where you suspect you’re crediting the inevitable: a brand-term budget, a broad retargeting audience, a blanket discount.
- Design a holdout. A randomised control, or a geo holdout if user-level isn’t feasible. The holdout is the whole point: it is your counterfactual.
- Compare lift, not totals, and let the result move budget. Then repeat on the next decision.
Attribution will still have a place for diagnostics and storytelling. But the decisions that move money should be governed by what genuinely changed. And that is incrementality.
This is how adBrain builds predictive & causal modelling into the systems we deploy. Want to put incrementality behind your growth decisions? Talk to us.
Note on evidence: the methods above (randomised experiments, geo testing, structural time-series causal impact, incrementality measurement) are drawn from established public experimentation and causal-inference literature. Any performance figures adBrain publishes are substantiated against a client’s own measured results.