Incremental attribution for ads

광고 증분 기여 측정

Separates conversions likely to happen anyway from added ad outcomes.

···
html
<div class="lift"><div class="heading">ATTRIBUTED ≠ INCREMENTAL</div><div class="bars"><div><span>NO AD</span><i id="control"></i><b id="cv">40</b></div><div><span>EXPOSED</span><i id="exposed"></i><b id="ev">55</b></div></div><div class="delta">ADDED PURCHASES <strong id="difference">+15</strong></div></div>
css
.lift{width:min(86%,460px);font:800 clamp(10px,max(2vmin,1vw),16px)/1.2 sans-serif}.heading{border-bottom:2px solid var(--fg);padding-bottom:8px;letter-spacing:.08em}.bars{height:clamp(105px,28vmin,220px);display:flex;align-items:end;justify-content:center;gap:clamp(25px,9vmin,75px);border-bottom:2px solid var(--line);margin-top:clamp(8px,2vmin,18px)}.bars div{height:100%;width:clamp(70px,19vmin,140px);display:flex;flex-direction:column;align-items:center;justify-content:end;gap:5px}.bars span{font-size:max(10px,1vw,.8em);color:var(--muted)}.bars i{display:block;width:100%;background:var(--fg);transition:height .5s}.bars div:last-child i{background:var(--accent)}.bars b{font-size:clamp(15px,3vmin,24px)}.delta{display:flex;justify-content:space-between;align-items:center;gap:5px;margin-top:clamp(10px,3vmin,20px);color:var(--muted)}.delta strong{font-size:clamp(25px,7vmin,55px);color:var(--accent)}
js
const cases=[[40,55],[42,54],[39,57]];let n=0;function draw(){const [control,exposed]=cases[n++%3];document.querySelector('#control').style.height=control+'%';document.querySelector('#exposed').style.height=exposed+'%';document.querySelector('#cv').textContent=String(control);document.querySelector('#ev').textContent=String(exposed);document.querySelector('#difference').textContent='+'+(exposed-control)}draw();setInterval(draw,1450);

Standard attribution counts purchases linked to an ad. An incremental view tries to estimate the added effect after accounting for purchases that would have happened anyway. Meta announced an update to its incremental attribution model in 2026.

The demo compares fictional exposed and comparison groups and highlights only their difference. It illustrates a counterfactual comparison; it is not Meta's model or an actual experiment result.

Causal interpretation needs a credible comparison group and control of selection bias. Where possible, compare platform estimates with independent experiments.

When to use

Use it to assess whether attributed sales were added sales, with careful comparison design.

Open as page ↗