What is Incrementality?

Incrementality is the share of conversions or revenue that happened because of the ads, not merely after a click or impression. An attributed conversion answers “which touch sat next to the event.” An incremental conversion answers “would the event have happened if that touch had not.”

The gap matters in performance marketing. Account ROAS and last-click in a tracker count credited money well. They answer poorly which campaigns create demand and which intercept people who would have bought anyway. Retargeting, brand search, and warm display almost always look better in attribution than in an incrementality test.

Incrementality does not replace attribution: you still need attribution to pay the network, train bidding, and reconcile postbacks. A lift test exists so you do not scale a channel that steals conversions from organic and from the campaign next door. Brand lift (surveys on “saw the brand / feel more loyal”) is a neighbour, not the same method: it changes stated attitude; incrementality changes observed action.

Holdout

The classic experiment splits the audience at random into exposed and holdout. Exposed sees the normal ads. Holdout sees none, or sees a neutral PSA (a public-service ad of the same format — to keep the auction and frequency, but remove the offer). Then conversion rate, revenue, or installs are compared across groups in the same window.

Lift is the difference in conversion rates, not “how many conversions the account drew.” If exposed CR is 4.0% and holdout is 3.2%, the increment is 0.8 pp, not the entire 4%. Meta Conversion Lift and similar Google / TikTok lift studies are industrial wrappers on this design. Volume is required: 20 conversions a week will not separate noise from effect. Leakage (one user in both groups, a shared cart, one household) biases lift down.

A PSA holdout costs more than “just pause 10%”: you pay for impressions that will not sell. It keeps auction competition intact, so the estimate is closer to what happens if you turn the channel off — not to “the budget simply flowed into another campaign.”

Geo-lift

When user-level randomization is impossible — CTV, OOH, parts of the walled garden after ATT, reach campaigns without a stable ID — the test is run on geography. Some regions (or cities) are treatment, others control. Controls are matched on historic sales, seasonality, and media weight, or a synthetic control is built from a basket of regions.

Geo-lift measures increment at market level: orders, revenue, installs by GEO, not by click ID. Weak spots are those of any quasi-experiment: travel spillover, national promos, news, a bad “NYC vs Dallas” match. You need weeks, not two days, and a pre-registered observation window — otherwise it is easy to pick a cutoff that flatters the chart.

For a buyer, geo-lift is how you stress-test prospecting, video, and upper funnel where account view-through over-credits and last-click zeroes the same activity. It does not replace daily campaign ROI; it is a calibration, quarterly or when the mix changes.

iROAS

iROAS (incremental ROAS) is incremental revenue divided by ad spend. The formula matches ordinary ROAS, but the numerator is only money the test treats as causal. If the account paints ROAS 4.0 and the holdout shows that three of four credited orders would have happened anyway, iROAS is about 1.0.

Relative to ROI: when only ad spend sits in cost, iROI ≈ iROAS − 1. A channel with reported ROAS 3 and iROAS 0.6 is a losing source of new demand even if last-click is “green.” The reverse also happens: reach video with weak attributed ROAS can post a positive iROAS because sales surface in search and direct a week later.

There is no separate wiki page for iROAS: it is a metric of the same experiment, not its own discipline. The same applies to geo-lift and holdout — they are methods for measuring incrementality, not catalog entities of their own.

Where attribution is systematically wrong

Last-click hands the close to retargeting and brand queries: the person was already on site or already searched the name. View-through hands the impression to people who would have converted inside the same view window anyway. Account data-driven is closer to truth than last-click, but it is still correlation along a path, not a control. MMM (media mix) estimates channel contribution on aggregates and complements the test; it is a different model, not “another holdout.”

A practical minimum without a full lift study: do not rank channels only on account conversions with mismatched windows; treat brand search and retargeting as harvest, not as proof of increment; when you pause a channel for 1–2 weeks, watch whether total revenue falls, not only that campaign’s attributed revenue. Cruder than a holdout; cheaper than believing ROAS.