The campaign produced $33.9K you would not have had

Against a matched control of the same size, exposed households bought 8.2% more. That is a real effect, but read the limits below before taking it to the planning meeting — CTV carries most of the uncertainty and almost none of the return.

Incremental sales

$33.9K

95% CI $17.0K–$50.8K

Incremental buyers

2,940

cost per incremental buyer $4.83

iROAS

2.4

on $14,200 media

New to brand

31%

of exposed buyers

Exposed / control

420,000

matched households each

Exposed against matched control

Exposed420,000 hh

Control420,000 hh

Method: incremental sales are exposed-group sales minus matched-control sales over the same eight weeks, not a modelled attribution. Control matched on prior-year category spend, trip frequency and store cluster. Confidence interval from 500 placebo tests on the pre-period.

Where the return came from

Select a channel for detail

Channel	Exposed hh	Media spend	Incremental sales	iROAS

What this does not prove

The control is matched, not randomised. It balances the factors we can observe in the loyalty file; it cannot balance what the file does not record.
Every interval on this screen is the same interval, expressed in whichever unit the tab is showing. It is not sampling error alone: at 420,000 households per arm sampling error would be roughly four times narrower. It is the spread of the placebo distribution — the same comparison re-run against 500 control groups drawn where no campaign ran — which is what a matched design has instead of randomisation, and it is the wider and more honest of the two.
Channel reach sums to 528,000 against 420,000 unique exposed households, because 108,000 households saw more than one channel. The per-channel incremental figures sum to the total because each is estimated against its own matched control and the overlap is then apportioned proportionally to single-channel response. That is an attribution choice, not a measurement: a household reached by onsite and CTV contributes to both rows at a fraction of its lift. A different rule would move money between rows without changing the $33.9K total.
Eight weeks cannot separate genuine incrementality from purchases pulled forward. The next quarterly run will show whether the exposed group dipped afterwards.
CTV exposure is modelled at household level rather than observed. Its interval is wide enough that an iROAS of 1.1 is not reliably different from 1.0.
Households below the 100-household cell minimum are suppressed, so thin store clusters are absent from the comparison entirely.

What to do with this

Move the CTV budget. $4,100 returned 1.1× and the interval does not clear break-even. Onsite display returned 3.1× on comparable spend. Shifting that budget is the decision this screen exists to support.
Scale onsite, but test the ceiling. 3.1× at 268,000 households does not mean 3.1× at 500,000. Raise spend by half, not double, and re-measure before committing the rest.
Check for pull-forward next quarter. Eight weeks cannot separate incremental sales from purchases that would have happened anyway, slightly later. If the exposed group dips in the following quarter, this number was partly borrowed.
Take the 31% new-to-brand rate to the retailer — and be precise about what it is. New to your brand is not new to the category: most of those 31% were buying a competitor in the same aisle last month, which moves share between suppliers without adding a dollar to the retailer’s category. The number their media team actually responds to is new-to-category, and this run did not compute it. It is a separate template, it needs category-level purchase history rather than yours, and it needs their permission because it exposes competitor behaviour. Ask for it before the joint business plan, not during it.

Proof of this run Template, datasets, policy, enclave quote

runcr-9b42ef17 · 2026-10-09 15:04:33 UTC

templatecampaign-incrementality v3.0 · sha256 2d71ac…5b08

retailer datameridian.loyalty.tx-60w · sha256 4a81d0…bb02 · 1,841,206 households

brand datanorthfield.exposure.laborday · sha256 c3e9f1…7a44 · 1,480,902 rows

policymin-cell 100 hh · min-segment 2,000 hh · signed by Meridian Markets, 2026-08-14

query budget13 of 40 used this quarter · requester: j.alvarez@northfield

enclaveIntel TDX · quote verified · measurement 1f8c44…90de

outputsha256 b18de0…cc93

Demonstration screen built on synthetic data with the structure and distributions of a regional grocery loyalty file. Retailer, brands and figures are invented. What is real is the method: the matched-control design, the suppression rule, and the manifest format.