The Budget Scale-Down Test: DIY Incrementality
Your platform’s reported ROAS includes sales you would have made anyway. Retargeting reclaims demand that already existed, prospecting catches people already drifting toward checkout, and the attribution model staples credit to the last click it happened to witness. The number that actually decides whether to scale is incremental ROAS — revenue that exists only because the ad ran — and many teams never measure it because they assume it needs an expensive vendor study. It doesn’t. A disciplined budget scale-down against a clean baseline gives you a directional read for free.
For the neighboring economics, compare When MMM, MTA, and Incrementality Each Break Down and use Blended ROAS Is Not Incrementality: The Limits to validate the measurement decision.
Reported ROAS vs. incremental ROAS
Attribution answers a generous question: “of the conversions I can see, which ones touched my ad last?” Incrementality answers the only question that matters for budget: “how many of these conversions would not have happened without the ad?”
The gap between the two is the part of your spend that’s buying conversions you already owned. It’s widest exactly where reported ROAS looks best — warm retargeting, branded prospecting, lookalikes built on recent purchasers. A campaign reporting a 6.0 can have a true marginal contribution near 1.0 because it’s harvesting intent rather than creating it. You can’t see that gap in any dashboard. You have to provoke it.
Why you can run this without a vendor
The rigorous methods — geo holdouts, ghost-ad serving, platform conversion-lift studies — are real, but they want scale, clean geographic separation, or a managed setup many teams can’t trigger on demand. The scale-down test is a within-account quasi-experiment. You don’t withhold ads from a region or a randomized audience; you change one input — spend — and watch what the total business does. It’s less clean than a true randomized holdout, so treat the output as directional, not a certified lift number. But directional truth beats a precise lie, and the reported ROAS on a retargeting campaign is frequently a precise lie.
The method, step by step
1. Pick the unit and the suspicion. Choose the campaign or audience you suspect is over-credited — in many cases your highest reported-ROAS line, because that’s where the harvesting hides. Decide what you’re testing before you touch anything.
2. Establish a clean baseline. Measure 2–4 weeks of stable, blended performance before the change. Blended means account-level: total revenue over total spend, i.e. MER, not the per-campaign ROAS the platform reports. Incrementality only shows up at the account level, because pausing one campaign quietly reroutes demand into others. A clean baseline means no recent budget swings, no creative refresh mid-window, no promo distorting volume, and no learning-phase churn in the campaigns you’re holding steady.
3. Make exactly one deliberate change. Cut the target’s budget by a meaningful amount — roughly 30–50%. Too small a cut hides inside daily noise; too large a cut resets the learning phase and corrupts the read. The point is one clean lever, large enough to move the blended number above noise.
4. Freeze everything else. No new creative, no audience edits, no bid-strategy changes, no budget shuffles on the campaigns you’re using as your steady backdrop. Every other change is a confound you can’t subtract out later.
5. Hold and measure on total outcomes. Run the reduced state long enough to clear the optimization lag and accumulate real signal — broadly 2–3 weeks. Watch blended revenue, MER, and total conversion volume, not the platform’s per-campaign ROAS. The platform will faithfully report that the cut campaign is still “efficient.” That’s the illusion you’re testing.
The math
Incremental ROAS at the margin is just the change in total revenue divided by the change in spend:
Marginal iROAS = Δ blended revenue ÷ Δ spend
Worked read: you remove 40% of a campaign’s budget. The platform had it at a reported 4.0. Over a clean window, blended account revenue falls by roughly 1.8× the spend you pulled. Your incremental ROAS at the margin is about 1.8 — less than half the reported figure. The difference between 4.0 and ~1.8 is the incrementality tax you’d been paying to buy conversions you mostly already had.
Two outcomes worth naming:
- Revenue barely moves. You cut a third of the spend and the blended top line shrugs. Marginal contribution is near zero — that budget was almost pure harvesting, and it’s your first reallocation candidate.
- Revenue drops more than the spend you pulled. Marginal iROAS comes in above 1.0 and possibly above your break-even multiple. That spend is doing real work; the cut was a mistake and you restore it.
Both are wins, because both replace a guess with a measured slope.
What corrupts the read
| Confounder | What it does | Guard |
|---|---|---|
| Learning-phase reset | A budget cut re-enters optimization; early data is noise | Keep the cut moderate; discard the first several days |
| Seasonality / demand shift | Outside demand moves and gets blamed on your change | Compare against the immediate baseline, not last quarter |
| Baseline contamination | A creative or audience change sneaks in mid-test | Freeze every other lever for the whole window |
| Signal lag | The optimization engine needs enough recent conversion signal to re-stabilize | Allow 2–3 weeks before reading |
| Delta too small | The change hides inside daily variance | Cut enough to move the blended number visibly |
| Demand re-routing | Paused demand reappears in another campaign | Always measure blended/MER, never the single campaign |
The re-routing point is the one many teams miss. Cut a retargeting campaign and a chunk of those buyers simply convert through prospecting, organic, or direct instead. Per-campaign ROAS will tell you the retargeting line “lost” revenue. The blended line tells you the truth: much of it never left.
Reading the result honestly
This is a directional instrument. Treat the output as a band, not a decimal. A marginal iROAS that lands “clearly below break-even,” “around break-even,” or “clearly above” is enough to act on — chasing a number to two places implies a precision a single quasi-experiment doesn’t have. Benchmarks like “enough recent conversion signal to stabilize” or “20–40% of spend frequently turns out non-incremental” are planning ranges to orient you, not laws and not numbers to quote back as fact. Your account has its own slope; the test exists precisely because nobody else’s average predicts it.
Run the cycle on your suspected harvesters first, then rotate. Done quarterly, a handful of scale-down reads rebuilds your budget around marginal contribution instead of reported credit — which is the difference between scaling what works and scaling what merely takes credit.
This is exactly the kind of slope Bach AI watches for, flagging when a high reported-ROAS line looks like harvesting rather than incremental lift — though it stays read-only until you approve any change.
The takeaway
You don’t need a vendor or a perfect holdout to know whether your spend is incremental. Pick your most suspicious campaign, cut its budget 30–50%, freeze every other lever, hold for 2–3 weeks, and divide the change in blended revenue by the change in spend. The slope you get back is a directional incremental ROAS — measured, account-specific, and free. Reported ROAS tells you what the platform wants to claim. The scale-down test tells you what your spend is actually buying.