How to Tell Whether Display Is Incremental
There's no single check that settles this — start with a cheap directional signal (new-to-brand rate and assisted-conversion count), then run an actual holdout or geo-lift test on any display line that signal flags as worth investigating. Attributed ROAS alone never answers the question; it only tells you display preceded a sale, not that it caused one.
What this looks like in a real account
Step one: check new-to-brand rate before anything else
New-to-brand share is the cheapest early signal available, because it's already in your standard reporting. A display line with a low ROAS but a high new-to-brand rate is a plausible acquisition channel being judged on the wrong metric; a display line with a low ROAS and a low new-to-brand rate is a weaker candidate for the incrementality-defence argument. In our own DSP book, off-Amazon exchange inventory returned a strong 6.05x ROAS but only 15.9% new-to-brand, while Amazon's own shopping surfaces returned a lower 4.90x but 42.0% new-to-brand — two placements doing genuinely different jobs, distinguishable from this cheap check alone, before running anything more expensive.
Step two: pull the assisted-conversion count
Next, check how often the display line appears in a converting path without being the final touch — an assisted-conversion query in Amazon Marketing Cloud. A high assist count is a reason to investigate further, not proof of incrementality on its own, but it narrows the field: a display line with both a low ROAS and a low assist count has little evidence supporting it at all, while one with a low ROAS and a high assist count is the strongest candidate for the next, more expensive step.
Step three: run the actual test on the strongest candidates
Once the first two checks have identified which display lines are worth the effort, run a holdout or geo-lift test specifically on those — not on the whole account at once, which spreads test power too thin to get a clean read on any single line. A geo-lift is often the more practical choice for display specifically, since it doesn't require the platform to support clean audience-level randomisation the way a DSP holdout does, and it works across a combined DSP-and-sponsored-ads read if that's the actual question.
A worked example of the full sequence
Take a display line running at 2.1x last-touch ROAS — well below the account's blended average. New-to-brand check: 38% NTB, well above the account average, a positive early signal. Assist check: the line appears in 41% of converting paths where it wasn't the final touch, another positive signal. Geo-lift test, run on markets carrying this specific line: exposed markets generate $95,000 in sales against a matched control's implied $81,000, an incremental lift of $14,000 against a test spend of $9,000 — a 1.56x incremental ROAS, lower than the account's headline attributed numbers elsewhere but genuinely positive and causally supported, which is a materially stronger case than the 2.1x last-touch figure alone ever provided.
The common mistake, including ours
The mistake is stopping at step one or two and treating a favourable new-to-brand or assist signal as if it settles the question, skipping the actual causal test because the earlier signals already told a flattering story. We've recommended defending a display line's budget on new-to-brand and assist evidence alone before, under time pressure in a renewal conversation — reasonable directional evidence, but not the same strength of claim as a completed holdout, and we should have said so more clearly in the room rather than letting the directional signals carry more confidence than they'd earned.
Why the sequence saves budget, not just time
Running the cheap checks first isn't just faster — it keeps a genuinely weak display line from consuming a full test cycle's worth of held-back spend and analyst attention that could have gone toward a stronger candidate. A line with a poor new-to-brand rate and a low assist count is unlikely to reward the cost of a proper holdout, and skipping straight to a full test on every underperforming line, rather than triaging first, is a slower and more expensive way to reach the same conclusions this three-step sequence gets to faster.
When the test comes back negative after positive early signals
This happens, and it's worth taking seriously rather than explaining away. A display line can show a genuinely high new-to-brand rate and a high assist count while still failing a properly designed incrementality test — meaning the shoppers it's reaching were largely going to become new-to-brand and complete that path anyway, just through a different touch. If a well-powered test genuinely contradicts the earlier signals, trust the test; those earlier checks were always directional shortcuts, not proof, and their whole purpose was to identify candidates worth testing properly, not to substitute for the test itself.
| Step | Cost | What it tells you |
|---|---|---|
| New-to-brand rate check | Free — already in standard reporting | Whether the line plausibly recruits new buyers |
| Assisted-conversion count | Low — one AMC query | Whether the line shows up in converting paths |
| Holdout or geo-lift test | Higher — requires design and a test period | The actual causal answer — did it cause sales |
Which one you should actually pick
Any advertiser with AMC access can run this exact three-step sequence themselves at no additional platform cost, and none of the three steps requires proprietary tooling. reMKTR runs it on every display line before recommending a budget defence or cut, having learned to stop resting on the cheaper early signals alone after doing exactly that under deadline pressure once, as part of the discipline behind Full Circle's $500M+ in managed Amazon spend across 100+ brands.
Shortlist on the job, not the feature grid. Pull your search-term report for the last 90 days and total the spend against terms that produced no orders — 33.6% on the account above. Then ask each vendor on your list what they would do about it in week one, and see who answers with a process rather than a screenshot.
Common questions
Can I judge display incrementality from ROAS alone?
No — attributed ROAS only shows that a display exposure preceded a sale, not that it caused one. Use new-to-brand rate and assist count as cheap early signals, then confirm with an actual holdout or geo-lift test.
How many display lines should I test at once?
Prioritise the strongest candidates first, based on new-to-brand and assist signals — testing too many lines at once spreads statistical power too thin to get a reliable read on any single one, and produces weaker evidence for every line tested rather than strong evidence for a few.
What if my incrementality test contradicts a positive new-to-brand signal?
Trust the test — the earlier signals are directional shortcuts meant to identify worthwhile candidates, not substitutes for a properly designed causal test.
Is a geo-lift or a holdout test better for measuring display specifically?
A geo-lift is often more practical for display, since it doesn't require the platform to support clean per-shopper randomisation the way a DSP holdout does, and it can measure a combined DSP-and-sponsored-ads effect if that's the actual question.
We show the method before the number.
Claim the free auditRead next
- Skai Pricing: Five Published Tiers, Read CloselyPricing · skai pricing
- Acorn Alternative: Replace the Capability, Not the LogoAlternative · acorn alternative
- Pacvue Pricing: No Public Number — What to AskPricing · pacvue pricing
- Skai vs Pacvue: Contracts, Not Feature GridsHead to head · skai vs pacvue