Amazon DSP Incrementality: How to Prove Programmatic Spend Actually Created Sales

Your DSP dashboard says 6x ROAS. Your CFO asks a reasonable question: how many of those sales would have happened without the ads? Most brands can't answer it. The dashboard was never built to.

That gap is the single most expensive measurement problem in retail media right now. Attributed ROAS credits DSP for every purchase inside the attribution window, including purchases from shoppers who were already going to buy. Retargeting a customer who has your product in their cart is the cleanest example. The ad gets credit. The sale was never at risk.

Incrementality is the correction. It asks what changed because the ad ran, not what got credited after it ran. Roughly 71% of advertisers now name incrementality as their most important retail media KPI, and it's the reason a growing share of DSP budget conversations happen with finance in the room rather than marketing alone. If you're running or evaluating Amazon DSP management at a meaningful spend level, this is the measurement question that decides whether the channel keeps its budget next year. If you want it answered for your account rather than in the abstract, book a call with our team.

Why ROAS Overstates What DSP Did

Sponsored Ads fire on intent. Somebody types a query, your ad appears, they buy. The causal story is short and mostly believable.

DSP fires on identity. It reaches a shopper because of who they are and what they've browsed, across Amazon, Prime Video, Fire TV, Twitch, and open-web inventory. The shopper may buy three weeks later after four other touches. Attribution assigns that sale somewhere, and inside DSP's own reporting it tends to assign it to DSP.

Three structural reasons attributed ROAS runs high on DSP:

Retargeting harvests. Your highest-ROAS audience is usually people who viewed a product and didn't buy. A meaningful fraction were coming back regardless.
View-through counting. DSP credits impressions that were served and never clicked. That's defensible for brand measurement and indefensible as a proxy for cause.
Long windows. A 14-day window captures more purchases than a 7-day window. The extra purchases aren't more incremental, they're just more captured.

None of this makes DSP a bad channel. We run it across our portfolio and it earns its place. It means the number on the dashboard is the wrong number to defend a budget with.

What Does Incrementality Actually Measure?

Incrementality measures the difference in purchase behavior between shoppers who saw your ads and comparable shoppers who didn't. Everything else is bookkeeping.

The practical form is a holdout test inside Amazon Marketing Cloud. You define an audience eligible to see your DSP ads, suppress the ads from a randomly assigned slice of it, let both groups run for a fixed period, then compare purchase rates. The difference is lift. Divide the incremental revenue by the spend and you get incremental ROAS, which is a smaller and far more honest number than attributed ROAS.

If you haven't set AMC up yet, our Amazon Marketing Cloud guide covers the access and instance basics. This post assumes you're already in.

The Three Test Designs, and What Each One Proves

Not every question needs a holdout. Picking the wrong design wastes a quarter.

Design	What you do	What it proves	Cost to run
Randomized holdout	Suppress ads from a random share of an eligible audience for a fixed window	True causal lift for that audience. The strongest evidence available.	Real. You give up sales from the held-out group.
Geo split	Run DSP in one matched set of regions, pause it in another	Channel-level lift including off-Amazon and halo effects	High. Needs scale and clean region matching.
Exposed vs unexposed analysis	Compare shoppers who happened to see ads against those who didn't, after the fact	Directional signal only. Selection bias inflates it.	Low. No suppressed spend.

The third design is where most brands stop, because it's free and it produces a flattering number. It's also the one we trust least. Shoppers who got exposed to your ads are frequently shoppers Amazon's models already judged likely to buy. Comparing them to everyone else measures the targeting model, not the ads.

Run the after-the-fact analysis to decide whether a real test is worth it. Don't put it in a board deck.

How to Run a DSP Holdout Test in AMC

Pick one audience, not the whole account. Retargeting and prospecting have completely different incrementality profiles. Testing them together returns an average that describes neither.
Size the holdout honestly. A 10% holdout on a small audience produces a control group too thin to read. We'd rather hold out 20% of one audience for six weeks than 5% of everything for two.
Fix the window before you start. Four to six weeks covers most consideration cycles in consumable and mid-ticket categories. Longer for high-consideration products.
Freeze everything else. A pricing change, a Prime event, a new Sponsored Brands campaign, or a listing refresh mid-test will contaminate the read. This is the step brands skip and then wonder why the result is incoherent.
Measure purchases, not clicks. Query AMC for purchase rate, units, and revenue per shopper in each group. Include downstream purchases on non-advertised ASINs.
Report incremental ROAS alongside attributed ROAS. Show both. The gap between them is the most useful diagnostic you'll produce all quarter.

Expect incremental ROAS to land well below attributed ROAS. That isn't a failure of the campaign. It's the first time you've seen the real number.

The Halo Problem: Incremental Sales Land on ASINs You Didn't Advertise

A DSP campaign pushing one hero product routinely lifts the rest of the catalog. A shopper sees the ad, lands on your brand store, and buys a different size, a bundle, or an adjacent product entirely. Amazon publishes no official halo formula, so this revenue is invisible unless you go looking for it.

In AMC you find it by connecting ad exposure to all downstream brand purchases rather than to the advertised ASIN alone. AMC's purchase-signal lookback now extends several years, which makes upper-funnel work far more measurable than it was two years ago.

We've found halo is where the argument for DSP is usually won or lost. A campaign that looks marginal on advertised-ASIN revenue often looks clearly positive once catalog-wide purchases are included. Brands that never run the halo query end up cutting DSP budget on incomplete evidence.

When Incrementality Testing Isn't Worth It

Being straight about this matters more than selling the method.

Skip the test if your DSP spend is small enough that the holdout group won't reach statistical significance. Below roughly $10,000 to $15,000 a month, DSP generally isn't producing enough signal to optimize on, let alone to split into test and control. Our post on DSP minimum spend covers where that floor actually sits now that Amazon removed the self-serve minimum.

Skip it if you're in a launch window. Holding out impressions while trying to establish rank is the wrong trade.

Skip it if you can't freeze the variables. A test run through a chaotic quarter produces a number you'll have to caveat into uselessness.

And be realistic about what a single test buys you. It measures one audience, in one window, for one catalog. It doesn't generalize to next quarter's creative or a new category. Incrementality is a habit, not a certificate. The brands that get value from it re-test one audience a quarter and build a picture over a year.

What This Looks Like in Practice

Across our client portfolio, the pattern is consistent: retargeting shows the highest attributed ROAS and the lowest incrementality, prospecting shows the reverse, and the correct budget answer is almost never the one the dashboard implies. Teams that reallocate on attributed ROAS alone tend to pile budget into retargeting until it saturates, then conclude DSP stopped working.

The same discipline applies off Amazon. We use holdout logic on paid social for the same reason, which we covered in our Meta ads incrementality testing post.

For a sense of what disciplined ad decisions produce when the measurement is right, our strategic budget increase case study walks through an established brand that grew sales 109% in a single month by widening its targeting and sales window rather than adding budget blindly.

Frequently Asked Questions

How long should an Amazon DSP incrementality test run?

Four to six weeks for most consumable and mid-ticket categories. Anything shorter risks missing the consideration window; anything much longer invites contamination from seasonality and catalog changes.

What's a good incremental ROAS for Amazon DSP?

There's no universal number, and anyone quoting one is guessing at your margin structure. The useful benchmark is your own contribution margin: if incremental ROAS clears the point where an extra dollar of spend returns more than a dollar of gross profit, the channel is working. Compare it against your blended efficiency target rather than against someone else's case study.

Do I need Amazon DSP managed service to run incrementality tests?

No. AMC access comes with DSP regardless of whether you run self-serve or managed. What you need is somebody who can write the queries, size the holdout, and keep the test clean for six weeks.

The Short Version

Attributed ROAS tells you what got credited. Incrementality tells you what you caused. If DSP is a real line in your budget and you've never run a holdout, you don't yet know which one you've been reporting. Run one audience, one window, one clean test, and find out.

If you'd rather have an operator design and run it with you, book a call with our team or read how we structure Amazon DSP management engagements.