What Is Amazon DSP? (And What It Isn't)

Two Unrelated Things Share the Acronym 'DSP'

Type "Amazon DSP" into Google and you get two completely different answers mixed together. One is Demand-Side Platform — the advertising tool this page is about. The other is Delivery Service Partner, the franchise-style program where independent business owners run fleets of Amazon-branded delivery vans and hire drivers. Same three letters, nothing else in common.

If you searched "how much does Amazon DSP make" hoping to learn what a delivery driver or DSP owner earns, that's the courier program, and it's a separate business entirely — pay is set by the individual DSP owner, not Amazon, and varies by region and contract. Amazon doesn't publish owner profit figures, and we wouldn't invent one here. If that's what you meant, Amazon's own hiring pages cover driver pay and DSP ownership applications directly.

Everything below is about the advertising platform: what it buys, how it's priced, and what the numbers mean when a campaign runs.

What the Advertising Platform Actually Does

Amazon DSP is programmatic media buying. Instead of bidding on search terms the way Sponsored Products does, you buy display, video and audio ad space — on Amazon, on Amazon-owned properties like IMDb and Twitch, and on third-party sites and apps through the open web — targeted using Amazon's shopper and purchase data rather than keywords.

That's the core difference from Sponsored Ads: Sponsored Ads catch demand that already exists (someone searching for a product). DSP builds and re-engages audiences — shoppers who viewed a product but didn't buy, past purchasers due for a repeat, or lookalike audiences built from Amazon's own signal. It's a different job, not a better or worse one.

Access used to require a managed-service minimum; self-service DSP is now open to more advertisers directly through the Amazon Ads console, though it still assumes someone on your team can build audiences, set frequency caps, and read the reporting.

A Worked Example: What the Numbers Actually Look Like

Definitions are cheap. Here's what a real DSP book looked like across 30 advertisers we ran in July 2026: 6.04x return on ad spend, 78.4 million impressions at a $4.00 CPM, and a blended $1.42 cost-per-click. On the conversion side: a $5.49 blended cost per acquisition across 57,137 attributed purchases, 20.1% of them from shoppers new to the brand.

Read those together, not apart. A $4.00 CPM on 78.4 million impressions is roughly $313,600 in media spend at that stage of the book. A $5.49 CPA against 57,137 purchases means that spend is doing real acquisition work, not just impressions for their own sake — and the new-to-brand share tells you how much of it is expanding the customer base versus re-selling existing buyers. Any DSP report that gives you ROAS without also giving you new-to-brand percentage is hiding half the story.

These figures are one book, one month, scoped honestly — they're not a promise about your account. What they show is the shape of a healthy DSP campaign: CPM, CTR, CPA and new-to-brand rate moving together, not one metric carrying the whole pitch.

How Amazon DSP Pricing Works

Amazon doesn't publish a flat DSP rate card, and neither does almost anyone who resells it — pricing is usually one of: a percentage of managed ad spend, a flat platform or seat fee layered on top of media cost, or a quote set against your specific spend and scope. If a page you're reading quotes you a specific dollar figure with no tier attached, treat it as unreliable; these numbers change by region, by account size, and without notice.

Dr. DSP, the managed product this page's publisher runs, doesn't publish a number either — for the same reason. It's a demo, the first 30 days free, and pricing set on the call against your real media budget and scope, with Orbit, the reporting and management software, included at no extra cost. That's a structural choice, not a claim of being cheaper than anyone charging a percentage of spend — percentage-of-spend is a legitimate model, including one we use ourselves elsewhere.

The thing to ask any vendor, managed or self-service, is: what's the full fee stack at your actual spend level, not their lowest quoted tier?

The Mistake Almost Everyone Makes With DSP Attribution

Last-click attribution cannot prove a DSP campaign added anything — it counts a sale whenever DSP touched a shopper before purchase, even if that shopper was going to buy anyway. We've built last-click dashboards ourselves and only later realized they were counting the same purchase twice: once as a DSP view-through, once as a Sponsored Products click, on the exact same order.

The honest fix is a holdout or matched control: hold a comparable audience out of DSP exposure and compare what happens. Reconciling DSP and Sponsored Ads data inside Amazon Marketing Cloud is the only way to stop that double-count and see whether display actually moved incremental sales, or just claimed credit for demand Sponsored Ads was already capturing.

If your current DSP reporting only shows ROAS and view-through rate with no holdout and no AMC join, that's not proof of incrementality — it's a number that looks like proof.

What to Do When the DSP Numbers Look Wrong

Bad numbers happen. Before assuming the platform is broken, check the usual suspects:

• CPMs spiking suddenly: check for new competitors entering the same audience or a sudden narrowing of targeting that's shrinking available inventory.
• ROAS looks high but sales aren't moving: that's the last-click double-count above — check whether Sponsored Ads is getting credit for the same purchase.
• New-to-brand share is low: the audience is probably built too heavily on past purchasers; widen to prospecting segments.
• A test showed no lift: check the holdout was actually clean — frequency capping or audience overlap between test and control groups quietly ruins more incrementality tests than bad creative does.

Every change we make carries three things before it runs: the evidence behind it, a measurement plan, and a rollback trigger — because the fix not working is a normal outcome, not a failure to plan for.