Amazon DSP Agency: What the Job Is and How to Pick One

Two businesses share this acronym, and only one is on this page

Amazon uses DSP for two unrelated things and the search results mix them constantly.

• Demand-Side Platform — the programmatic advertising product. Brands and agencies buy display, video and audio placements on Amazon's own properties and across third-party sites and apps. That is this page.
• Delivery Service Partner — the courier franchise programme, where independent owners run vans, routes, drivers and dispatch.

If you came looking for delivery work or the franchise programme, logistics.amazon.com is where to go. The collision is Amazon's, not yours. Everything below is about media buying.

What the weekly job actually involves

A DSP agency is not a reporting service, and the difference shows up in what happens between reports.

• Audience construction. Building and maintaining segments from Amazon shopping behaviour — retargeting, in-market, lifestyle, competitor conquesting — plus your own first-party lists where they exist. The highest-converting audience we work with is people who added to cart and did not buy, including those who never bought from the brand at all, because a large share of cart abandonment on Amazon is interruption rather than rejection. That is a recurring pattern across our managed DSP accounts rather than a guaranteed result on yours.
• Supply decisions. Where impressions run, and more importantly where they stop running.
• Creative production and rotation. Display and video units in the right sizes, refreshed before fatigue. This is real production work, not a checkbox on a scope.
• Pacing and bid strategy against a target derived from your margins rather than a category benchmark.
• Measurement. Reconciling DSP against sponsored ads so the two stop claiming the same order, and testing whether display caused sales rather than observed them.

There is also a triage discipline that separates experienced buyers from new ones. Two DSP failures exist and they need opposite responses: a line item that will not spend, and a line item that spends and will not convert. For under-delivery the order is base bids and caps, then frequency, viewability and geography constraints — each loosened with the advertiser's agreement, because each is a brand-safety decision. For delivery without conversion the order is the placement report first, then bids, then creative, then the pixel. Starting with creative is how a month gets spent on the wrong problem. That is standing practice across our managed DSP accounts.

DSP is not three placements — one month touched 134 supply sources

Most explainers describe Amazon DSP as a handful of placements: the home page, the detail page, maybe Prime Video. That picture is why so many brands cannot work out what happened to their money.

In a 31-day pull across 27 advertiser seats in summer 2026 our book delivered on 134 distinct named supply sources, and the performance spread between them was wider than the spread between advertisers. That figure is computed directly from the Amazon DSP API for that window and that advertiser count — a portfolio aggregate, with no single advertiser identified, and not a claim about our whole book or about any other period.

Two examples of how far apart those sources sit. Off-Amazon exchange inventory returned more per dollar than Amazon's own shopping surfaces — 6.05x against 4.90x — but found far fewer new customers, at 15.9% new-to-brand against 42.0%. They are doing different jobs: the exchange converts people already in the market, Amazon's own surfaces recruit. Judge both on return alone and you will quietly starve acquisition. Meanwhile Alexa device inventory delivered 34.9% of all impressions for 3.0% of the spend, at a $0.32 CPM. It is the cheapest inventory in the account and the easiest to misread, because it inflates impression counts in any report that does not break out supply source.

So the single most useful thing to ask any DSP agency in a pitch is not for a case study. It is: show me a supply-source report from a live account, with spend, return and new-to-brand by source. Anyone buying DSP without one is buying an average of things that behave nothing alike.

The click trap, and why cheap clicks are the expensive mistake

Here is the finding from that same pull that changes how a report should be read.

Third-party mobile in-app video generated 48.3% of every click in the book and 3.5% of the sales. It ran at a 1.56% click-through rate against a portfolio average of 0.28%, and above 20% on some exchanges — a rate that is a thumb, not an intent signal. Those clicks cost $0.33 each, which is exactly why the segment looks attractive on a cost-per-click report, and they carried a $26.94 cost per acquisition against $6.83 across the portfolio. Verified from the API for 27 advertisers over those 31 days, stated as a portfolio aggregate.

The format, not the audience, is doing it. Same apps, same devices, same audience classes: display units in third-party mobile apps returned 6.29x while video units in the same inventory class returned 1.46x, at $5.52 against $26.94 per attributed purchase.

Two consequences worth carrying into a vendor conversation. First, before anyone concludes that programmatic video does not work on Amazon, separate Amazon's own video surfaces from third-party in-app video — they are not the same product. Second, any optimisation target built on clicks or click-through rate will walk your budget straight into that segment and report success while doing it. Ask a prospective agency what they optimise toward and what they exclude, and ask them to name a supply source they have blocked and why.

Seats, minimums, and the floor we will tell you about

Access to Amazon DSP is not open self-signup for everyone, which is the reason this category of agency exists at all. There are three routes.

• Amazon's managed service. Amazon runs the campaigns. Their own product page states this option typically requires a minimum investment of USD 50,000, varying by country. It suits advertisers with limited programmatic experience who want Amazon's own team on it.
• Self-service. You hold the seat and control everything after account creation. This needs a trader who lives in the platform, not a marketing generalist with a spare afternoon.
• An agency holding seats. Amazon DSP is available to brand advertisers, agencies and tool providers. Agency minimums vary and are generally more flexible than the direct managed route. Ask each one for theirs in writing.

Now the part almost nobody publishes. Below roughly $10,000 to $15,000 a month, we tell brands to put the money somewhere other than Amazon DSP. Under that level the audiences do not accumulate enough signal to separate a working line item from noise inside a reasonable window, and the reporting overhead is disproportionate to the budget. That is standing guidance we give prospects rather than a platform rule, and it is one of the few answers that costs us the sale.

The related question is time rather than money. Three weeks is the point at which most brands ask to pause, and it is close to the worst moment to do it, because the audience models are still resolving and pausing discards the learning as well as the spend. On one account a supplement brand asked to stop after three weeks and roughly $6,000, at which point its return had quadrupled inside a Prime week that had raised costs across the whole platform. One account, one occasion — but the useful thing to negotiate is not a contract length, it is a learning window agreed in advance.

Fee structures are usually a percentage of media spend, sometimes with a base. reMKTR charges a percentage of media spend and states it before you sign. What matters is that you know the number, what it is calculated on, and where it steps down.

What reMKTR is, with the window attached to every number

We run Amazon DSP as a managed service on our own seats, sold on outcomes rather than dashboards. We are part of the Full Circle group — a full-service Amazon management company with $500M+ in managed spend across 100+ brands — and we hold 109 live Amazon DSP advertiser seats.

The figures below come from a single live API pull covering 27 advertiser seats over 31 days, 15 July to 14 August 2026. That scope is not decoration. Different windows and different advertiser sets produce legitimately different numbers, and a vendor quoting one period's return next to another period's acquisition cost is telling you something that is not true about their own book.

• 5.12x return on $732,916 of spend across those 27 seats
• $6.83 cost per attributed purchase, across 107,376 purchases
• 31.4% new-to-brand
• 197.7 million impressions, a $3.71 CPM and a blended $1.31 cost per click

The number worth carrying into a budget meeting is not the return multiple. It is the $6.83, because that is the figure a brand can hold against what it pays to acquire a customer anywhere else.

One more shape from the same pull, because it decides where review time should go: 18.3% of live line items consumed 80% of the spend, and the one in five line items that produced no attributed purchases at all burned just 0.7% of the budget. DSP waste is small and visible because you choose every line item. Sponsored-search waste is automatic and enormous because you do not choose the search terms. The two channels need completely different controls, and a partner who runs them the same way is running one of them badly.

Five questions that separate agencies, and where others beat us

Every agency will show you a case study with a large multiple on it. These get past that.

• Show me a portfolio number with its window and advertiser count. A return multiple without a denominator is not evidence. Ask for the median account as well as the average.
• Show me a supply-source report. Spend, return and new-to-brand by source, from a live account.
• How do you separate DSP from sponsored ads? If both channels report the same purchases, your blended return is inflated and nobody has noticed. That reconciliation happens in Amazon Marketing Cloud or it does not happen.
• When did you last run a holdout, who designed it, and what did it show when it failed?
• Who is on my account weekly, and how many others do they hold?

Now the part where we do not have the field to ourselves. Tinuiti runs a published incrementality practice — geo splits and audience holdouts, productised under its own name. That is substantially the argument we sell as our differentiator, they publish it openly, and any prospect comparing us should raise it. Where the question stays useful is one level down and it applies to us equally: who designs and computes the test, and is the design written into the scope of work before spend starts? A holdout run by the party being paid on the result is not really a holdout.

Podean is the other name worth knowing. It bought Ad Advance in February 2026 explicitly for retail-media and DSP depth, added Amerge in April 2026, and reaches 260 staff across 17 countries. If your brief spans many markets and languages under one contract, that footprint suits you better than ours does. And Amazon's own managed service is the right answer if you want Amazon's team on it and can clear the stated minimum.

Three times not to hire a DSP agency at all

We would rather say this here than four months into an engagement.

Your sponsored ads are leaking. Display amplifies whatever the account already does. If a large share of search spend is going to terms that produce no orders, that budget returns more there first, and display gets more efficient afterwards because the retargeting pool is cleaner. Dr. PPC is built for exactly that job and it is the honest first stop.

Your inventory cannot support the demand. Driving display traffic to items that go out of stock mid-flight wastes the spend and costs rank that has to be bought back at a worse rate. If stockouts, storage surcharges and fee bands are what is really eating margin, no media partner fixes that — Dr. Stock is the sibling product for inventory, fees and supply chain.

You have a capable in-house trader. Then buy the seat, not the service. You will pay less and move faster, and a managed engagement will feel like a layer between you and your own account.

There is a middle option too. If you want this capability as a product your own team operates rather than a service you buy, Dr. DSP is the same work packaged the other way round, with the first 30 days free. Orbit is included with it, and Orbit is not a DSP — do not read it as a substitute for the platform.