Programmatic Media Buying, Explained With Real Numbers

What Programmatic Media Buying Actually Means

Programmatic media buying is the automated buying and selling of digital ad space through software, not people. Instead of a media buyer emailing a publisher for a rate card, an algorithm bids for each individual ad impression in the time it takes a page to load, usually under 400 milliseconds. Amazon DSP, Google DV360 and The Trade Desk are all programmatic platforms; they differ in what inventory they reach and what data they use to decide a bid.

The mechanics are always the same three pieces. A demand-side platform (DSP) represents the advertiser and decides what to bid. A supply-side platform (SSP) represents the publisher and decides who gets to bid. Real-time bidding is the auction that runs between them, impression by impression. Programmatic is the process; display, video and audio are the ad formats riding on top of it, which is where most explainers get muddled — display can be bought programmatically or the old-fashioned way, and programmatic buys video and audio just as often.

One clarification worth stating plainly: if you landed here because you're a delivery driver researching Amazon's DSP program, that's Delivery Service Partner, a courier franchise, and this page won't help you. This is about Demand-Side Platform, the ad-buying kind. Same three letters, unrelated businesses.

How It Works, With Real Numbers

Walk one campaign through the pipeline. An advertiser sets a budget and targeting rules inside a DSP. A shopper loads a page or app; the publisher's SSP auctions that single impression to every connected DSP in milliseconds. Each DSP scores the impression against its advertiser's rules — has this shopper bought a competing product, do they match the target audience, is the placement brand-safe — and submits a bid. Highest bid wins, the ad renders, and the auction resets for the next impression, millions of times a second across the exchange.

Here's what that looks like on paper. Across 30 advertisers on Amazon DSP in July 2026, the media bought produced 78.4 million impressions at a $4.00 CPM, a blended $1.42 cost-per-click, and 57,137 attributed purchases at a blended $5.49 cost per acquisition, with 20.1% of those purchases from shoppers new to the brand, for a 6.04x return on ad spend. Every one of those numbers is a lever: raise the CPM ceiling and you win more of the good inventory but pay more for it; tighten the audience and CPA usually drops but volume drops with it too.

That scale only means something next to the seat count behind it. Full Circle and reMKTR run the numbers above are drawn from 30 of them in one month, not a single hero account.

What a Programmatic Buy Actually Costs

Programmatic spend breaks into layers, and almost every published price only shows you one of them. The advertiser is quoted a media budget, but that budget gets split before it ever reaches an impression: what the publisher gets paid, what the platform keeps for running the auction, what third-party data costs if you use it, and what a management fee costs if someone else is running the account.

The layer that decides your real cost is almost never the one printed on the homepage. A flat base fee looks affordable until you learn the percentage-of-spend rate sitting behind it — at $100,000 a month in spend, one percentage point is $12,000 a year, and that figure rarely appears next to the base fee in marketing copy. This isn't an argument against percentage-of-spend pricing; plenty of legitimate shops charge that way, including managed services generally. It's an argument for asking what the percentage actually is before signing anything.

The Mistake Most Buyers Make

The single most common mistake in programmatic buying is trusting last-click attribution to prove the ads worked. Last-click credits whichever touchpoint happened right before a purchase, so a DSP impression that ran a shopper past a product a dozen times before a sponsored ad closed the sale gets zero credit for the exposures that built the intent. That's not a fix you can make by picking a different attribution window — it's a structural flaw. Last-click cannot prove incrementality, and it never could.

The only ways to answer whether display spend added sales that wouldn't have happened anyway are holdouts, running the campaign against a comparable group that sees no ads, and matched controls, comparing similar audiences with and without exposure. On Amazon specifically, that means reconciling DSP and sponsored ads inside Amazon Marketing Cloud, because without deduplication the same purchase gets counted once by each channel and the combined number is fiction. This is a mistake worth admitting rather than hiding: it's easy to report a strong DSP ROAS that, once reconciled in AMC, turns out to be mostly the same conversions sponsored ads had already claimed. The fix isn't better creative. It's checking the overlap before believing the number.

When The Numbers Don't Add Up

If a programmatic number looks wrong, work the pipeline backward before touching the campaign. A CPA that suddenly doubles is usually one of four things: the audience segment refreshed and got broader, a competitor's DSP started bidding harder for the same inventory and your win rate dropped, a creative went stale and click-through fell, or the reporting window changed and you're comparing unlike periods.

If a fix doesn't work, don't stack a second fix on top of it. Roll the first change back, confirm the number returns to baseline, then test the next hypothesis alone. Stacking changes is how accounts end up with a strong quarter nobody can explain and nobody can repeat. And if a setting you meant to check is already correct — frequency capping, brand safety exclusions, a supply path already restricted to a private marketplace — say so and move on. Re-toggling something that isn't broken wastes a testing cycle you don't get back.