Programmatic 101: How Programmatic Media Buying Actually Works

What Programmatic Actually Means

Programmatic advertising is buying and selling ad space by auction instead of by phone call or insertion order. A piece of software called a demand-side platform (DSP) bids on an ad impression the instant a page or app loads, a supply-side platform (SSP) represents the publisher selling that impression, and an ad exchange sits between them running the auction. The whole cycle — request, bid, win, serve — happens in the time it takes a page to render.

That's the mechanism. Programmatic buying and programmatic media describe the same thing from two angles: buying is what the advertiser does through the DSP, media is what actually runs once the auction closes — display banners, video pre-rolls, connected TV spots, audio ads. None of that is a separate ad format. It's a buying method that can carry almost any format.

A data management platform, or increasingly a clean room like Amazon Marketing Cloud, feeds the DSP the audience information it bids against — who this shopper is, what they've bought before, whether they've already seen this ad this week. Strip away the acronyms and programmatic is three things: an auction, a bidder, and the data that tells the bidder what to bid.

One Auction, Step by Step — With Real Numbers

Here's what happens on one impression, then what a real book of business looks like once you run millions of them.

A shopper opens an app. The publisher's SSP sends an ad request — anonymized signals about that shopper — to an exchange. The exchange fans it out to every connected DSP. Each DSP decides in milliseconds whether the impression is worth bidding on and how much, based on the advertiser's targeting and budget. Highest bid wins, the ad renders, and the cycle resets on the next page load. No human picks the winner.

Numbers make this concrete. Across 30 advertisers running Amazon DSP in July 2026, the book bought 78.4 million impressions at a $4.00 CPM — what it cost, on average, to win a thousand of those auctions — with a blended cost-per-click of $1.42. Downstream of the clicks: 57,137 attributed purchases at a blended cost-per-acquisition of $5.49, 20.1% of them from shoppers new to the brand, producing a 6.04x return on ad spend. CPM tells you what the auction cost. ROAS and new-to-brand rate tell you whether winning those auctions was worth it. Most beginner explainers stop at the auction and never connect it to that second number.

Why Last-Click Attribution Can't Prove Any of This Worked

Here's the part most beginner guides skip: winning the auction and getting the click doesn't prove the ad caused the sale. Last-click attribution hands credit to whichever touchpoint happened right before checkout — often the touchpoint that would have converted anyway. It can't separate a shopper who bought because of the ad from one who was already buying and just saw it on the way.

The only honest way to answer that is to hold a group back. A holdout — a matched slice of the audience that sees no ads, or a matched control group with equivalent characteristics — lets you compare what actually happened against what would have happened anyway. The gap between the two is the real, incremental effect. Everything else is correlation dressed up as a result.

There's a second counting problem specific to Amazon: DSP and sponsored ads both run against the same shopper, and each platform's own reporting will happily claim credit for the same purchase. Reconciling in Amazon Marketing Cloud, where impression and click-level data from both sit in one place, is the only way to see whether display added a sale sponsored ads wouldn't have gotten, or just took credit for one it already had.

The Mistakes Most Buyers Make — Including Ones We've Made

A few mistakes show up in nearly every account.

• Buying CPM as if it were quality. A cheap CPM on junk inventory isn't a bargain, it's impressions nobody sees. CPM tells you what you paid, not what you got.
• Running programmatic with no holdout. Without a control group there's no way to separate display's real contribution from what would have sold anyway. It's the most common reason a display line item gets cut for 'not working' when nobody actually measured it.
• Double-counting across DSP and sponsored ads. Adding each platform's self-reported conversions and calling the total revenue overstates results, sometimes badly, because both can claim the same purchase.
• Leaving frequency caps too loose. We've done this — a cap set for reach early in a campaign, left unreviewed for weeks, ends up serving the same shopper the same ad a dozen times while budget that could have reached someone new gets eaten by repetition.

When the Numbers Look Bad: What to Check First

Before you touch a campaign that looks like it's failing, check these:

• Is the holdout big enough to trust? A small holdout can show a false swing in either direction. Check the sample size before believing the result, good or bad.
• Did a setting already change? Frequency caps, bid strategy, and budget pacing get adjusted by someone else on the account more often than people expect. Check the change log before assuming the strategy failed.
• Does the attribution window match the sales cycle? A seven-day window on a considered purchase will always look worse than the campaign actually performed.
• If the first fix didn't work, don't reverse it blind. Check whether it was even given time to run against its own measurement plan before rolling it back. A lot of 'it didn't work' is really 'we didn't wait.'

A useful discipline, whether or not you use a managed service: write down the evidence behind a change, the measurement plan that will judge it, and the rollback trigger that ends it — before it runs, not after. It turns gut-feel into an actual decision.