How to Build a Retail Media Strategy That Proves Its Own Results

What a retail media strategy actually covers

A retail media strategy is the plan for how you spend money to get products in front of shoppers on and around a retailer's platform — Amazon, mostly, but the logic carries to Walmart, Target, and Instacart too. It has three layers most brands run in some combination: sponsored ads (search and product placements paid per click), DSP display (programmatic ads bought through Amazon's demand-side platform, on-site and off), and off-site amplification (using the retailer's shopper data to reach people elsewhere on the internet).

The strategy part isn't picking those three things — everyone picks those three things. The strategy part is deciding how much goes to each, in what order, and how you'll know afterward whether the money worked. Most guides to retail media stop at definitions and formats. The actual job is the third piece: proving the spend caused the sale, not just that a sale happened near it.

The budget layers, with real numbers attached

Budget splits should follow the shopper's stage, not a rule of thumb. Sponsored Products carries the bottom of the funnel — people already searching for something like your product. DSP display carries the middle: people who viewed your product or a competitor's and didn't buy yet. Off-site and streaming carry the top: people who don't know you exist.

Here's what that split looks like when you follow it through to real output. Across a 30-advertiser Amazon DSP portfolio measured in July 2026, the numbers ran like this: 78.4 million impressions at a $4.00 CPM, blending to a $1.42 cost-per-click across the whole book — the $0.41 CPC figure quoted everywhere in this category is online-video only, not the blended number. That spend produced 57,137 attributed purchases at a blended $5.49 cost per acquisition, with 20.1% of those purchases coming from shoppers new to the brand. Return on ad spend across the whole portfolio, not the best-performing line item, was 6.04x.

The reason to report the whole-book number instead of a highlight is that a single campaign can post almost any ROAS you want if you pick the right week. A strategy is judged on the blended number, because that's the number that pays the bill.

Why last-click attribution lies to you, and what to use instead

Last-click attribution assigns the sale to whichever ad the shopper clicked right before buying. It has one structural problem it can never fix: it can't tell you whether that shopper would have bought anyway. If someone searches your brand name, clicks a sponsored ad, and buys — last-click gives that ad full credit, even though the sale might have happened without it.

Two things actually answer the incrementality question. Holdout tests hold back a matched group of shoppers from seeing the ad and compare their purchase rate to the group that saw it. Matched controls pair similar audiences with and without exposure and measure the gap. Neither is exotic — they're standard experimental design borrowed from clinical trials, applied to media.

There's a second, quieter problem: sponsored ads and DSP display both claim credit for the same sale if you measure them in separate dashboards, because a shopper can be exposed to both before buying. Reconciling that — deciding which channel actually gets the credit rather than letting both claim it — has to happen in one measurement layer that sees both, not two systems that don't talk to each other.

The mistake almost everyone makes, including us early on

The most common mistake is treating the budget split as the strategy and stopping there. A close second — one we've made ourselves early in a client relationship — is reading DSP performance off last-click and concluding display 'isn't working' when in fact it's assisting sponsored ads sales that get all the credit under last-click rules. Cutting display on that evidence removes a channel that was doing its job; you just weren't measuring it correctly.

The other frequent error is copying a budget split from a case study without checking whether the underlying audience size supports it. A brand with a small addressable audience on Amazon will burn through a display budget's efficient reach fast and start paying for frequency nobody needs. The split that worked for a high-volume brand isn't automatically right for a smaller one.

When the numbers come back bad: what to actually do

When a retail media number looks bad, check three things before you touch the budget. First: is the number even measuring what you think? A ROAS drop can mean the campaign got worse, or it can mean a tracking or attribution-window change moved conversions somewhere else. Second: is this a setting that was already correct, and the 'fix' about to break it? Turning off a display line because ACOS looks high sometimes removes an assist channel rather than a wasteful one — check whether sponsored ads conversion rate also drops the week after you cut it. Third: if you already tried a fix and it didn't move the number, stop iterating on the same lever. A CPC increase that didn't lift impression share usually means the auction dynamics changed, not that you didn't bid high enough.

The general rule: don't act on a single week of last-click data. Look at a full attribution window, check whether the drop shows up across the whole portfolio or one campaign, and only then decide whether it's the strategy or the measurement that's wrong.

Where reMKTR fits

Everything above — the layers, the blended numbers, the holdout logic — works the same whether you run it yourself, through an agency, or through a managed service. reMKTR runs Amazon DSP as a managed service and reconciles it against sponsored ads inside Amazon Marketing Cloud, specifically to stop the double-counting problem described above, replacing last-click with holdouts and matched controls. We're part of the Full Circle group, which has managed more than $500M in Amazon spend across 100+ brands. That's one way to run this strategy. It's not the only way — a brand with the in-house team and AMC access to build its own holdout tests doesn't need anyone's managed service to do it right. The discipline matters more than who executes it.