The most reliable way to measure retail media is to model it inside a marketing mix model, alongside TV, social, search, price and in-store promotion, using sales from all channels rather than only the retailer’s own. Retail media network dashboards report attributed sales, not incremental sales. They can’t see what would have happened without the ad, and they can’t see effects outside their own walls. A mix model can do both, if retail media is set up properly as an input.
Why network reporting overstates impact
Retail media networks, from Amazon to supermarket loyalty programs, can connect ad exposure to purchases through their own customer data. That closed loop is useful. It’s also the problem.
Attribution is not incrementality. A sponsored product ad shown to a shopper who already had your brand in their basket gets credit for the sale. So does one shown to a loyal buyer who purchases every week. The dashboard counts those sales as ad-driven anyway.
Each network marks its own homework. Attribution windows, view versus click rules and definitions of “halo” differ from network to network. Adding up reported sales across Amazon, a grocery network and a drugstore network usually produces a total that no finance team would recognize.
The view stops at the retailer’s door. A grocery network can see sales in its own stores. It can’t see whether the same campaign lifted sales at a competitor down the road, or simply moved purchases from one retailer to another.
Retail media often rides on other activity. Retail media is frequently bought in the same weeks as a price promotion, a feature display or a TV burst. When they run together, the network claims sales the promotion or TV drove.
How to set retail media up in the model
Getting retail media right is mostly a data design problem.
Split by network and by format. Sponsored search on a marketplace behaves very differently from onsite display or offsite video bought through a retailer’s audience data. Lumping them into one “retail media” line hides those differences. Where volumes allow, give each major network and format its own variable.
Use delivery, not just spend. Impressions help the model separate genuine exposure from price changes in the auction. Two weeks with the same spend can buy very different reach.
Model sales by retailer. This is the step that unlocks the most insight. If your sales data is split by retailer, the model can estimate a network’s effect on its own sales and, separately, its spillover into other retailers. Hypothetically, a grocery network campaign might lift sales in that chain while pulling some from a rival chain. That’s a shift, not growth. National totals can’t show the difference.
Keep price and promotion in the model. Because retail media and trade promotion are often planned together, both must be present. Without promotion depth and timing, the model has no way to split credit, and retail media inherits the promotion’s effect.
Watch for collinearity. If retail media only ever runs during promotions, the data can’t tell them apart, no matter how good the model is. Varying the timing, even a few weeks of retail media without a promotion, gives the model the contrast it needs.
The offline halo
For brands selling in physical stores, a large share of retail media’s value may show up offline. A shopper sees a sponsored product on a grocery app, then buys it in store. Or a marketplace ad raises awareness that converts at a supermarket days later.
A network’s own reporting may capture part of this through loyalty cards. A mix model sees it from the other direction: it relates retail media delivery to total sales, including in-store sell-out and sales at other retailers. With carryover built in, it can also pick up effects that arrive a week or two after exposure. This is where cross-channel incrementality modeling earns its keep: it estimates what retail media added across the whole business, not just inside one retailer’s reporting.
The halo can be negative too. Retail media on one marketplace may cannibalize a brand’s own direct-to-consumer site. A model that includes all sales channels will show that. A network dashboard never will.
Interactions with TV, social and search
Retail media usually sits at the bottom of the funnel. It captures shoppers who are already in the aisle, physical or digital. Much of that intent was created elsewhere: a TV campaign, a creator partnership on social, a search ad the week before.
That means retail media’s measured return often depends on upper-funnel activity. A model that estimates channel interactions can show whether retail media works harder when TV is on air, which is exactly the insight a brand needs before shifting budget from TV to retail media.
A common trap follows from this. A brand sees strong retail media ROAS, cuts TV to fund more of it, and watches retail media returns fall a few months later, because the demand it was harvesting has thinned out.
Validating the numbers
A mix model gives a top-down estimate. Pair it with bottom-up checks where you can. Some networks offer holdout-based tests. Regional pauses, switching a network off in some store clusters but not others, also give a clean read. If the test and the model broadly agree, you can plan with confidence. If they disagree, that gap is worth investigating before the next budget round.
What to ask before the next retail media renewal
Is retail media split by network and format in your model? Are sales available by retailer and by channel, online and in store? Is promotion data in the model? Have you ever run retail media outside a promo window? And does the model estimate how retail media interacts with TV and social?
If the answers are yes, the network dashboard becomes one input among several, not the final word. That’s where retail media measurement should be.
Editor’s Note: The opinions expressed here by the authors are their own, not those of impakter.com — Photo Credit: DC Studio
