Retail media reports the best return in the budget. Amazon, the grocers' networks and the marketplaces all show a closed-loop ROAS: ads served on the retailer's site, matched to purchases on the retailer's site. The number is real and it is measured inside one wall. It says nothing about whether those shoppers were on their way to buy anyway, nothing about the TV campaign that sent them to the marketplace, and nothing about the sales retail media drives in your own webshop. Marketing mix modelling puts retail media next to every other channel on one yardstick: incremental sales per unit spent, across all the places you sell. This guide covers the two conditions that make that work, and how to read the result.
Retail media is a channel like any other in an MMM: weekly spend, its own response curve, read against the same sales series as TV and Meta. Two conditions make the result meaningful:
Then the comparison holds: the retailer's ROAS and Meta's ROAS were never comparable, and the model's return per unit is.
Closed-loop reporting attributes a purchase to an ad when the shopper saw or clicked the ad before buying on the same retailer. It is accurate about that. It is silent about two things. First, the shopper who searched for your product on the marketplace had often already decided; retail media sits at the end of the path, where attribution rewards channels for being present rather than for being the cause. Second, the decision was frequently made elsewhere: a TV flight, a creator's post, a search on Google. The retailer's report credits none of that, and your Meta and Google reports do not see the marketplace sale at all.
An MMM looks at the whole picture from the outside. It relates total weekly sales, across webshop, stores and retailers, to weekly spend in every channel, after seasonality, price and promotions are accounted for. Retail media gets a curve, TV gets a curve, and the two are read on the same scale. The full guide to marketing mix modelling explains the method; this guide is about retail media specifically.
The outcome the model explains has to include the sales retail media drives. For a brand selling through its own site, stores and marketplaces, that means marketplace and retailer sell-out, weekly, in the same series as everything else. Most marketplaces and retail media networks report brand sales weekly; grocers report sell-out through their data services. Get the feed automated, because a model that learns from a sales series with a hole in it credits the missing weeks to the wrong channel.
Where the business wants to know where the sales landed, webshop and marketplace sales can be modelled as separate outcomes, at the cost of more data per model. A model on total sales gives the cleanest answer to "how much should we spend on retail media?"; separate outcomes answer "where did it land?". Our guide to MMM by business model goes through that choice for retailers and marketplaces.
| Retail media type | Spend data | In the model as | What to watch |
|---|---|---|---|
| Marketplace search and display (Amazon and similar) | The retailer's ads reporting, weekly, by campaign type | Its own channel; search and display split when each is big enough | Branded product search behaves like brand search: it captures demand other channels created |
| Grocer and retailer networks | Network reports, weekly | One channel per network when big enough; otherwise grouped as retail media | Spend is often bundled with trade terms; separate the media money from the trade money |
| Off-site retail media (retailer audiences on open web and social) | Network reports, weekly | Joins the on-site channel, or display, by what it behaves like | It reaches people away from the point of sale, so it carries over longer than on-site ads |
Two habits keep the spend honest. Separate media spend from trade spend: a joint business plan with a retailer often bundles promotions, listing fees and media into one agreement, and only the media belongs in this channel. And split branded product search from generic where the reporting allows, for the same reason brand search is split on Google: the two do different jobs on different curves.
Odins takes retail media spend in through the platform connections where the network offers one, among 600 plus integrations across search, social, programmatic, affiliate and retail media, and through a pipeline where it only comes as reports, so it lands as weekly spend in the same structure as every other channel.
Retail media sits closer to the purchase than any channel except brand search, which is exactly why its reported return is high and its incremental return is in doubt. The model answers in two ways.
First, through variation. Where retail media spend has changed level, the model can see whether total sales moved with it or whether the retailer's sales simply shifted from organic to paid. A channel whose marginal return is low and whose curve is flat is capturing demand rather than creating it, and the model will say so with a range.
Second, through the rest of the mix. When TV and retail media run in the same weeks, the model credits each by how sales moved relative to each one over the whole history, not by who was last in the path. A retail media channel that only ever spends in TV weeks, at a fixed share, cannot be separated, and the fix is a plan: a period where one moves and the other does not.
When the range stays wide, test
Retail media is a good channel to test, because it can be moved in steps without a relearning period. Hold it at a higher level for four to eight weeks in a planned window, with the rest of the mix steady, and let the model read the difference in total sales. The result calibrates the model, so the retail media estimate and the TV estimate stay on one scale instead of a test number sitting beside a model number.
The model gives retail media a response curve with a confidence range, an average return per unit spent and a marginal return at today's level. Compare on the marginal return, because that decides where the next unit goes. A high average and a flat curve means the channel is well funded and the next unit should go elsewhere. Retail media's carry-over is short, so the horizon matters less here than it does for TV. The comparison with TV still has to be made on a horizon where TV's brand share counts, or TV loses by construction. Brand effects in the model explains how that is handled.
What the model will not do is pick keywords, products or audiences inside the retail media account. That is the platform's job. The model sets the level and the share; the account team runs what happens inside it.
Odins builds and runs the model, collects the spend and the sales every week, and delivers a monthly recommendation with a confidence range per channel. Companies like CDON, Nettbil, Hyre and Megaflis run their marketing budgets on it.
Yes, on one condition: the sales retail media drives are in the outcome the model explains. With marketplace and retailer sales in the series, retail media gets a response curve read on the same scale as every other channel.
It is accurate about what it measures: ads on the retailer's site matched to purchases on the retailer's site. It is not a measure of incremental sales and it is not comparable with another platform's ROAS. Use it to run the account; use the model to set the budget.
A model on total sales will show whether retail media spend moves total sales, across the webshop and the marketplace. To see where the sales land, model webshop and marketplace sales as separate outcomes.
Separate the media money from the trade money before it goes into the model. Listing fees, promotions and trade terms are not advertising and should not be in the channel.
Platforms that take marketplace and retailer sales into the outcome, model each sales channel's media separately and collect the data for you. Odins builds the model this way for retailers and brands that sell through more than one channel.
Read the complete guide to marketing mix modelling, see which data connections are available, or book a demo and bring your retail media reports.