There’s More to Retail Media Than You Think

The digital ad industry loves to talk about retail media, and it should. It delivers for commerce platforms (i.e., retailers, marketplaces) and advertisers, while ticking all the right boxes in terms of identity, attribution and privacy-safe signal.

For commerce platforms, it’s high-margin revenue that can even be accretive to GMV and a positive user experience. For advertisers, it’s valuable inventory with users in a discovery and purchase mindset, rich first-party data, and logged-in users. That makes it an ideal environment for deterministic attribution, targeting data application, predictive modeling, and real-time optimization. Forecasts on the TAM are aggressive with some at $101 billion in 2022, and well above those of CTV (~$20 billion in the U.S.).

What’s missing in all the dialogue is a clarification that the models deployed by the most successful and profitable commerce platforms are outcome driven (return on ad spend) with user-level predictions, real-time optimization, and self-service tools for simple access to large counts of advertisers. Advanced machine learning models enable this, and while it falls under the generic moniker of retail media it needs to be highlighted as a distinct approach.

Amazon.com, Shopee, Etsy, eBay, Flipkart, Grab, Alibaba, Coupang, Expedia, DoorDash, and Mercado Libre to name a few, leverage this superior approach to retail media, and it’s paying off. To understand this, let’s double-click into Amazon’s ad platform, which drove $38 billion in 2022. First off, it's fueled by roughly 2 million Amazon merchants (or suppliers) in the Amazon Marketplace. These are small to midsize advertisers looking for simple and automated tools to promote their products. And they require performance-driven ad solutions where users can easily understand the revenue impact of their spend.

Given a large merchant base that requires simplicity and measured performance, Amazon built strong machine learning models trained on first-party data, which make impression-level decisions about which ad to serve which user and at what cost. Its model is closer to the automated ad systems of Meta and Google than it is to the more commonly understood approach of retail media, which is human-designed awareness campaigns with manually selected data elements and post-campaign analysis.

In order to evaluate how these models differ, there are a few vectors to consider:

• Self-Serve or Managed Service: How is the campaign set up and managed? Is it done by the advertiser (or agency on its behalf) with simple, scalable self-serve tools in an external interface provided by the platform, or is it done by a media sales team on an internal tool?
• ROAS or Broad Awareness: Is the goal to drive measurable revenue within the next month or so or to get your message in front of large groups of consumers to influence potential purchases in the future that aren't necessarily tied to a given campaign?
• Machine Predicted or Manual Targeting: Are machine-learned models predicting an individual’s or impression’s likelihood of hitting a goal, or are humans selecting data segments (first or third party) and other targeting elements?
• Real-Time Optimized or Post-Campaign Analysis: Are the user predictions or targeting models continuously updating while a campaign is running or are they analyzed afterward and leveraged for future campaigns?

The more common approach by commerce platforms is to provide campaigns that are managed service, ROAS or awareness, manual targeting, and post-campaign analysis. This is effectively the traditional digital publisher sales model, but with retail media’s deterministic measurement and considerably improved targeting data. It works well since it taps into an established sales motion, and can pull from shopper marketing dollars. Since this can also be extended offsite, it’s not even constrained by the retailer's inventory. It’s not a surprise and makes sense that several are launching this solution.

The approach Amazon and a few others take is to enable campaigns that are self-serve, ROAS, machine predicted, and real-time optimized. This works well when there's a large group of suppliers (or merchants) that actually provide products and if there's a strong relationship between them and the retailer. The merchants are a captive set of advertisers ready to engage with. The greater depth and breadth, the greater the incentive for them to advertise, especially when there's a clear measurement of return. This is especially pronounced in marketplaces where the entire SKU base is from merchants, and there's an existing merchant portal tool where they interact with the retailer.

While both models offer ROAS, it’s important to call out that the machine predicted and real-time optimized approach just performs better. If performed well enough and transparently reported, companies may consider ROAS as a cost of goods (i.e., core to how each product is distributed) and less of a marketing expense. Humans are only so good at selecting the optimal data to apply. Even when they leverage machine learning or data science techniques to do so, they're ultimately applying the learning to coarse groups of users (segments are typically linked together with boolean "or" statements and have ID counts of hundreds of thousands to millions). Machine predicted makes an inference on each impression based on real-time and historical user events (e.g., browse, search, purchase, etc.). Also, these models continuously and automatically learn and can update on an hourly basis. They iterate and adjust more in a few hours than humans will do over several campaigns. This isn't even chess vs. checkers.

These two retail media models aren't mutually exclusive. There are "awareness" ads on amazon.com, and outside of the core retail app Amazon sells awareness campaigns through a brand-facing sales team on Prime CTV and its DSP. Meta and Google obviously also sell awareness too, but those campaigns don't comprise the bulk of its revenue nor their advertisers. It doesn't seem to be the focus of their product and engineering investments — as seen by the development of Performance Max and Advantage+ Shopping.

Therefore, when you read about retail media or if you're a marketplace planning to implement ads, it's important to be cognizant of the different models and to distinguish their defining characteristics. Think about which approach is best given your business model and goals. To invest in "retail media" but to stop short of real-time optimization or machine learning is to abandon so much of the outsized promise of retail media. The companies that have and will continue to gain market share in this space certainly know these distinctions and have placed their bets accordingly.

Bill Michels is responsible for Moloco’s Retail Media Platform, where he oversees both product and business development functions.