One of the most cited holdout experiments in digital advertising ran on eBay's own branded search spend. Researchers from Berkeley and Chicago Booth worked with eBay to pause its branded paid search ads across a set of markets and measured what actually happened to sales, rather than trusting what the ad platform's own reporting claimed. Naive attribution-style calculations on that same spend implied returns north of 1,600 percent. The experimental result, once the researchers controlled for the fact that people searching "eBay" mostly intended to land on eBay anyway, was a return of roughly negative 63 percent. The full study is published through NBER for anyone who wants the methodology, not just the headline number.
That gap between a claimed 1,600 percent return and an experimentally measured loss is not a tracking bug. It is the honest difference between what a platform can see and what a channel actually causes, and no amount of fixing pixels or deduplicating server events closes it. Attribution, even done well, measures touchpoints it can observe. It cannot measure what would have happened without them.
Marketing mix modeling is built for exactly that question. It is also, in most of the content written about it, sold as something every ecommerce brand needs immediately, which is not quite true. Here is what MMM actually does, what it costs in data and effort, and the point at which an ecommerce brand has actually earned the need for one.
What marketing mix modeling measures that attribution cannot
Attribution, whether last-click, data-driven, or a third-party attribution layer, works from the bottom up: it observes individual clicks, sessions, and conversion events, then assigns credit for each one to a touchpoint. That approach has a hard ceiling. It can only credit what it can see, which means it structurally undercounts brand awareness, TV, podcast ads, out-of-home, and increasingly anything on a browser or device where a consent signal, an ad blocker, or a platform's own privacy defaults break the chain between impression and purchase. Even dedicated attribution tools like Triple Whale inherit this ceiling, since they still build up from the same observable clicks and pixels GA4 does.
MMM works from the top down instead. It takes weekly or monthly totals, spend by channel, revenue, price changes, promotions, seasonality, and fits a statistical model that estimates each input's contribution to the outcome. It never needs to see an individual user, which is precisely why it keeps working as consent banners, ad blockers, and shrinking cookie windows quietly thin out what any pixel-based tool can see: there is no pixel to lose, no cookie to expire, no consent banner to decline.
The trade-off is resolution. MMM cannot tell you which ad creative underperformed this week the way a platform dashboard can. It answers a different, more strategic question: if this channel's budget moved up or down by a meaningful amount, what would happen to total revenue. That is a budget-allocation question, not a campaign-optimization one, and conflating the two is where a lot of MMM adoption goes wrong.
The data marketing mix modeling actually needs
Every serious source on this converges on a similar range: expect to need somewhere between 12 and 24 months of weekly spend and revenue history before a model has enough to work with, with more history helping the model separate genuine seasonality from channel effects rather than confusing the two. Recast's own guidance on data requirements for marketing mix models is a useful reference point here, since it comes from a vendor whose entire product depends on getting this threshold right, not from a general blog post.
The detail most beginner content skips: volume of history matters less than variation within it. A model can only estimate a channel's effect by watching what happens when that channel's spend changes. A channel held at a nearly constant budget for two years gives a regression almost nothing to separate its contribution from the overall revenue trend, no matter how many data points that flat line contains. A newer channel with nine months of real flights, pauses, and deliberate spend swings can be more useful to a model than three years of "set it and forget it" budgeting on an older one.
That single fact reframes how to prepare for MMM. The useful preparation is not waiting passively for a calendar to fill up. It is making sure the media plan itself creates the variation a model will eventually need: staggered flights instead of always-on budgets, documented promotional calendars, and at least occasional deliberate pauses on channels that never get turned off otherwise.
Who has actually earned the need for one
MMM is expensive to do properly, whether that cost shows up as a subscription to a modeling platform, an analytics team's time running the open-source alternative, or both. That cost is easy to justify for a brand spending seven figures a year across eight channels including offline media, and much harder to justify for a brand running two paid channels on largely flat budgets.
A rough, honest set of prerequisites before MMM earns its place on the roadmap:
- Multiple channels running simultaneously, ideally five or more, so there is an actual allocation decision to inform.
- Combined marketing spend high enough that a meaningful reallocation, five or ten percent of budget, represents real money worth the modeling effort to protect.
- A documented history of deliberate spend changes, not just steady autopilot budgets, since that variation is what the model actually learns from.
- Tracking and attribution fundamentals already solid. If platform-reported numbers don't reconcile with actual store revenue, fix that gap before adding a second, more complex model on top of the first, unresolved one.
Below that bar, the honest move is to keep the budget and the analytical effort on the layer underneath: clean event tracking, a dashboard built on blended MER rather than platform-reported ROAS, and periodic holdout tests on the one or two channels where a founder's gut and the platform's own reporting disagree most. An analytics audit is the fastest way to confirm that foundation is solid before spending anything on a modeling layer built on top of it. Our marketing analytics service sets up that measurement foundation and, for brands that have already cleared the prerequisites above, scopes what a mix model would take to build and maintain.
Building one: buy, use open source, or wait
Three real paths exist once a brand clears the bar above.
The first is a paid MMM platform. These wrap a modeling engine, usually a variant of Bayesian regression, in a managed interface with automated data ingestion, and they exist specifically because raw modeling output is hard to interpret without a background in statistics.
The second is open source. Google publishes Meridian, a Bayesian MMM package built by its own marketing science team, and Meta publishes Robyn, built the same way inside Meta's marketing science division. Both are free and actively maintained by the platforms that built them. The real cost of the open-source path is not the software license. It is needing someone who can clean the input data correctly and read a Bayesian posterior without over-trusting a single run.
The third path, and the right one for most brands reading this, is to wait and prepare. Spend the next year making sure spend and revenue data is clean and complete, introduce deliberate variation into channels that have been running flat, and keep validating attribution against reconciled store revenue in the meantime. A model built on a clean, well-varied twelve months of data will outperform one rushed onto flat, uninstrumented history regardless of which package or vendor runs it.
Where this leaves the ecommerce brand not ready yet
Marketing mix modeling is the right tool for a specific, later problem: proving whether the overall channel mix is right once attribution alone can no longer answer that question. It is not the right first investment for a brand still reconciling GA4 against Shopify, and no amount of statistical sophistication fixes a model built on numbers nobody has verified yet.
The realistic sequence is boring but works: verify tracking, fix attribution and reconciliation, build the spend variation a future model will need, then model. Skipping to the last step early does not save time. It just moves the same unresolved measurement problems one layer deeper, dressed up in more convincing math.
