Last week I closed with a concrete promise: "how standard models end up giving paid channels credit for revenue that was going to come in regardless."
Today I'll deliver on that promise with a number that surprised us when we first measured it, and with the explanation of why it shows up exactly where it shows up.
The phenomenon is clearest during Black Friday, during the launch of a product that goes viral, during a window when a brand appears in the press. Paid spend rises and organic demand rises at the same time, both responding to the same underlying market attention. The model receives the two movements as a single observable signal of revenue, and it has to split the credit between the two sources.
Even at this level of description, you can see why the math matters.
What the math does by default
A standard MMM resolves the ambiguity in the following way. When it observes revenue and spend rising in the same period, the first hypothesis is causal: spend drove revenue.
So where is organic demand in the equation? The model estimates it from prior history. But if that baseline was estimated under conditions different from the current period, the baseline will look too low.
Paid ends up receiving credit for revenue the baseline should have absorbed. Paid gets over-reported.
Think about it
Picture an audio recording in a room where several people are speaking at once. The microphone captures a single combined waveform. If you have to assign words to people, you can try a few things.
• Assume the person closest to the microphone said more.
• Split evenly.
• Cross-reference with video, tone of voice, the context of the conversation.
What matters is that the recording alone does not decide. The assignment that appears in the report reflects the assumption you used to disambiguate.
That is exactly what happens with paid and organic at peak. The equation alone does not decide; the structural assumption does.
The correlation problem at peak
Under normal conditions this phenomenon stays contained. Paid and organic do not move together with as much intensity. The baseline has time to stabilize, and the decomposition the model picks does not drift too far from the true one, even if only by chance.
At peak, the scenario changes.
Organic demand rises from seasonality, press mentions, social buzz. All of it enters the model at the same time. And that is before counting the paid media campaigns running over the same window.
The model prefers to attribute the peak lift to the variable it has structured most clearly, because that is the only path through which it can adjust parameters and improve fit. Attribution gets biased toward paid channels at exactly the windows where the cost of being wrong is highest.
When that biased measurement comes back into the system as the basis for the next budget decision, the next allocation gets built on a portrait of the channel that is more generous than reality can sustain.
Identifiability, now seen from measurement
In the previous issue I described identifiability as a property of the equation: the same data series can have infinitely many equally plausible decompositions, and the answer that appears in the report ends up depending on the model's internal assumptions more than on the raw data.
I noted earlier that measuring organic and paid is a structural choice. If the model's structure was built for conditions where paid and organic move independently, the default choice over-allocates to paid.
But if the structure has priors that reflect how marketing actually works in real life, with seasonality, demand cycles, and sensitivity to external events, the default choice changes.
The stronger the correlation between paid and organic in any given period, the more weight the structure carries in the final number. The ROAS you see is partly a property of the data, partly a property of the model.
3x the error in measurement
Here is the number I promised.
Across ten independent datasets, the average error of standard models against the truth was 3x that of a model built with priors appropriate to each brand. The methodology lives in the OMEN paper.
How many brands are making decisions based on that gap? That is the number that surprised us.
On BFCM specifically, the direction of the error is consistent. Standard models tend to inflate the credit paid channels receive, and the effect compounds when that measurement passes into the next layer of the system, which is the recommendation for how to allocate the next dollar.
This is not a problem you can fix by adding more data. The structure has to change for the answer to change.
A deeper look at the numbers from the paper
Measured against ten independent datasets, in conditions where the right answer was known because we built it into the experimental design:
• Measurement: standard models tend to over-attribute revenue to paid channels. Average absolute error against ground truth is on the order of 3x that of Prescient's model.
• Forecasting: when predicting incremental lift during peak periods, the error of standard models lands at roughly 2x to 4x that of Prescient's, depending on the baseline against which it is compared.
• Optimization: average error against the true optimum lands at 32-45% for standard models and at 5.6% for Prescient's. At BFCM, standard models recommended up to 81% overspend versus the optimum, while Prescient's model stayed within roughly 1%.
Today's biased measurement is the basis on which tomorrow's forecast and next week's recommendation get calculated.
What to look for in your own report
If you want to see whether this bias is operating in your own data, there are three questions worth asking your team this week.
• During the last observed peak, what percentage of incremental lift ended up attributed to paid versus the organic baseline? If the answer is that paid took 70% or more of the lift in a brand with active press, viral launches, or strong category seasonality, the baseline estimation deserves a second look.
• What happens to paid attribution when you compare a peak week to a week immediately before or after? If reported paid efficiency in peak looks meaningfully better than in adjacent weeks, part of that improvement may be absorbing demand that was already rising for reasons unrelated to spend.
• How does the model respond when you adjust the organic baseline explicitly? A model whose budget recommendation barely changes is leaning on a rigid decomposition. A model that responds with proportionate changes is one whose structure is acknowledging the ambiguity and resolving it with priors, not by default.
These three questions do not require a model change. They are diagnostic. They show where the current structure is placing the weight, and where that placement stops being appropriate.
The Halo
Exclusive insights, every week.
Subscribe to The Halo for sharper marketing thinking.