One question that can come up when working with a Marketing Mix Model is why the totals shown for individual segments do not always add up perfectly to the total shown under All Segments.
For example, you may see conversions attributed to All Segments that are higher than what you get when you manually add together each individual segment. You may also notice that Non-Paid appears to have little or no activity in the UI even though you know organic or non-paid conversions exist in the underlying data.
At first glance, that can look like a data sync issue or an attribution bug.
In reality, it comes down to how the model and the resulting UI work together to create a usable predictive model.
What the UI is doing
The easiest way to understand this is to look at the order in which the data is processed.
- All Segments first combines every row by channel, including both paid segments and Non-Paid traffic.
- The system then applies an active-channel rule. In general, a channel needs to have spend greater than zero and meet a minimum share of total spend in order to be included in the model.
- Paid channels such as Paid Search, Paid Social, or Display generally have spend associated with them, so they survive this filtering process.
- If Non-Paid conversions are also associated with those same channel names, those conversions can become part of the overall totals because the channel itself qualifies as an active channel.
- When you look at Non-Paid by itself, however, those rows have no media spend. Because the same active-channel rules are being applied, those channels may no longer qualify for inclusion.
- Each individual paid segment also only contains its own rows. When you manually add those segment totals together, the Non-Paid conversions that were included in the All Segments view are not picked back up.
That means All Segments is not necessarily the same thing as “the sum of every individual segment.”
Instead, it represents the model’s view of eligible channels after the underlying rows have been combined and the active-channel rules have been applied.
What happens to conversions from completely non-paid channels?
There is another layer to this.
Some Non-Paid conversions may be associated with channels such as Referral, Email, Direct, Other, or similar zero-spend sources.
Because those channels do not have qualifying media spend, they can also be excluded from the modeled All Segments total.
This means there can actually be three different numbers in the underlying data:
- The raw total number of conversions across every row
- The conversions included in the All Segments modeled channels
- The conversions shown when individual paid segments are added together
Those numbers serve different purposes and should not necessarily be expected to match.
Is this a reporting attribution model or a data bug?
This is an important distinction.
The difference is not necessarily caused by missing data, a broken segment filter, or a data synchronization issue. It can simply be the result of the active-channel rules interacting with the way Non-Paid traffic is stored in the underlying dataset.
We could technically change the product rules so that zero-spend conversions behave differently. For example, we could choose to always display Non-Paid conversions separately, exclude them entirely from paid channels, or create additional rules around how they appear in All Segments.
However, there is a reason we need to be careful about doing that.
Paid media does not always result in a conversion that is directly recorded as coming from a paid source. A user may first interact with an advertisement and later return organically, through direct traffic, or through another non-paid source before converting.
If we automatically remove all Non-Paid conversions from the modeling process simply because there is no spend directly associated with that row, we risk removing some of the downstream impact that paid media may have helped create.
Why the model takes this approach
The important thing to remember is that an MMM is fundamentally a predictive modeling tool, not a traditional reporting tool.
A reporting tool is usually designed around reconciliation. If you have five categories, you generally expect those five categories to add perfectly to the total.
A predictive model has a different objective.
Its job is to understand the relationship between media investment and business outcomes well enough to help answer questions such as:
- What happens if we increase or decrease the overall media budget?
- What happens if we move spend between channels?
- Which channels appear to have the greatest opportunity for incremental investment?
- How might different media allocations affect future conversions, leads, or sales?
Because of that, the way data is prepared for the model should prioritize the accuracy, stability, and usefulness of the model itself, even when that means the UI does not behave exactly like a traditional reporting dashboard.
TL;DR
If individual segment totals do not add up exactly to All Segments, that does not automatically mean conversions are missing.
In many cases, the difference consists of real Non-Paid conversion rows that are treated differently depending on whether the system is evaluating all channels together or looking at an individual segment.
The underlying data is still there. The difference comes from applying the model’s active-channel rules to different groups of rows.
That distinction is important when using MMM results: the goal is not necessarily to make every reporting total reconcile perfectly. The goal is to build the most useful predictive model possible for understanding how media investment impacts future performance.