Reward Model
A movement-based approach to attribution from aggregated marketing signals
Marketing attribution often starts from a simple question:
Which channels appear to move with business outcomes?
In practice, this question is difficult to answer. Marketing signals are noisy, channels are measured in different ways, and customer journeys are often incomplete. Some channels have impressions, some have clicks, some have spend, some have searches, and some appear only as platform-reported or business-specific activity.
The Reward Model is designed for this environment.
It provides a high-level, movement-based way to estimate attribution from aggregated marketing signals. Instead of relying only on a traditional media-mix interpretation, it focuses on whether changes in channel activity are repeatedly informative with respect to changes in the target.
The goal is to build a practical attribution baseline from aggregated data, especially when the analyst wants to understand which channels behave coherently with conversions, revenue, or another business outcome.
Why a reward-oriented view is useful
Many attribution models focus on the average relationship between channel activity and outcomes. That is useful, but it can hide an important practical question:
When the target changes, which channels tend to move in a compatible way?
A reward-oriented model starts from this intuition.
If a channel repeatedly moves in a way that is aligned with the target, that channel receives more support. If its movements are repeatedly uninformative or inconsistent, it receives less support.
This does not require the model to observe every individual customer journey. It works from aggregated time-series data, making it suitable for privacy-friendly attribution and for channels that are not visible in user-level paths.
The problem with relying only on size
Large channels often dominate raw marketing data.
A channel with many impressions or clicks can appear important simply because it has more volume. But volume alone is not the same as contribution.
A channel can be large but weakly informative.
Another channel can be smaller but more consistently aligned with changes in the business outcome.
The Reward Model is designed to consider both:
- the amount of channel activity observed over time;
- how informative that activity appears with respect to the target.
This makes it useful when the analyst wants an attribution baseline that is not based only on scale.
What the Reward Model does
At a high level, the Reward Model works with aggregated time-series data.
The input data may include:
- conversions by day or week;
- revenue by day or week;
- impressions by channel;
- clicks by channel;
- spend by channel;
- searches;
- visits;
- email activity;
- platform-specific engagement metrics;
- marketplace or offline indicators.
The model evaluates how each channel signal behaves over time relative to the target.
It asks questions such as:
- Does this signal tend to move in a compatible direction with the target?
- Is the relationship stable enough to be informative?
- Is the signal useful after accounting for timing?
- How should multiple signals belonging to the same channel be combined?
- Which channels deserve more or less attribution support?
The output is a channel-level attribution baseline.
A movement-based attribution perspective
The core idea is simple.
When a business outcome changes, the model examines whether channel signals show compatible movement patterns.
For example:
| Day | Conversions | Channel A signal | Channel B signal |
|---|---|---|---|
| Monday | 100 | 1,000 | 500 |
| Tuesday | 120 | 1,180 | 470 |
| Wednesday | 115 | 1,120 | 530 |
| Thursday | 140 | 1,350 | 510 |
In this simplified example, Channel A tends to move in the same direction as conversions. Channel B is less clearly aligned.
The Reward Model does not rely on this simple table alone, but this intuition is central: repeated compatible movements provide evidence that a channel is informative for attribution.
The model is not trying to prove causality from one movement. It is looking for repeated, structured evidence over time.
Timing matters
Marketing effects are not always immediate.
A channel may influence the target on the same day, after a short delay, or across a short time window. This is especially common in channels such as paid social, display, email, affiliates, or upper-funnel activity.
The Reward Model accounts for timing by evaluating how channel signals align with the target over possible lags.
This helps avoid a common mistake: judging a channel only by whether it moves at exactly the same time as the target.
A signal may be useful even if its effect appears with a delay.
Multiple signals per channel
Modern marketing data rarely has one clean variable per channel.
A single channel can be represented by several signals:
| Channel | Possible signals |
|---|---|
| paid search | impressions, clicks, cost, searches |
| paid social | impressions, clicks, cost, engagement |
| sends, opens, clicks | |
| direct | direct visits, branded searches |
| affiliate | clicks, orders, commission |
| marketplace | platform views, product views, orders |
The Reward Model supports this kind of structure by separating raw input signals from final attribution channels.
This means that several signals can contribute to the same final channel-level result.
For example, google_ads can be represented by both clicks and impressions. The model evaluates the evidence from those signals and returns attribution at the google_ads channel level.
This makes the output easier to interpret and closer to the way marketing teams think about budget allocation.
Signal compression and scale control
Marketing signals can be highly uneven.
Some channels may have millions of impressions, while others may have thousands of clicks or a much smaller business-specific signal.
To avoid letting very large raw values dominate purely because of scale, the model uses a lightweight signal-compression step. This acts as a simple saturation-like mechanism: very large signal values are compressed, while the information contained in the original signal is still preserved.
This helps the model avoid over-crediting high-volume channels simply because they are large.
The result is a more balanced attribution baseline, where both scale and signal quality matter.
Reward Model versus Automatic MMM
The Reward Model and Automatic MMM both work with aggregated marketing data, but they emphasize different views of contribution.
| Perspective | Main idea |
|---|---|
| Automatic MMM | Focuses on predictive signal strength from aggregated channel data |
| Reward Model | Focuses on movement coherence between channel signals and the target |
Automatic MMM is closer to a media-mix attribution view. It looks for predictive relationships between marketing signals and the business outcome.
The Reward Model is more movement-oriented. It asks whether a channel’s changes are repeatedly informative with respect to changes in the target.
These two perspectives can be complementary. One is useful when the analyst wants an aggregated predictive baseline. The other is useful when the analyst wants to understand which channels behave coherently with the outcome over time.
A high-level workflow
At a conceptual level, the Reward Model follows a simple workflow.
Step 1 — Collect aggregated data
The dataset contains one row per time period.
Example:
| timestamp | conversions | google_clicks | google_impressions | facebook_impressions | email_clicks | direct_searches |
|---|---|---|---|---|---|---|
| 2024-01-01 | 120 | 340 | 12000 | 18000 | 210 | 900 |
| 2024-01-02 | 135 | 360 | 12600 | 17500 | 230 | 950 |
| 2024-01-03 | 128 | 330 | 11900 | 18200 | 220 | 910 |
Step 2 — Map signals to channels
Each input signal is associated with a final marketing channel.
Example:
| Signal | Channel |
|---|---|
| google_clicks | google_ads |
| google_impressions | google_ads |
| facebook_impressions | facebook_ads |
| email_clicks | |
| direct_searches | direct |
Step 3 — Describe the type of each signal
Each signal can also be described by its measure type.
Example:
| Signal | Measure type |
|---|---|
| google_clicks | clicks |
| google_impressions | impressions |
| facebook_impressions | impressions |
| email_clicks | clicks |
| direct_searches | searches |
This allows the model to work with mixed data sources and still produce channel-level attribution.
Step 4 — Evaluate movement evidence
The model evaluates how each signal behaves in relation to the target over time, considering timing and repeated movement patterns.
Signals that are more informative receive stronger support in the attribution baseline.
Step 5 — Produce channel-level attribution
The final result is attribution by channel and time period.
This output can be used directly as an aggregated attribution baseline, or it can become part of a broader unified attribution process when customer journeys are also available.
Why this approach is useful
It works without customer journeys
The Reward Model can be applied when user-level journeys are missing, incomplete, or unreliable.
This is useful in modern measurement environments where tracking is fragmented and privacy constraints are increasing.
It is intuitive
The model is based on an intuitive question: does the channel move in a way that is informative with respect to the target?
This makes the output easier to explain than some black-box attribution approaches.
It supports mixed signal types
Different channels can be represented by different kinds of signals.
The model can work with impressions, clicks, searches, visits, spend, platform activity, or business-specific indicators.
It can highlight smaller but informative channels
Because the model does not rely only on raw volume, smaller channels can still receive attribution support when their signals are consistently informative.
It can complement other attribution approaches
The Reward Model can be used as a standalone aggregated-data attribution baseline or as part of a broader attribution framework that also includes customer journeys.
Reward Model and unified attribution
The Reward Model is especially useful as one possible aggregated-data layer inside a unified attribution framework.
When customer journeys are not available, it can provide attribution from aggregated signals.
When customer journeys are available, the aggregated baseline can be combined with journey-based evidence, such as a Markov model, to create a more complete attribution view.
This is useful because aggregated data and customer journeys answer different questions:
| Layer | Question answered |
|---|---|
| Aggregated reward layer | Which channels move coherently with the target over time? |
| Journey layer | Which channels play an important role inside observed paths? |
| Unified view | How can both sources of evidence be combined? |
The Reward Model therefore provides a practical bridge between time-series attribution and journey-based attribution.
When should you consider the Reward Model?
You should consider this approach if:
- you have aggregated marketing signals over time;
- customer journeys are missing or incomplete;
- you want a movement-oriented attribution baseline;
- you want to compare predictive and movement-based attribution views;
- you have many channels with different signal types;
- you want to include smaller channels that may be informative even if they are not high-volume;
- you want an attribution baseline that can later be combined with customer-journey evidence.
It is particularly useful when the analyst wants to understand not only which channels are large, but which channels behave in a way that is repeatedly aligned with business outcomes.
A practical example
Imagine a company investing in:
- paid search;
- paid social;
- email;
- affiliates;
- direct traffic;
- marketplace activity;
- offline campaigns.
Some channels have clicks and spend. Others have impressions. Some are visible only through platform-reported activity. Customer journeys may be available for only a subset of channels.
The Reward Model can use the aggregated signals to identify which channels show the strongest movement evidence with respect to conversions or revenue.
This gives the marketing team a channel-level attribution baseline that can be used for:
- budget discussions;
- channel comparison;
- reporting;
- measurement diagnostics;
- unified attribution with journey data.
Conclusion
The Reward Model is a practical attribution approach for aggregated marketing data.
It focuses on movement coherence: how channel signals behave in relation to changes in the business outcome over time.
By combining signal quality, timing awareness, scale control, and channel-level aggregation, it provides a useful alternative to traditional media-mix thinking when the analyst wants a more movement-oriented view of contribution.
For companies working with fragmented data, partial tracking, and mixed signal types, the Reward Model offers a clear and flexible way to transform aggregated marketing activity into attribution insight.