> ## Documentation Index
> Fetch the complete documentation index at: https://docs.dema.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# How Causal factor attribution works
> Learn how Dema uses causal evidence from incrementality experiments and platform-wide benchmarks to correct attribution values and reveal the true impact of your marketing.
Causal factor attribution bridges the gap between **correlation-based attribution** (MTA) and **true incremental impact**. It adjusts your attribution values using real experimental evidence, or Dema's benchmarked incremental factors when no experiments are available yet.
## Why standard attribution isn't enough
Multi-touch attribution (MTA) models assign credit based on observed user journeys. While useful, they have two fundamental limitations:
1. **Correlation, not causation** - MTA measures which channels appear in a conversion path, but can't tell you whether those conversions would have happened without the marketing spend. A channel may look highly effective simply because it captures users who were already going to convert.
2. **Click-based tracking** - MTA relies on clickstream data, which means it systematically undervalues impression-based and awareness channels (e.g., display, video, influencer). Channels that build consideration without generating a direct click receive little or no credit, even when they play a significant role in driving conversions.
**Key question:** *Is this channel actually driving new sales, or is it just getting credit for sales that would have happened anyway? And are impression-based channels being undervalued?*
Causal factor attribution answers both questions by adjusting attribution values, whether from MTA or ad platform reporting, with evidence from controlled incrementality experiments.
***
## How it works
Causal factor attribution applies **calibration multipliers** to each marketing channel's attributed contribution, using either the MTA value or the ad platform's own reported value as the base. These multipliers are informed by real experimental data and reflect how much of the attributed value is truly incremental.
Dema maintains a **platform-wide benchmark** of incremental ROAS by channel, built from all finalized incrementality experiments across the platform. Even if your business has never run an experiment, you get a data-driven starting point for each channel based on what Dema has observed across similar channels and markets.
This benchmark is displayed as a **bell curve distribution** showing the expected range of incremental impact for a given channel.
As you run incrementality experiments, the distribution **narrows and shifts** to reflect your specific business. Dema uses a Bayesian statistical model that combines the platform-wide benchmark (prior) with your merchant-specific results (likelihood). More experiments with stronger statistical significance contribute more weight.
The result: your calibration recommendations become increasingly tailored to your business over time.
Based on the distribution, you set a calibration multiplier for each channel and choose whether to apply it to the MTA value or the ad platform's reported value. For example, if incrementality experiments suggest that Meta social campaigns deliver only 70% of what MTA claims, you'd set a multiplier of 0.7 with MTA-based calibration. For channels where the ad platform's reporting is more reliable than click-based MTA, you can calibrate from the ad platform value instead.
The system applies your multipliers to the chosen base values (MTA or ad platform). Channels with multipliers below 1.0 have their contributions reduced; channels above 1.0 are increased. **Daily totals are always conserved**: when one channel's contribution changes, the difference is redistributed.
***
## Benchmarked incremental factors
One of the most powerful aspects of Causal factor attribution is that **you don't need to run experiments to get started**.
Dema aggregates results from incrementality experiments across the platform to build a benchmark distribution for each channel. This gives you:
* **An expected incremental ROAS range** for channels like Meta, Google, TikTok, and others
* **A confidence level** that reflects how much experimental data backs the benchmark
* **A starting point for calibration** that's grounded in real causal measurement
The benchmark distribution uses a Bayesian model that weights experiments by their statistical significance (p-value). Experiments with stronger results contribute more to the benchmark. Outliers are filtered using standard statistical methods to keep the distribution robust.
As you accumulate your own experiments, the distribution transitions from a broad platform-wide benchmark to a narrow, merchant-specific estimate:
| Scenario | Distribution shape | What it means |
| ----------------- | --------------------- | ------------------------------------------- |
| No experiments | Wide bell curve | Based entirely on platform-wide benchmark |
| A few experiments | Narrowing curve | Combining benchmark with your early results |
| Many experiments | Narrow, precise curve | Driven primarily by your own data |
***
## Calibration types
Causal factor attribution supports two ways to calibrate a channel:
Multiplies the existing MTA-attributed contribution by your chosen factor.
**Use when:** You trust the relative distribution of MTA values across campaigns and just want to scale the channel's total contribution up or down.
**Example:** MTA says Meta drove €10,000 in gross sales. With a 0.7 multiplier, the calibrated contribution becomes €7,000.
Replaces the MTA value with the ad platform's own reported value, then applies the multiplier.
**Use when:** The ad platform's reported metrics are a better starting point than MTA (e.g., for channels where Dema's tracking has limited visibility).
**Example:** Meta reports €12,000 in attributed sales. With a 0.6 multiplier, the calibrated contribution becomes €7,200.
***
## Total conservation
When calibration changes a channel's contribution, the system ensures that **daily totals remain constant**. The difference (delta) is redistributed to designated **source channels**, typically direct and other unattributed traffic.
This conservation principle means calibration doesn't change your total reported revenue or profit. It only changes **how that total is distributed across channels**. If Meta's share goes up, direct's share goes down by the same amount.
**Why source channels absorb the delta:**
* Direct and unattributed channels capture organic traffic
* Their contribution is inherently less certain than paid channel attribution
* Redistributing to these channels is the most conservative approach
If the delta exceeds what source channels can absorb, the system automatically caps the adjustment to prevent any channel from going negative.
***
## How it connects to other Dema features
Causal factor attribution works alongside Dema's other measurement tools:
| Feature | Role | Relationship |
| --------------------------------------------------------------------- | ----------------------------------------------------------------------- | ---------------------------------------- |
| [Attribution models](/guides/fundamentals/attribution) | Provide the base MTA values that get calibrated | Input to causal factor attribution |
| [Incrementality testing](/guides/incrementality-testing/how-it-works) | Generate the experimental evidence that informs calibration multipliers | Source of causal evidence |
| [Marketing mix modeling (MMM)](/guides/mmm/how-mmm-works) | Models long-term channel effectiveness at a strategic level | Complementary strategic view |
| **Campaign-level attribution** | Distributes calibrated channel-level values down to campaign/ad level | Downstream consumer of calibrated values |
**Best practice:** Run incrementality experiments on your highest-spend channels first. This gives you the most accurate calibration where it matters most, while relying on benchmarked factors for smaller channels.