Attribution analytics · Validation system

Where the attribution model and reality disagree.

A working prototype that validates marketing attribution against measured causal incrementality. Synthetic data with known ground truth. Geo-lift estimation with cluster-robust standard errors. A truth-check that exposes which channels are over-credited, under-credited, or accurate within measurement noise. Self-evaluation with precision, recall, and F1, because attribution decisions are categorical and should be graded as such.

Comparison chart: model claim vs measured incrementality per channel.
Channel-only shares of conversions. Last-touch attribution model claims (charcoal) versus measured incrementality from geo-lift experiments (forest). The gap above each pair, in percentage points, is the truth-check finding for that channel.
Walkthrough · Episode 01

Hi. What you're looking at is a tool that does something most marketing teams don't: hold their attribution model accountable to causal reality. Attribution is the bookkeeping that says "this channel drove that conversion." Incrementality is the harder question of whether the conversion would have happened anyway. Most teams treat attribution as if it answered both. It doesn't.

The chart up top is the entire project compressed into one image. The charcoal bars are what a last-touch attribution model claims each channel contributed. The forest-green bars are what a geo-lift experiment actually measured. Where the two bars disagree, the model is misallocating credit. Direct mail on the far left is the most-misallocated channel in this run. Getting only 7.7 percent of credit when reality says it deserves 27.

Below this panel, four headline numbers. Below that, six cards. Each card is a self-contained walkthrough of one piece of how the system works, written in this same plain-language voice. There is no required order. Click whatever looks interesting.

Aggregate over-attribution
3.2×
Last-touch attributes 16,426 conversions to channels; geo-lift measurement supports 5,144. The model has no concept of baseline conversions.
Largest mis-allocation
−19.5pp
Direct mail is under-credited by nearly 20 percentage points. The model gives it 7.7% of channel credit; reality supports 27.3%.
Recall on over-credit class
0.85
Across 50 simulated experiments, the truth-checker correctly flags 85% of channels that are actually over-credited. Direction errors are nearly zero.
Narrative direction accuracy
100%
Claude-graded narrative correctness across 18 channel evaluations. Same-family judge caveat applies; magnitude correctness was also 100%.
The system, end to end

Six components, one validation pipeline.

Synthetic data builds the answer key. The geo-lift engine recovers truth from public data alone. The comparison layer flags where the attribution model and reality diverge. The narrative layer translates the gap into plain English. The self-evaluation harness grades the truth-checker the same way you'd grade a clause classifier: precision, recall, F1, threshold sweeps, calibration.