Scope summary signal-steering. When three measurement methods assess the same Google Ads channel differently, that is a steering question, not a numbers question. Evidence-backed: Triangulated measurement and steering system developed as a prototype. Testable against known ground truth (whether the methods capture the value). Use cases: assembling methods, hypothesis tests and anomaly detection. Not evidenced: Production readiness. Functioning at scale on real data. Validation on real spend (deliberately out of scope).

Commercial · Marketing Measurement

signal-steering

When three measurement methods assess the same Google Ads channel differently, that is a steering question, not a numbers question.

Objective

Which figure do you trust for a Google Ads budget decision when three methods contradict each other?

The test case became a triangulated measurement and steering system: platform attribution, geo incrementality and marketing mix modelling applied to the same data. The objective was not a finished product, but a robust answer to a real budget-steering question.

Approach

Triangulating measurement on synthetic data with known, fixed ground truth.

A representative measurement and steering problem was reconstructed: platform attribution, geo incrementality and MMM on the same data; steering against incremental value rather than the platform figure; plus detection of faulty signals. The synthetic dataset is methodologically necessary. Only a known true value makes it possible to test whether the methods capture it. No real, third-party or live spend data.

Why build a prototype

A domain expert can now go much further than only a few years ago — as a means to an end, not an end in itself.

Developed with agentic AI: data generation, MMM (PyMC-Marketing), geo incrementality (DiD/GeoLift), allocation logic and interface were created within weeks, without a development team. Not a production-ready system, but far enough to test the steering question on a running model. The decision about which methods to triangulate and how to read the signal remained with the human.

Evidence-backed

  • Triangulated measurement and steering system developed as a prototype
  • Testable against known ground truth (whether the methods capture the value)
  • Use cases: assembling methods, hypothesis tests and anomaly detection

Not evidenced · out of scope

  • Production readiness
  • Functioning at scale on real data
  • Validation on real spend (deliberately out of scope)
Data (synthetic)
PMax + Meta · BigQuery
Budget steering
allocates the next euro
Three methods measure the same data
Platform (last click)
fast, overestimates
Geo experiment
true lift, robust
Marketing mix model (MMM)
long-term impact
Result They contradict each other. The robust signal comes from geo experiments plus the model.
Market solutions cover only one method each
Platform & bidding
Smart Bidding · PMax
Meta · Skai · Marin
Geo experiments
GeoLift · Measured
Haus · INCRMNTAL
MMM tools
Meridian · Robyn
Recast · Mutinex
Blended SaaS (Northbeam, Triple Whale) combines two methods but remains a black box: no access to your own steering logic.
Green = developed in-house Grey = market solutions & raw data

The entire system in one view. Below it, the decisive points step by step.

The problem

The most common figure lies

Every budget decision depends on one figure: what did this channel truly add? The platform's standard answer is Last-Click. For PMax, it reports an iROAS of 5.02. The true figure is 2.50. The platform claims revenue that would have arrived without it. Following that figure doubles down on a channel worth half as much.

PMax · the same channel, three measurements
The dashed line is the fixed ground truth. Last click reaches far above it; the experimentally grounded methods capture it.
The actual point

How can 2.50 be “ground truth”?

In reality, no one knows the true value. What a channel added is a counterfactual that never took place. Measurement therefore always happens in the dark. The prototype creates a world in which the truth is fixed: synthetic data with a deliberately defined effect. This makes measurement testable for the first time.

This is not circular. The methods are not given the answer; they must infer it blindly from the data. The world is fixed, not the result. As in a flight simulator: the physics are fixed, but the pilot still has to fly. A method that misses 2.50 here will miss it in reality as well, only without anyone noticing.

Maturity instead of a polished pass

A good system knows when it does not know

There was no experiment for calibrating Meta. The model then estimates 10.05, clearly wide of the mark. Rather than concealing this, it reports the blind spot itself through its own convergence diagnostic (R-hat ✗). The steering implication is concrete: the next geo test belongs on Meta.

The point is not the model. It is the decision behind it: which figure do you trust when three methods contradict each other, and where should the next budget go? The prototype makes this decision visible and testable before real money depends on it.

Contact