MMM data quality: use blocking gates, not a reassuring average.
Evidence status
What was reproduced, and what it does not prove.
The CSV and Python verify the six checks, the three blocking failures, the illustrative average of 96.1% and the allocation prohibition. They do not measure a real quality, an MMM or a causality.
- Technical specification of the MMM data-quality register, specification in French (opens in a new tab)
v1, 31 August 2026 · CC BY 4.0 · specification in French · Method document written by marketing-mix.net. The figures are illustrative: they are not market measurements.
Key distinctions
Three conditions before calculating.
Source reconciliation
Spend must be matched with the financial source using documented rules for fees, taxes, discounts, credits and production. An unexplained gap becomes a risk of artificial return.
Semantic continuity
A stable channel name can hide a change of perimeter, tool or buying method. Taxonomy and measurement breaks are events of the model, not mere cleaning problems.
Traceable correction
Every imputation or reclassification keeps the initial value, the rule, the author, the date and the impact. An opaque correction prevents reproduction and makes future monitoring impossible.
Method
Building a quality control that can genuinely stop the model
- 01Define the controls
For each table, write dimension, formula, threshold, frequency, owner and action; separate informative control, alert and blocking gate.
- 02Replay the history
Apply the controls to every period and locate breaks, delays, duplicates, impossible values and reconciliation gaps; measure the effect of the corrections.
- 03Decide by use
A gap can be acceptable for an overall trend and blocking for a weekly return by platform; connect each defect to the use affected.
- 04Monitor new loads
Automate the stable controls, log the exceptions and trigger a review on a change of source, of definition or on a threshold breach.
The control table with three blocking gates
Six dimensions are controlled, but financial reconciliation, series breaks and taxonomic stability can each block the model on their own.
Teaching scenario: an average quality of 96.1% is still a no-go
In this fictional teaching register, the dataset covers 104 weeks. Completeness: 99.8%. Spend reconciliation: 98.7%. Taxonomy stability: 94%. Freshness: 92%. Seven series breaks are undocumented.
The simple average of the four percentages reaches 96.1%, but three gates fail: reconciliation required at 99%, taxonomy at 97% and at most two unresolved breaks. The illustrative financial gap represents €52k.
The exploratory model can serve to prepare the data, but allocation is blocked. The decision waits for the €52k to be reconciled, the breaks to be documented and a taxonomy above 97%.
Acceptance conditions
What must be true to act.
- 01
Every control has a formula, a threshold, an owner and an action.
- 02
Blocking gates are not offset by an average.
- 03
Corrections keep the initial value and their justification.
- 04
Authorised uses are explicitly connected to the residual defects.
Limits
What this analysis does not prove.
- Technically clean data can remain causally insufficient.
- Thresholds must vary with the exposure of the decision.
- Automation detects undeclared changes of business meaning poorly.
References
Works cited.
- Wang, R. Y. & Strong, D. M. (1996), “Beyond Accuracy: What Data Quality Means to Data Consumers”, Journal of Management Information Systems, 12(4), 5–33.
- Batini, C. & Scannapieco, M. (2016), Data and Information Quality.
- Chan, D. & Perry, M. (2017), “Challenges and Opportunities in Media Mix Modeling”, Google Research.