Better Attribution Starts with Better Inputs
A sophisticated attribution model can still give you a sophisticated version of the wrong answer.
In marketing, we spend a lot of time analyzing data. We need to spend just as much time understanding how that data was created/collected/sampled/aggregated, and what happened before it reached the model.
What context shaped the results? Which signals were captured, lost, duplicated, or delayed?
Which values were observed, and which were modeled? What assumptions are already embedded in the inputs?
By the time data reaches an attribution model, it may have passed through tracking systems, identity matching, conversion adjustments, and predictive models. Every step can introduce errors that affect the final interpretation.
A tracking issue can look like declining demand.
A change in modeled customer value can look like deteriorating campaign performance.
Duplicate conversions can make an ineffective strategy look successful.
Feed corrupted signals into a model, and you compromise its conclusions, regardless of how advanced it is. The consequences are real: budgets move, effective campaigns get cut, and teams optimize toward a distorted picture of the business.
Before asking, “What does the data tell us?” ask, “Can we trust what went into it”. Understanding the context and validating the signals are part of the analysis. They are what make the analysis useful.
I am amazed on how easy is to blame data models/attribution instead of trying to better understand the data (input) the models get to work!
View the original LinkedIn post ↗