Understanding customer journeys
At the core of any attribution model is the customer journey, which is the series of interactions a user has with marketing touchpoints before converting. These touchpoints can include paid ads, organic search visits, email clicks, or social media engagements. The order and timing of these events matter: a touchpoint early in the journey may play a different role than one right before the conversion.
A well-designed attribution model should account for:
The sequence of touchpoints not just which ones occurred
The time lag between interactions and the eventual conversion
The type of engagement like clicks, impressions, page views, and so on
Multi-touch attribution (MTA) models, especially those powered by machine learning, can learn how the combination and arrangement of events influence outcomes.
Incrementality and correlation
A common trap in attribution modeling is assuming that just because a touchpoint appears frequently before conversions, it caused them. This is the difference between correlation and incrementality.
Correlation reflects a statistical association between a touchpoint and a conversion.
Incrementality measures the causal effect of a touchpoint, or in other words, whether removing it would reduce conversions.
MTA, especially when paired with methods like incrementality testing or triangulated modeling, helps you move beyond correlation and closer to understanding causal impact.
Biases and blind spots in attribution
Even advanced models face several challenges when trying to fairly assign credit:
Survivorship bias: Attribution models that only include converting journeys miss critical context. For example, if a touchpoint appears in converting paths but also frequently in non-converting ones, its real impact may be overstated. Including non-converting journeys is essential for accuracy and Funnel's LSTM-based model takes this into account.
Lower-funnel over-crediting: Single-touch models, like last-click, tend to favor bottom-of-the-funnel channels that appear close to the point of conversion, for example branded search. These models ignore the earlier interactions that nurtured the user.
Tracking limitations: Modern privacy practices, like cookie consent or ad blockers, and multi-device behavior, like mobile-to-desktop transitions, introduce blind spots. If a user interacts on one device and converts on another, those interactions may go untracked. Funnel addresses this through:
Preference for deterministic identifiers, like hashed email
Handling of opt-out traffic via extrapolation
Modeling based on observed patterns across similar users