The problem with last-click isn't complexity
Last-click is simple. Usually wrong. It hands all the credit to the final touchpoint and calls it a day, which means paid search harvests credit for conversions that social or display or email actually generated. Every channel wants to be last. That pressure shapes optimization in ways that hurt the portfolio even as it flatters the individual channel report.
The deeper problem: platforms report attribution in ways that favor their own channels. Google's default attribution model lives inside Google. Meta's lives inside Meta. Neither has any incentive to give credit to the other. Running both without calibration gives you two competing scorecards and no ground truth.
Where MTA is useful and where it isn't
Multi-touch attribution tells you how credit distributes across touchpoints inside a winning path. That's a real question. But MTA on its own cannot tell you whether a channel caused anything. A touchpoint appearing in every converting path might be essential. It might be ambient noise that shows up because it has the widest reach. The model can't distinguish between the two without external calibration.
The models themselves matter less than that constraint. Linear, time-decay, position-based, and data-driven all answer the same question (how to divide credit) and none of them can prove causation. Picking the "right" model without incrementality to anchor it is an argument about which flavor of correlation to use.
What a calibrated MTA actually looks like
The approach we use at PCG has two layers.
The first is platform-native: Google Ads conversion data, GA4 path analysis, standard model comparisons across channels. This is the baseline. It tells you what the platforms see and where their defaults are likely to be wrong.
The second is a Bayesian attribution model, built when the program is large enough to justify it, and calibrated against at least one clean incrementality test per major channel. The incrementality test establishes ground truth for causal lift. The Bayesian model uses that anchor to assign credit that reflects what actually happened, not what the platform wants to claim.
Calibration is the word that matters. An MTA model not tied to any causal test is a sophisticated-looking guess. Calibrated against a geo holdout or a ghost-ad study, it becomes a tactical lens you can act on.
Where the real leverage is
In the accounts we audit, the most common problem isn't a bad attribution model. It's broken conversion architecture upstream. Server-side events misfiring, duplicate conversion actions, enhanced conversions not passing, value data dropped at the tag. If the inputs are wrong, the model produces confident nonsense regardless of how it's built.
Clean conversion architecture first. Platform-native MTA second. Bayesian calibration third, when the data and the spend support it.
The goal isn't a perfect model. It's a model that's honest about what it knows, anchored to real-world causal data, and good enough to make better decisions than the platform default. That bar is achievable. Most accounts aren't meeting it.