Politikos Shop.

With the account fully automated and scaled, a new question surfaced: where should the next incremental euro go? The default last-click attribution model — the one baked into most ad platforms — credits the final channel before conversion with the entire sale. That single assumption quietly underweights upper-funnel campaigns that assisted the conversion but never closed it, and it over-credits brand and retargeting channels that would have converted anyway.

For an account already running at scale, this is not a rounding error. It translates directly into real budget decisions: campaigns that appear "inefficient" under last-click get defunded, even when they are the discovery engine feeding the entire funnel. Over time, the account converges toward a narrow slice of bottom-funnel activity and stops growing the top of its own pipeline — an invisible ceiling that manual optimization cannot break.

The goal was to replace the platform default with something closer to reality: a probabilistic model that captures the actual causal contribution of each touchpoint, and a decision layer that continuously reallocates budget to where expected profit is highest. Anything less would lock the account into the same last-click blind spots that limited the previous year.

The solution combines two probability-theory models. Markov Chain treats the customer journey as a stochastic sequence of states — impression, click, visit, conversion — and computes the removal effect of each touchpoint. Pull a channel or campaign out of every historical path and measure how much conversion probability drops: that drop is its true causal contribution. The output is a credible attribution weight for every channel and campaign, grounded in observed data rather than platform default heuristics.

On top of attribution sits the allocation problem, which is a classic Multi-Armed Bandit: given a budget to distribute across a set of choices (channels, campaigns) with uncertain payoffs, what allocation maximizes expected profit? The bandit continuously balances exploration — sending a small share of budget to under-tested campaigns to learn their real ROI — with exploitation — pouring the majority into proven winners. Over time it converges on an allocation that beats any static budget split.

Operationally, the two models form a closed loop. Markov Chain runs weekly on the latest path data, producing updated attribution weights. Those weights feed the Multi-Armed Bandit, which adjusts budgets daily across the account structure. Every Monday the account looks slightly different — because every week the data tells a slightly different story about where expected profit actually lives.