How to choose and configure a heuristic model when you can't run experiments

How to choose and configure a heuristic model when you can't run experiments

The question: you lack the data volume or budget for Markov, Shapley, or lift tests. Among the simple rule-based models — first-touch, last-touch, linear, time-decay, position-based — how do you choose deliberately instead of by default?

The selection playbook:

— Map your sales cycle length. Short impulse cycles tolerate last-touch; long considered cycles demand a model that credits the opener.
— Match the model to your actual question. Asking 'what creates demand'? First-touch or position-based (the 40-20-40 'U-shape' that rewards first and last). Asking 'what closes'? Last-touch. Asking 'what sustains a long nurture'? Time-decay, which weights recent touches more.
— Set time-decay's half-life from your real conversion lag, not the default. A 7-day half-life on a 60-day cycle erases the top of the funnel.
— Run two heuristics in parallel and study where they disagree — the disagreement maps your funnel's shape.

The nuance: every heuristic is an assumption wearing a number. None is more 'correct' than another; each encodes a different belief about how touches matter. They describe correlation patterns you imposed by choosing the rule — they prove no causation.

Bottom line for practitioners: pick the heuristic that matches your sales-cycle length and the specific question you're answering, configure its parameters from measured lag, and never present its output as truth. A consciously chosen rule beats an inherited default — but it's still a convention, not a measurement.