METHOD
Simple rules bound the story. Markov chains and Shapley values argue from paths. An evaluator ranks them. A marketing professional with extensive experience in marketing analytics still has to press send.
Each model below: what it does, when to use it, then how it is computed. First-click is not a toy bound — it is the right model when you are funding discovery.
Simple rule
All credit to the last touch. Easy to explain. Still the number most meetings start with.
Use it as the shared baseline — and when the path is genuinely short: branded search, remarketing, checkout.
Gives every conversion to the last session before it happened. Last-click stays on the page so last-touch ROAS is still in the meeting. It tends to credit closers (Direct, branded Paid Search) and miss assists. A reporting habit — not proof that a channel caused the sale.
Simple rule
All credit to the first touch. This is how a converting journey actually started.
Use it when discovery matters: new markets, prospecting, upper-funnel brand, first-time buyers. First-click is often the right bound when you are deciding whether to keep a channel that almost never closes.
Gives every conversion to the first session in the lookback. It shows openers last-click hides — often Organic Search, Paid Social, and Display. The number moves if the lookback or Direct mapping changes. Treated as the ceiling on introducer value, not as the only plan.
Simple rule
Equal credit across every step in the converting path.
Use it when the path is a team sport — mid-funnel nurture, content plus paid, long B2B cycles — and you want a conservative midpoint between first and last.
Splits credit evenly across every step. Softens last-click’s closer-bias without picking a fade speed. Long messy paths (bots, consent duplicates) can stretch it. Used as a midpoint check against the path models, not as the recommendation.
Simple rule
Later touches get more credit. How fast it fades is a setting agreed with you.
Use it when recency is a real signal: short purchase cycles, promo windows, retargeting that is supposed to close.
Gives more credit to touches closer to the conversion. Older touches fade on a half-life you set (default: a week). That setting is written in the appendix so the number can be rebuilt. It still does not say what would happen if a channel were removed.
Path model
Asks what would happen if a channel disappeared from the path.
Use it when you need to know which channels are load-bearing — the ones converting journeys actually pass through, not just the ones that happen to be last.
Maps how people move between channels, then asks: if this channel were gone, how many conversions would disappear? That is the removal effect. It needs full journeys from BigQuery — the Data API cannot do it. The report shows the removal next to the path steps that drive it.
Path model
Fair credit from every combination of channels, with a range so you can see the uncertainty.
Use it when channels work as a coalition — paid plus email plus organic — and you need a number you can defend as equitable rather than sequential.
Looks at every mix of channels on converting paths and splits credit so no channel is favoured just for being first or last. Few channels: exact split. Many: an estimate with a range, printed on the page. If the range overlaps last-click, no move is recommended.
Numbers first, language model second.
The narrative may only mention figures that already sit in the payload. The rubric can promote Markov chains or Shapley values over last-click; it cannot rewrite their shares.
The numbers have to add up before the story is trusted.
Shares sum to 1.0. Removal effects are non-negative. No channel is credited more conversions than exist. A model that fails coherence cannot lead the PDF.
A jump this week is not automatically a budget move.
Spearman rank correlation versus the trailing eight weeks, plus a per-channel share-swing threshold. A swing coincident with a consent or export gap is labeled data-quality, not a signal.
If it would not change a meeting, it does not lead the report.
The ranking must imply a hold, move, or test a Head of Growth could take this week. Path models that merely restyle last-click stay in the grid; they do not get the decision sheet.
Every model is wrong in a way that is written down.
Consent-mode undercount, Direct inflation, view-through gaps, lookback truncation. Written into the appendix every week. The rubric can demote a model whose known bias dominates the disagreement.
