Multi-Touch Attribution Reconciled to Finance

Below roughly three hundred conversions a month on a channel pair, an algorithmic attribution model is fitting noise and will tell you a different story every week. That threshold decides the engagement. Above it, I build a path-level model in your warehouse and test it against geo holdouts. Below it, I say so, install position-based rules everyone can read, and put the budget into incrementality tests instead. The position I take: Attribution that cannot be reconciled to recognized revenue is an opinion with a chart attached. Every model I ship lands next to a table tying modeled revenue to what finance booked for the same period, and if the two sit more than about two percent apart the work is not finished. Platform-reported conversions are treated as one more claimant, never as the answer. Meta and Google both count the same order, and a stack that lets them each keep it is not an attribution system. Collection: Click identifiers are the foundation and are almost always broken. gclid, gbraid, wbraid, fbclid, ttclid and msclkid captured on landing, written to a first-party cookie with a lifetime chosen deliberately rather than inherited, and copied onto the CRM record the moment a lead or account is created. Server-side conversion delivery through the Google Ads API, the Meta Conversions API and the TikTok Events API, routed through one sending service rather than three integrations that drift apart over a year. Modeling: Path tables in BigQuery or Snowflake, sessionized by rules a channel lead can read out loud. Removal-effect Markov as the default, because its output survives being questioned in a room; Shapley where the path count is small enough to compute exactly. Lookback and decay windows come from your own repeat-purchase curve, not a platform default. Every credit split ships with its stability across a rolling window, because a number that moves eight points week to week is a number people will stop believing. Validation: A model that has never been contradicted has never been tested. Where volume supports it I run geo holdouts and compare measured lift against what the model claimed for the same channel; where it does not, I schedule a spend pause and read the gap. Disagreements get documented rather than smoothed away. Knowing that the model overstates branded search by a known margin is worth more than a model that claims it does not. Regulated categories: Hashed email is the only user identifier that leaves your systems, and in health and wellness even that is gated on consent state before it reaches an ad platform. For igaming operators, the conversion payload carries an order value and nothing about what was wagered, and jurisdiction enters the model as a dimension rather than a filter applied after the fact. Not included: No mobile app install attribution and no MMP replacement. If your growth is install-led, an SKAdNetwork specialist is the correct hire and I will say so on the first call. No view-through credit, in any tier, for any channel. No dashboard build either: output lands as warehouse tables on a schedule, and your BI tool reads them. Who this is not for: Businesses without a stable unique customer identifier in the CRM. That gets fixed first, it is not a small job, and attribution built on top of a broken key produces confident nonsense. Teams who need the model to agree with the platform numbers for a board deck. And anyone expecting an answer in two weeks — the first credible read comes after a full purchase cycle plus one holdout.