Attribution alignment between marketing and the CRM pipeline

The disagreement is rarely about the model. Marketing counts a lead at form submission, sales counts it when a meeting is held, and those two numbers sit 30 to 45 percent apart before anyone has said the words first touch or last touch. This engagement settles the definitions first, in writing, with both function heads present, and builds the reporting afterwards. Where the definitions cannot be settled, no model repairs it, and you hear that in week one instead of paying for a quarter to find out. Where the numbers actually break: Opportunity contact roles sit empty, so nobody knows which humans were involved in the deal. Lead-to-contact conversion drops the original source. A GCLID captured on the landing page never reaches the CRM record, so paid search appears to have produced nothing and gets cut. Campaign influence in Salesforce runs on a default attribution window that nobody chose deliberately. In banking and insurance the deal closes through a broker or a branch, where the final touch belongs to someone outside your tracking entirely. These are plumbing failures wearing the costume of a measurement debate, and they are fixed with engineering rather than with a model. The threshold I hold to: Below roughly 200 closed-won opportunities per quarter, an algorithmic multi-touch model, Markov or Shapley or whatever a vendor has named it, is fitting noise and will move credit between channels on sample size alone. At that volume you get a rules-based model you can defend to a CFO, a self-reported source field on the form, and holdout tests where the spend justifies one. Above it, algorithmic weighting earns its place. You learn which side of the line you are on before you buy, from your own numbers. What gets built: Source and touch capture that survives the lead-to-contact conversion. Contact roles enforced at the stage where they are genuinely knowable rather than requested at the point of close. A touch table in the warehouse joining marketing events to CRM objects on a stable key. Offline conversion upload back to Google and Meta so bidding optimises toward closed revenue instead of form fills, which is regularly the single change with the largest effect on media efficiency. Self-reported attribution kept as a deliberately independent second signal, because where it disagrees with the model the disagreement is itself information. The part nobody wants to schedule: A written definition of a qualified lead, signed by both functions, with the stage transition that proves it and the response time each side commits to. Half of this work is facilitation and I run those sessions personally rather than handing over a template. I will not: Build a model whose conclusion is specified in advance because a channel's budget is under review. Rebuild your lead scoring. Take a side between the CMO and the CRO, since the report says what the data supports, including when it supports neither position. Nor will I produce one number allocating credit across brand, sponsorship and paid social with false confidence, because those sit on different evidence and the report says so plainly. Not for you if: Your CRM stages are aspirational rather than descriptive and deals move between them by end-of-quarter convention. Attribution built on that measures the convention. Fix stage discipline with sales leadership first and hire me afterwards. What you leave with: The definitions document, the touch and conversion models in dbt with tests, offline conversion pipelines running, one attribution view instead of four competing ones, and a written statement of what this measurement cannot tell you, which is the page clients quote back to me a year later.