Attribution Model Intake and Baseline Lock (Onboarding)

Attribution Model Intake and Baseline Lock (Onboarding)

1. 01

   Map the client's actual revenue path before selecting any attribution method

   Document every touchpoint from first impression to closed revenue, including offline calls, partner referrals, and CRM stages. A B2B pipeline with a 90-day sales cycle needs a different model than a DTC store with a 3-day repurchase window.

2. 02

   Pull 90 days of raw channel spend and revenue into one table the client can inspect

   Reconcile ad platform totals against bank statements and the payment processor. Discrepancies above 5% usually trace to currency handling, refunds, or duplicate conversion events and must be resolved before modeling.

3. 03

   Classify each channel as measurable, partially measurable, or dark

   Dark channels include word of mouth, AI assistant referrals, and offline events. Naming them at intake prevents the client from later asking why those conversions appear unattributed.

4. 04

   Agree on the primary success metric and the reporting cadence in writing

   Pick one number the retainer is judged against, whether blended CAC, incremental ROAS, or pipeline-sourced revenue. Confirm who on the client side signs off on the monthly read.

5. 05

   Set a baseline period and freeze it

   Choose a 60 to 90 day window with no major campaign launches or pricing changes, export the numbers, and store them. Every future lift claim gets measured against this locked baseline.

6. 06

   Document the methodology in a one-page client-facing note

   State which model type is in use (multi-touch, media mix, or incrementality test), what data feeds it, and what it cannot see. Platforms such as Ruler Analytics and SegmentStream support several model types, so the note must name the specific configuration.

7. 07

   Schedule the first incrementality test before the first quarterly review

   A geo holdout or spend-holdout test on the largest channel gives the agency a causal number to defend the model. Without one, the client is trusting a black box.