How to build a campaign attribution pipeline

What matters here

• Stitch to the account, not only the person. B2B deals involve several people and person-level attribution misses most of the influence.
• Publish more than one model. First and last touch disagree by design, and the gap between them is information instead of a problem.
• Join real spend. Cost per opportunity typed into a spreadsheet is an estimate somebody will challenge.
• Reconcile to a pipeline number finance recognises. An attribution report that cannot be tied back is a report nobody acts on.
• Keep the raw touches. Every model change is a reprocess, and you will change models.

Who this is for

You run marketing operations or revenue operations. Every quarter attribution is rebuilt in a spreadsheet, and every quarter somebody disputes it with a different model.

How it works in practice

What happens between a click and a defensible pipeline number.

1. 1

   Touches land raw and keep everything

   Channel, campaign, timestamp, person and account where known. Raw survives every model change.

2. 2

   Touches are stitched to a person and an account

   Anonymous to known where possible, and person to account, because B2B buying is a committee.

3. 3

   Spend is joined from the ad platforms

   So cost per outcome is derived from real numbers, not typed into a slide.

4. 4

   Several models run side by side

   First touch, last touch, linear and a weighted one. Each published, none presented as the answer.

5. 5

   The result reconciles to CRM pipeline

   Total attributed pipeline ties back to a number finance already accepts.

6. 6

   The gap between models is reported

   Because where they disagree is where the interesting conversation is.

Paste these into Claude Code or Codex with the Tray Headless plugin installed. Each stage runs on its own. The systems named in them are the worked example rather than a requirement, and every prompt says so.

Once per project, run /tray-workflows:set-workspace to pick the workspace these build in. Point it at a sandbox first.

1. 1

   Set up and land the touches raw

   You will change the model, and raw is what makes that cheap.

   Headless skills build-workflow

   Copy

   Use build-workflow. The systems in play are Marketo, Salesforce and
   Snowflake, or whatever we run in those seats.
   
   Land every touch raw into the warehouse: channel, campaign, medium,
   source, content, timestamp, the person if known, the anonymous id, and
   the landing page.
   
   Do not transform on the way in and do not drop anything that looks
   unused. Every attribution model change is a reprocess of this table, and
   you will change models. Discarding a field now means re-instrumenting
   later.

2. 2

   Stitch to the person, then to the account

   B2B buying is a committee and person-level attribution sees one member.

   Headless skills build-workflow

   Copy

   Use build-workflow. Stitch in two stages.
   
   Anonymous to known: when somebody identifies themselves, backfill their
   prior anonymous touches within a sensible window. That backfill is
   usually where first touch actually lives.
   
   Person to account: through the CRM contact, then by email domain, then
   through the account matching rules we already use.
   
   Then roll touches to the account. A deal with five people touched by
   marketing separately is one buying decision, and attributing to
   individuals splits credit across a committee in a way that means
   nothing.

3. 3

   Join real spend

   A typed cost per opportunity is an estimate somebody will challenge.

   Copy

   Pull spend from the ad platforms and join it to campaigns on a shared
   identifier, not on a campaign name that somebody typed differently in two
   places.
   
   Where a channel has no platform spend, take committed cost from the
   marketing budget instead of leaving it blank. An events channel with no
   cost attached looks impossibly efficient and discredits the whole report.
   
   Then derive cost per lead, per opportunity and per closed deal. Derived
   from two real numbers, never typed, so when somebody challenges it the
   answer is a query, not a recollection.

4. 4

   Publish several models, not one

   Where the models disagree is the actual insight.

   Headless skills tray-gotchas

   Copy

   Use tray-gotchas, then compute several models over the same touch set:
   
     First touch, which flatters awareness channels
     Last touch, which flatters closing channels
     Linear across all touches in the window
     A weighted model with the weights written down and owned by somebody
   
   Publish all of them side by side. Never present one as the answer.
   
   The gap between first and last touch for a channel is the interesting
   number: a channel that looks strong on first and weak on last is doing
   awareness work, and a report that shows only one of those will get it
   defunded or overfunded.
   
   Also handle: touches outside the lookback window, deals with no touches
   at all, which is a stitching failure instead of a sales-sourced deal,
   and self-reported attribution where you collect it.

   Deals with zero touches are the diagnostic worth watching. A rising count almost always means stitching has broken, not that outbound suddenly got better.

5. 5

   Reconcile to a number finance recognises

   Otherwise it is interesting and nobody acts on it.

   Copy

   Reconcile every run: total attributed pipeline against total CRM pipeline
   for the period. They will not match exactly, and the difference has to be
   explainable, not hand-waved.
   
   Report the unattributed share explicitly rather than hiding it. A model
   that quietly drops 30% of pipeline is worse than one that shows the gap.
   
   Then report by channel: spend, pipeline and closed under each model,
   cost per opportunity, and the first-versus-last gap. Plus deals with no
   touches, trended, because that is the health of the stitching rather than
   a finding about outbound.

6. 6

   Validate, then hand the weights over

   Because model weights are a marketing decision, not an engineering one.

   Copy

   Run the per-step schema checks and the whole-workflow audit before this
   touches production.
   
   Then open the same workflow in Tray Build so marketing operations can
   adjust the lookback windows and the weighted model in the visual canvas.
   Those are marketing decisions and they should be arguable without a
   sprint.

What it connects to

Touches, spend and outcomes live in three places and have to meet in one.

Named systems are the ones most teams run, not the only ones that work. Each is an authentication in your Tray workspace, referenced by name, so the workflow never holds a credential. Where we have a connector page, the name links to it.

Questions people ask

Related guides