Attribution on a 180-Day Cycle With 38% Identity Stitching
A B2B SaaS company sells a $40K annual contract through a 90–180 day cycle. Six channels are tracked: paid search, LinkedIn, organic, webinar, partner referral, and direct. Identity resolution stitches ECID to CRM ID for only 38% of persons; the rest stay anonymous. Direct traffic accounts for 41% of opportunity-creation touchpoints. Webinars live in a separate dataset that arrives two days late and carries a different person identifier. Marketing leadership wants to reallocate 30% of paid budget and has asked for a channel-level ROAS they can defend to the board. Last-click attribution currently credits paid search with 52% of pipeline.
1. Choose an attribution model and lookback window for this business, and justify the choice against the sales cycle and the stitching gap.
Read the worked answer
No single model survives a 180-day cycle with 38% stitching. Choose the model whose weaknesses are visible.
- Model: position-based (40% first touch, 40% last touch, 20% spread across the middle) at person level, reported alongside last-click rather than replacing it. Last-click handing 52% of pipeline to paid search is a warning sign, not a result - search captures demand created elsewhere and sits late in the funnel.
- Lookback: 90 days minimum, 180 if the opportunity data supports it. Anything shorter systematically deletes upper-funnel touches - webinars, LinkedIn - and pushes credit onto search and direct.
- Do not lead with a data-driven model until stitching is materially above 38%; it will learn anonymous behaviour rather than person behaviour, and it will look authoritative while doing so.
- Publish both models with the delta. The gap between them is the actual insight, and it is what protects the analyst when the number is challenged.
2. How do you treat direct traffic, the late-arriving webinar dataset, and the unstitched 62% without quietly inventing credit?
Read the worked answer
Direct (41%) is the consequential one. In B2B, direct at opportunity creation usually means untracked: dark social, offline conversations, redirects stripping parameters, or a CRM-entered opportunity with no web touch. Break direct down by landing page and referrer to separate true direct from attribution loss, and report the split rather than labelling it all 'brand'.
Webinars. Join on CRM ID, not ECID, and timestamp webinar touches at the event date rather than the ingest date. Otherwise the two-day lag pushes touches outside the lookback window and silently deletes the channel from upper-funnel credit. Prefer a lookup-style enrichment onto the person over a late event stream.
The unstitched 62%. Report an explicit unattributed bucket, and show channel shares twice: stitched persons only, and all persons. The difference is a measurement-quality metric and belongs on the dashboard next to ROAS.
Never redistribute anonymous credit proportionally without labelling it an assumption - that is how a defensible model becomes an indefensible one.
3. Leadership wants to move 30% of paid budget. What do you actually recommend, and how do you frame the limits of attribution?
Read the worked answer
Give the board two numbers and one caveat, not one number.
- Primary: pipeline influence - the share of pipeline with at least one touch from each channel, plus cost per influenced opportunity. This is far more robust to model choice than attributed ROAS and survives cross-examination.
- Secondary: position-based attributed ROAS, with the 38% stitching limitation stated in the footnote rather than discovered later.
- Caveat: attribution tells you where credit landed, not what would have happened without the spend. It cannot underwrite a 30% reallocation on its own.
Recommendation: move in tranches. Shift 10%, hold for one full cycle length (180 days), and measure incremental pipeline via a geo holdout or an audience-level holdout. If a holdout is commercially impossible, propose a media-mix model or a matched-market test as the tiebreaker.
Framing matters: the analyst who says 'here is what attribution can and cannot support' keeps credibility. The one who produces a single ROAS number and watches it fail will not be asked again.