Attribution Modelling
Working out what each touchpoint contributed, choosing a model that fits your sales cycle, and being clear about what attribution can and cannot tell you.
Overview
Attribution allocates credit among the touchpoints you can see. It is useful for steering and it is not a measure of cause. Last click undervalues everything that happens early; first click undervalues everything that closes; and no model can account for the touchpoints that were never tracked at all.
The practical approach is a model consistent enough to steer with day to day, combined with periodic experiments that answer the causal question properly. Arguing about which attribution model is correct is usually a substitute for running a holdout test.
Who this is for
- Companies whose channels each claim the same conversions
- Businesses with a long sales cycle and no view of early touchpoints
- Teams making budget decisions on last-click data
How we approach attribution modelling
The specific pieces of work a typical engagement covers. Scope is agreed up front — nothing here is a surprise line item later.
Model selection
Matched to your sales cycle length and channel mix, with the trade-offs of each model explained rather than asserted.
Cross-device and cross-session joining
Stitching journeys where you have a login or a CRM identifier, which is the only reliable way to see a long consideration period.
Offline conversion import
Bringing closed-won data back from the CRM, so attribution reaches revenue rather than stopping at the form.
Incrementality testing
Geographic or audience holdouts that measure what a channel added. This is the part that answers the actual question.
Model comparison
Seeing your data through several models at once, which reveals how much a budget decision depends on the model rather than the results.
Reporting and governance
One agreed model for regular reporting, so decisions are not re-litigated each month on a different basis.
From first call to measured result
The same sequence every time, so you always know what happens next.
Map the journey
Establish what touchpoints exist and which are actually observable. The gap between those matters more than the model.
Model
Implement and compare models against your own data, then agree one for regular use.
Test
Run holdout experiments on the channels where attribution is least trustworthy, usually the upper-funnel ones.
Apply
Feed the findings into budget allocation, and repeat the experiments periodically as the mix changes.
Outcomes, not deliverables
A pile of artefacts isn't progress. These are the changes the work is meant to produce — and what we report against.
Credit where it belongs
Multi-touch models stop last-click quietly defunding the channels that start the journey.
Answers attribution cannot give
Holdout tests measure causation, which no attribution model can, however sophisticated.
Reporting that stops being argued
One agreed model means monthly reviews are about decisions rather than methodology.
Visibility to revenue
Offline conversion import connects marketing activity to closed business rather than to enquiries.
Common questions about Attribution Modelling
The things people ask before they get in touch. If yours is not here, ask us directly.
Which attribution model should we use?
For short cycles, a data-driven or position-based model is usually adequate. For long B2B cycles, attribution alone is insufficient and needs offline conversion data plus experiments. The important thing is choosing one and staying with it, because switching models mid-year makes comparison impossible.
Is multi-touch attribution still possible with privacy changes?
Less completely than before. Cross-site tracking has largely gone, so multi-touch now works best where you have your own identifier such as a login or a CRM record. For everything else, incrementality testing has replaced what attribution used to attempt.
Do we need media mix modelling?
It becomes worthwhile at larger spend levels and with several years of data. Below that, the model is fitting noise. A well-run programme of holdout tests answers most of the same questions for far less.
How often should we run holdout tests?
One channel at a time on a rota, so each significant channel is examined once or twice a year. Continuous holdouts on everything cost too much reach to be practical.
Thinking about Attribution Modelling?
Tell us what you are trying to change. If we are not the right fit we will say so, and point you somewhere better.
Looking at the wider picture?
Attribution Modelling usually sits alongside other work in Data, Analytics & Measurement. Browse the full area to see what it connects to.