Timo and I ran a workshop on Attribution in the Warehouse, together with John Wessel, from RudderStack, who sponsored the session. This is a recap of what we covered. The [full recording is here](https://partnerwithpropel.com/workshops/attribution-in-the-warehouse/)if you'd rather watch it.

#### Attribution usually fails in the meeting, not in the model

Here is a scene most marketing teams will recognise. You are in a planning meeting for the next quarter. Someone pulls up a dashboard. Someone else says they are not sure the conversions can be trusted, or that the attribution model is not fair to a particular channel. An hour later you have talked about tracking gaps and consent and models, and the thing you came to decide, which initiatives to run next, is still undecided.

The most common example is a channel you know works but cannot prove. You have heard it in customer interviews. Paid social, or a podcast, keeps coming up. But in the meeting the numbers do not back you, so you end up fighting for budget on a channel finance has already written off as paid search. A lot of the conversations that start with Propel begin exactly there: marketing knows the podcast is working, finance sees only paid search on the dashboard, and nobody believes marketing.

When teams try to fix this, they usually reach for one of two levers: change the attribution model, or capture more data. Neither is wrong on its own. But we see the same thing happen even after a team gets tracking near-perfect:

> The question does not go away.

That is the tell. If better tracking and a new model do not settle the argument, the problem is not the model. It is where attribution sits.

#### Attribution belongs inside your marketing strategy, not in a box beside it

Attribution fails when it sits as an isolated problem, a magic box you point at marketing after the fact. The better position is inside the strategy itself. You take the strategy, break it into the initiatives you actually want to run, and decide how you will measure each one as you plan it, not afterwards.

Take the podcast example. If you decide to launch a podcast, the moment to design its measurement is when you design the initiative, not a quarter later when it fails to show up on a dashboard. That is also when you can build the things that make it measurable at all. Attribution stops being one central source of truth and becomes something you apply per initiative, with the context each one needs.

#### So the answer to "what's the better model for us?" is that there isn't one. Every model does a specific job. Once you know what job each one does, you can apply it to the right question. What you need instead is a place to bring the data together. We call it a marketing intelligence layer, and it rests on three things.

First building block: what marketing provides

![What marketing provides: clear UTMs, initiatives and mapping, and the attribution rules. The context the data team doesn't have.](/images/blog/attribution-in-the-warehouse-a-recap-of-our-workshop/image-001.png)
_What marketing provides: clear UTMs, initiatives and mapping, and the attribution rules. The context the data team doesn't have._

This is the part most often left out, because these tasks get pushed to the data team. But marketing owns context the data team does not have, and three pieces of it matter.

**Clear capture.** UTMs are still the backbone, and in most channels they can and should be automated rather than maintained in a sheet that people copy and paste from, because that is where errors come from. Barbara has written a full walkthrough of [how to automate your UTM parameters](https://www.021newsletter.com/p/dynamic-url-utm-parameters-what-are). Capture also covers survey design: the "how did you hear about us" survey at signup or purchase, and earlier identification points like webinar or download forms where you can ask a few questions and identify people sooner. Surveys are underrated on the data side too, because they are direct, stated data from the user rather than an observation you have to interpret. The one discipline that makes them work: keep the options current. Launch a podcast strategy and forget to add podcast as an option, and you have designed the answer out of your own survey.

**Meaning.** The same raw journey can be read through more than one lens. Channel is the familiar one: paid social, retargeting, where the budget goes. Initiative is closer to the strategy, usually time-bound, and closer to what marketing actually pours work into. Marketing can map UTMs, landing pages and form submissions to both, which also lets you enrich touchpoints that arrive with no source. Capture is a project you finish. Meaning is a practice you keep up.

**The rules.** Our most contested take is that attribution modelling should not be owned by the data team, because they do not have the context on initiatives, journeys and the blind spots of each channel. Marketing knows, for instance, that a podcast listener will never click a UTM, so the rule for crediting that touch has to be a marketing decision. The one condition is transparency: the rules are written as something everyone can see, follow and challenge.

#### Second building block: what data provides

![What data provides: resolved identity, focused touchpoints and conversions. The half that joins anonymous and known traffic into one journey.](/images/blog/attribution-in-the-warehouse-a-recap-of-our-workshop/image-002.png)
_What data provides: resolved identity, focused touchpoints and conversions. The half that joins anonymous and known traffic into one journey._

The other half is an equal share of work, and it is the data team's.

**Resolved identity** is the core. Most of your traffic is anonymous and some of it is known; identity resolution is what joins the two so you can see a whole journey. Assign anonymous IDs the moment someone hits your site, then stitch that traffic back to the user once they are known, ideally in the warehouse. Start simple and tool-based, which gets you most of the way, and only fight for the last hard portion of matches if it earns its keep.

**Focused touchpoints.** A touchpoint is any moment in the journey you can derive marketing meaning from. The default is a visit. But a dedicated landing page can count even when the visitor arrives with no UTM or referrer, say from a WhatsApp group, because the page itself is assigned to an initiative. And a survey answer can become a virtual touchpoint: take the "how did you hear about us" response, place it in the journey, and treat it on equal terms with the rest.

This is where modelling in the warehouse pulls ahead of an event-analytics tool. You can infer touchpoints from sources those tools ignore. On a current client project, one source is advisors who invite clients by email through a dashboard, so the invite itself becomes a matchable touchpoint. Another is a credit-card partnership where the partner passes the last four digits of the card, enough to tie a purchase back to that partnership. None of that lives in GA4. Modelled in the warehouse, all of it can become a touchpoint you attribute against.

**Conversions** are what you optimise toward. Often one core event, but for a B2B SaaS it might be a chain: lead created, opportunity created, opportunity won. With the touchpoints in between, you can build attribution across the whole path or just one segment of it.

#### Third building block: what the layer unlocks

Put those together, across warehouse systems, CRMs, every touchpoint and conversion, with identity resolution underneath, and you have one customer, one journey, that marketing can read from different angles. Break it down by channel to see where the ad budget is working. Break the same journey down by initiative to see how the strategy is performing.

![One journey, read two ways: by channel to see where the budget works, by initiative to see how the strategy performs.](/images/blog/attribution-in-the-warehouse-a-recap-of-our-workshop/image-003.png)
_One journey, read two ways: by channel to see where the budget works, by initiative to see how the strategy performs._

Because the data sits in one place, you can run different attribution lenses over it and let each answer its own question. First click shows the earliest touch you can capture. Last click is the classic closer. And where you have written rules, you can make deliberate, auditable calls: when these conditions hold, this conversion is paid social, and you can prove why for every single conversion. The layer also feeds activation and the agent-driven work everyone is now curious about. It only works because the rest is in place first.

#### The takeaway: if you can't explain the number, no one will use it

The thread running through all of it: if you cannot explain how a number was built, well enough to trace it back and show your working, nobody will trust it, and you are back in that planning meeting arguing about the model. This is the case against black-box attribution, including GA4's data-driven model. You cannot see why a channel got the credit it got, and there is an awkward conflict of interest in an ad platform grading its own homework.

One client ran UA and GA4 side by side for a month, same data, same data-driven model on paper, and the results diverged wildly. GA4 showed affiliates as near-dead, and the team nearly cut all affiliate spend over what turned out to be a model change nobody could interrogate. A model you can audit beats one you cannot, every time. A multi-touch model you can audit is better still, because it is both explainable and more complete.

If you want to talk through what a marketing intelligence layer would take for your own setup, book a call with us.

Propel builds bespoke marketing measurement setups, including warehouse-based attribution, identity resolution and the marketing intelligence layer that sits under them, for teams who want their measurement to fit their marketing rather than the other way round.