What AI Should Actually Kill: Marketing Attribution
My most recent corporate position was VP of Marketing. This is where I had to face my arch nemesis: attribution. Why do I have such a dislike for attribution? Because it undermines marketing to the point of making it ineffectual.
I don't think attribution ever really worked, and won't, because it's answering a question that doesn't have the kind of answer it's asking for.
My theory about where it actually came from is a read on the mechanism, not documented history. But it is nearly the same experience I have come across in all my work. Finance started asking marketing to justify itself the way it justifies every other line item: you bought this tool, what did it do for you. That's not an unreasonable question on its own. It's the question you'd ask about a piece of equipment, a hire, a lease. Money went out, show me what came back. Marketing, without a better answer, built one. Attribution is the answer marketing gave to a question finance was always going to keep asking, because no one was willing to say the harder thing: that question assumes something about how a business works that isn't true.
The pressure part checks out even if the origin story is mine. The CMO Survey has tracked marketing leaders reporting increased CFO scrutiny on ROI ticking up from just over half to nearly two-thirds of respondents in a single year, and outlets covering the martech industry have started saying what we all knew inside: that marketing used to be treated as overhead and is now being evaluated the way any capital investment is, against a return threshold it must clear. So whether or not there's a clean origin story, the logic behind it is there and getting worse.
Why do I care about this? Mostly because it involves the work of humans and that work should be respected, not scrutinized for every penny recouped - what a Scrooge McDuck way to act. A SaaS company isn't a set of discrete inputs you can isolate and score. It's an organism. Pull on one thread and the effect surfaces somewhere else entirely. A piece of content someone read eight months ago changes how a rep frames a call, but there isn't a direct link back. A support interaction rebuilds enough trust that it shows up as an expansion deal eleven months later, with no record tying the two together anywhere. Attribution asks which single part caused the outcome. Most of the time, in a system built this way, no single part did. The system did.
This has been going on long before AI entered the conversation, and perhaps why marketers are SO burnt out. Sales teams that don't use the CRM the way marketing needs them to, so the trail goes cold the moment a lead is handed off. Niche markets where the people doing the evaluation never touch your website, because they're in a Slack channel or an industry group or a phone call you have no way to track. In both cases the attribution report is built. It just isn't describing anything real. It's a number confident enough to survive a board meeting, which is a different thing entirely from a number that's true.
This is where AI could genuinely help, and where I think most companies are about to waste the opportunity in front of them. The obvious move, and I'd bet it's already happening in more marketing Slack groups than not, is to point AI at the same fragmented, undefined data and ask it to build a faster attribution model. More touchpoints consumed, a slicker dashboard, a model that updates in real time all the time. That's the same theater with better production values and called progress.
The actual fix removes the need for attribution rather than improving it. If a company has data definitions (BIG assumption), meaning everyone agrees what qualified pipeline or an active account actually means, and traceability, meaning a data point can be followed from the system in which it originated through every transformation to the decision it eventually informs, nobody needs to model influence anymore. You can see the path. AI no longer guesses which touchpoint deserves credit and starts showing what actually happened, in sequence, with the connections intact. And then, we can stop stuffing content with keywords and trying to reverse engineer which prompts get that content ranked and read.
AI's real effect on attribution has nothing to do with better prediction. It just removes the reason the guessing existed in the first place, and only for the companies willing to do the unglamorous work first: define the data, trace the path, treat the business like the interconnected system it is instead of a collection of parts that can be measured in isolation. Everyone else will build a faster version of the same theater and wonder, in another two years, why the number in the board deck still doesn't match what the sales team already knows to be true.
If you want to see what that unglamorous work actually looks like, I've put together the two documents I use to start it — AI Intelligence Layer Definitions template — over on Launch Actually.