# Before AI optimises your campaigns, fix what you're measuring

> Canonical: https://deliverbetter.co.uk/insights/ai-marketing-attribution/
> Published: 28 Sep 2026. Last updated: 30 Sep 2026.
> Author: David Bain, Deliver better.

## In short

Sort out what you're measuring before you let AI optimise it. Join the CRM, web analytics and finance data so one journey can be followed from first touch to sale, agree what counts as a lead and a sale, and treat every attribution figure as an estimate with its limits written down.

Every AI pitch in marketing at the moment promises to optimise something. Bids, budgets, send times, audiences. And it can. The models are genuinely good at finding the pattern in a mess of channel data and moving money towards what’s working.

The catch is what “working” means. An optimisation engine believes your data. If your CRM says one thing, your web analytics says another, and neither matches the sales figures, AI won’t resolve that for you. It will pick one version, act on it with total confidence, and spend your budget accordingly. The old problem was a slow, biased guess about what drives sales. The new problem is a fast, confident one.

So the useful order is the other way round from how it’s usually sold. Sort out what you’re measuring first. Then let AI optimise it.

 

 

 

 
 
 
 
222 What is AI-assisted marketing attribution?
 

 
 Attribution is working out which marketing activity actually brought in a customer, so you can decide where the next pound should go. AI-assisted attribution uses a model to share the credit for a sale across all the touch-points that contributed, rather than crediting only the first click or the last one. Done well, it gives you a more realistic picture of what’s driving revenue. Done badly, it gives you a more sophisticated way to be wrong.

The honest bit first. No attribution model is perfect, AI or otherwise. Some of the journey is invisible, some of it is offline, and some of it is people who would have bought anyway. The aim isn’t a model that’s right to two decimal places. It’s a model that’s good enough to make better budget decisions than habit, with its limits written down. Writing those limits down is what makes finance trust it, not the other way round.

 

 

 

 
 
 
 
222 The data has to agree before the model can help
 

 
 Most attribution problems aren’t modelling problems. They’re plumbing problems.

In a lot of marketing teams the customer data lives in three places that don’t quite match. The CRM holds part of the story. Web analytics holds another part. The sales or finance system holds what actually got paid. Leads get tracked from first contact to sale by hand, if at all. Ask which channel drove last quarter’s revenue and you get three answers and an argument.

AI can’t fix that by being clever. It needs the systems joined up so a single customer journey can be followed from first touch to sale. That usually means getting the CRM to reflect how you really sell, connecting the journeys to it, and agreeing a shared definition of a lead and a sale that marketing, sales and finance all sign. Dull work, mostly. It’s also the work that makes everything after it possible. When we connected HubSpot, customer journeys and an AI-assisted attribution model for one team, the model was the easy part. Getting the data to agree, and getting the people to agree on what the data meant, was the job.

 

 

 

 
 
 
 
222 Where AI earns its place in campaigns
 

 
 Once the measurement holds up, AI has real value across a campaign, not just at the end of it.

It optimises media in real time, adjusting targeting, placement and bids across channels faster than any team could by hand, to stretch the budget further. It spots patterns in customer data that point to who’s worth reaching and who’s about to leave. And it tests creative at a scale that changes the odds. JPMorgan Chase signed a five-year deal with Persado precisely because a model that generates and tests thousands of message variants can find copy that outperforms the human-written version, sometimes by a wide margin. Their own chief marketing officer framed the result as more humanity in the marketing, not less, because the team spent its time on the ideas rather than the fortieth subject-line tweak.

That’s the pattern worth copying. AI does the variation and the measurement. People do the judgement about what’s worth testing and what the numbers actually mean.

 

 

 

 
 
 
 
222 The number that looks like a fact
 

 
 One warning, because it catches people out. When a model gives you an attribution figure, it looks like a measurement. It isn’t. It’s an estimate with a confidence level, dressed up as a decimal.

Treat it like one. Keep a human read on the numbers, especially when the model recommends something that would have been obviously daft a year ago. Watch for the channel that gets all the credit simply because it’s the easiest to track. And be as interested in what the model can’t see as in what it can. A tool that never surprises you and never admits a gap isn’t giving you the full picture. It’s giving you the convenient one.

 

 

 

 
 
 
 
222 What to do this week
 

 
 Take one question you can’t currently answer with confidence. Something like which channel brought in your most valuable customers last quarter, or what a lead from each source is actually worth. Then trace back what you’d need to answer it: which systems hold the pieces, whether they agree, and where the journey breaks. That gap, not the model, is your first job.

Get that right and AI-assisted attribution stops being a black box and starts being a better argument in the budget meeting. Which is the whole point.

 

 

 

 
 
 
 
222 Can you report on clicks but not on revenue?
 

 
 If you can report on clicks and opens but not on revenue, tell me what you’d like to be able to answer. We’ll work out how.

 

 
 
 [
 
 Start a conversation
 
 ](/contact)
 

 
 [
 
 Try the free tools
 
 ](/free-tools)
