Attribution is the most politically charged topic in analytics. Everyone agrees it matters. Nobody agrees on how to do it. And the result, in most organisations I've worked with, is that attribution models are chosen not for their accuracy but for their narrative convenience.
The number that gets chosen is the one that makes the most people in the room feel good. And that number is almost always lying to you.
How attribution gets corrupted
It happens in stages. First, someone builds a dashboard. The dashboard shows channel performance. Everyone looks at it in the weekly review. And very quickly, the channels that "look good" on the dashboard get more budget, and the channels that look bad get cut — regardless of whether the model is actually measuring what it claims to measure.
Last-touch attribution is the most common offender. It gives 100% of the credit to the last interaction before conversion. Which means email newsletters, retargeting ads, and branded search — things that operate late in the funnel — look like growth drivers. And the content, the PR, the social posts that actually moved someone from unaware to interested get zero credit.
Last-touch attribution doesn't measure what drove the sale. It measures what happened to be standing next to the customer when they finally decided to buy. Those are not the same thing.
The red flags I watch for
After 20+ years running analytics at enterprise scale, here are the signals that tell me a number is performing for the presentation, not for the business:
It always goes up. Real performance metrics have variance. If a metric only ever improves, someone is either cherry-picking the timeframe, adjusting the definition, or measuring an input they control rather than an outcome they don't.
It's disconnected from revenue. Engagement metrics, session metrics, page view metrics — all of them are fine as diagnostic tools. None of them are business metrics. If a team is reporting engagement as if it were performance, ask what the conversion rate from that engagement is.
Nobody can explain the spike. Ask about any significant movement in any key metric and watch what happens. If the team can explain it precisely — "it was the campaign we ran on the 14th, here's the attribution window" — you have a healthy analytics culture. If they say "it was a good week," you don't.
The model hasn't changed in two years. Attribution models should evolve as your customer journey evolves. A model built when you had three channels shouldn't be running unchanged when you have nine channels and a completely different customer mix.
What better looks like
I've worked with teams that do this well, and the difference is usually not technical sophistication — it's culture. They treat attribution as a hypothesis, not a truth. They use multiple models simultaneously and look for the story that's consistent across all of them. They invest in incrementality testing — controlled experiments that tell you what would have happened without the intervention.
And critically, they have someone in the room who is willing to say: "I think this number is flattering us, and we should look at it differently."
That person is rarely popular in the short term. They're almost always right.
The one change that would help most teams immediately
Stop presenting attributed revenue as if it were actual revenue. It isn't. Attributed revenue is a model's estimate of revenue causation. Present it as such. "Our model attributes $X to this channel" is a different sentence from "this channel drove $X." One is a hypothesis. One is a fact. Only one of them is true.
When you start making that distinction consistently, the conversations get harder. And the decisions get better.