← Insights

Attribution you can defend

Every attribution model is wrong. The useful question is which decisions you are willing to make with it, and which you are not.

Sarthak Jain, Co-founder · · 5 min read

Attribution debates go in circles because both sides are right. Last-touch overweights the final click. First-touch overweights the discovery channel. Multi-touch distributes credit using weights someone invented.

None of them are true. They are lenses, and the mistake is treating whichever one is installed as a measurement rather than a point of view.

Decide what the number is for

Attribution supports two very different decisions, and they need different rigour:

Where to spend more, at the margin. This tolerates a rough model. You are comparing channels against each other under the same lens, so a consistent bias mostly cancels out.

Whether a channel is worth existing at all. This does not tolerate it. A channel that assists constantly and closes rarely looks worthless under last-touch and indispensable under first-touch. Killing it based on either is a coin flip.

Use attribution for the first. Use a holdout or a genuine pause for the second — turning something off for a period and watching what happens is crude, unpopular, and far more honest than any model.

Record what you could not see

In B2B the decisive touchpoints are frequently invisible: a conversation in a Slack group, a recommendation from a former colleague, a model naming you in an answer with no referrer at all.

Which means self-reported attribution — a single open "how did you hear about us?" field on the enquiry form — is often the highest-signal data in the whole stack. It is unstructured and biased and it still catches things the pixel never will.

Collect both. Where they disagree, that gap is information.

Write down the model's blind spots

Whatever model you land on, document what it systematically undercounts, next to the dashboard. Dark social. Word of mouth. AI answers. Long consideration cycles that exceed the cookie window.

This one paragraph does more for decision quality than another integration, because it stops people treating the chart as complete.

Consistency beats sophistication

An imperfect model applied consistently for a year tells you more than three sophisticated models applied for a quarter each. Pick one, write down what it cannot see, and keep it long enough to have a trend.

If this describes your situation, start with a baseline.

We document where you stand before recommending anything - so whatever we claim at 90 days has a before to be measured against.