Attribution Models in Fintech: Why Last-Click Robs Your Best Channels

An attribution model isn't a setting in a report. It's a decision about where your money goes. Because the model determines which channel gets "credit" for a conversion, and you then allocate budget on the basis of that credit. Pick a model and you've effectively picked who to praise and who to cut.

And most people sit on last-click by default, without even realizing that this default systematically distorts the picture.

How last-click lies

Last-click hands all the credit to the final touch before the conversion. In fintech, with its long funnel, that's almost always branded search or retargeting — the bottom of the funnel.

Here's what follows: branded search looks like a genius (it's the one "closing" conversions), while the channel that introduced the person to the product three weeks ago looks useless. You cut the top of the funnel as "inefficient," the flow of new people dries up, and a month later that same genius branded search sags too, because the brand queries have run out. Nobody new heard about you.

It's a classic trap. Last-click pushes you to kill the very channels that fill the funnel.

First-click and position-based

First-click is the opposite extreme: all the credit to the first touch. It's useful when you need to evaluate awareness channels and understand what genuinely brings in new people. But it undervalues the bottom of the funnel, the part that closes.

Position-based is the compromise: more weight on the first and last touch, less on the middle. For a long funnel that's often more sensible than either extreme.

Data-driven — what's inside the black box

Google and GA4 push data-driven as the "smart" model: the algorithm works out each touchpoint's contribution from your own data. The idea is good. But it's a black box, and it only works with enough conversion volume — on small datasets the model is unstable and produces noise, then passes that noise off as insight.

A contrarian take: data-driven isn't "the best model" the way it's presented. It's the best model given a large, steady stream of conversions. At low volumes it's more honest to take position-based and understand its logic than to trust a black box learning from ten conversions a week.

What no model does

And here's the main thing every attribution argument forgets. Any model is just a way of slicing the same pie. It shifts credit between channels, but it never tells you what would have happened if the channel hadn't existed at all.

And that is the only genuinely important question: if I switch this channel off, do I lose conversions or do they simply flow into another one? That's answered not by an attribution model but by incrementality tests — geo-lift, holdout groups. You switch the channel off for part of the audience and watch whether conversions drop.

A take people don't like: arguing about attribution models forever is a way of not running incrementality tests. A model will show you a pretty breakdown, but only a holdout will give you the truth about which channel actually creates sales and which one is merely taking credit for them. One honest geo experiment is worth a year of arguing about first-click versus data-driven.