Beyond last click: why your attribution model is quietly lying

There's a moment in almost every budget meeting where the whole problem shows up. Marketing presents the dashboard: paid social drove 4,000 conversions last quarter, search drove 3,000, the platforms are very pleased with themselves. Then someone from finance asks the only question that matters:

"Yes – but how many of those customers would have bought anyway?"

The room usually goes quiet. Not because the question is unfair, but because the tool being used to answer it – the attribution model – cannot, even in principle, answer it. And most marketing budgets are being allocated on exactly that broken foundation.

Last-click attribution is simple to describe: whatever touched the customer last gets 100% of the credit for the sale. Clicked a brand-search ad on the way to checkout? Search gets the sale. Clicked a retargeting banner? Retargeting gets the sale.

It's like crediting the cashier with the entire meal because they're the one who handed you the plate. Brand search and retargeting are the cashiers of digital marketing – they stand at the end of journeys that other channels created, and they collect credit for demand they didn't generate.

So when a last-click report tells you a channel "drove" a conversion, what it actually measured is where demand was collected – not where it was created. Those are different things, and the difference is where budgets go to die.

The usual rebuttal is: "We moved beyond last click years ago. We use multi-touch attribution." Linear, time-decay, position-based, U-shaped, data-driven – pick your flavour.

It feels rigorous. It isn't.

Every multi-touch model is still an arbitrary carve-up of the same correlational data. Choosing 40-20-40 over last-click isn't finding the truth; it's choosing a different story. The weights are conventions, not evidence. "Data-driven" versions just learn the conventions from your historical data – which was generated by the same biased process.

The deeper problem is structural. Any attribution model can only see touchpoints that happened and correlate them with conversions. It cannot see the counterfactual: what would have happened without the spend. And the people your retargeting reached were already on your website. The people who clicked your brand-search ad were already searching for your name. These audiences convert at high rates no matter what you do – so every path report has selection bias baked into it from the start.

Attribution answers the question: what was present before the conversion? Budget allocation needs the answer to a different question: what did the spend actually cause? Confusing the two is how a channel can look like a growth engine while being, in reality, a very expensive toll booth on a road customers were already driving down.

There is only one way to answer a causal question: compare against a world where the thing didn't happen. That means an experiment.

The cleanest version is the holdout. Take a set of comparable regions or audiences, keep spending in most of them, switch the channel off in the rest, and keep everything else as equal as you can. The difference in outcomes between the two groups is the lift – the part of the result the spend genuinely caused. No credit-assignment rules, no assumptions about journeys. Just treated versus control, like any other science.

For the portfolio view, marketing mix modelling does the complementary job: it uses variation in spend and outcomes over time to estimate each channel's contribution, including channels you can't easily experiment on. The model is slower and coarser than an experiment, which is why the two work best together – experiments calibrate the model, and the model fills the gaps between experiments.

Neither is exotic. Both have been standard practice in serious measurement teams for years. What's rare isn't the method – it's the willingness to find out.

An ASX-listed insurer I worked with had the classic setup. Platform-reported conversions said paid social was the growth engine, and the budget had followed the story. Finance had doubts but no counter-evidence – just a feeling that the numbers were too clean.

We ran geo holdouts. A large share of the "conversions" the platforms were claiming was demand that arrived anyway – people who were already coming, intercepted at the last step and invoiced for the privilege. Mix modelling confirmed the pattern at the portfolio level.

The result: $6.2M of annual spend reallocated away from channels that were harvesting demand and towards channels that were creating it – worth +27% in incremental ROAS. Not because paid social didn't work. It did – at a fraction of the claimed level. The budget was the problem: it had been sized on the claim, not the evidence.

That's the part that stays with you. Nothing about the data changed. What changed was the question.

A few tells that your attribution is flattering your spend:

The first step doesn't require a programme, a platform or a consultancy. Pick one channel, one set of geographies, one month. Run a single clean holdout. It's uncomfortable exactly once – the first time a favourite channel comes back smaller than its dashboard. After that, nobody who has seen the difference wants to go back.

If a number can't survive a holdout, it's not a result. It's just a story – and last click is one of the oldest, most persuasive stories in marketing. Budgets deserve better than stories.