Your Attribution Model Is a Budget Decision You Never Made

Early in my career in marketing analytics, my team ran a position-based attribution model. We chose it deliberately. It gave credit to the first touch, the last touch, and the steps in between.

Then we hit a wall. We could not track users across platforms. Mobile app attribution broke first, so installs fell back to last-click.

That looked like a technical fix. It was not.

Over the following months, our budget for top-of-funnel channels declined. Nobody proposed the cut. Nobody approved it. We noticed much later, and by then the shape of our spending had changed.

Here is what I believe now. Choosing an attribution model is not a measurement decision. It is a funding decision. Your model decides which work can be proven, and proven work is the work that keeps its budget.

1. Attribution decides what you can defend, not just what you can see

Budgets are not set by dashboards. They are set in meetings.

In those meetings, somebody has to defend a line item. The easiest line to defend is the one with a clean number attached. Paid search converted. Retargeting converted. Here is the report.

The hardest line to defend is the one that created demand months earlier. There is no clean number. There is a story.

So teams fund what they can attribute deterministically. Deterministic means you can trace one person from one touch to one sale. It is a high bar. Only some channels clear it.

Performance channels clear it easily, because they sit closest to the sale. Brand, content, partnerships and community rarely clear it at all.

That is why performance gets over-funded. Not because it works better. Because it is easier to prove, and proof is what survives a budget review.

💡 Key Takeaway: Your attribution model does not only measure work. It decides which work can be defended, and work that cannot be defended loses its budget.

2. The channels closest to the sale take credit they did not earn

This is not a theory. It has been tested.

Researchers ran large-scale field experiments at eBay and published the results in Econometrica. Brand keyword ads produced no measurable short-term benefit. Returns from paid search were a fraction of what non-experimental estimates suggested.

For non-brand keywords the picture was worse. Most of the spend reached frequent users who were going to buy anyway. Average returns were negative.

Uber found something similar in 2017. It switched off roughly two thirds of an annual digital budget of about 150 million United States dollars. User acquisition did not drop in any significant way.

The mechanism was simple. Ad networks had inserted themselves into the last click. They claimed credit for installs that were already happening.

Both cases point the same way. Last-click does not report contribution. It reports proximity.

3. Your tool already made the decision, and the number is an estimate

You might assume modern tools solved this. Read the documentation.

Google states that all its attribution models exclude direct visits from receiving credit, unless the whole path is direct. That is a funding rule. Nobody at your company chose it.

The same documentation notes that conversions can be reattributed for up to seven days. It also notes that data-driven attribution may draw on aggregate data from your data-sharing settings.

None of that is measurement in the way people picture it. It is a model with opinions.

The other objection is that mobile measurement has moved on since 2015. It has. That is the problem.

The deterministic link between a person and their behaviour across platforms is gone for most users. Modelling filled the gap. Google now ships an open-source marketing mix model called Meridian, built for what it calls privacy-durable measurement.

My old fallback was visible. Installs went last-click and I could watch it happen. Today the fallback is invisible. A model fills the gap and hands back a confident number.

💡 Key Takeaway: You are not reading a measurement. You are reading an estimate, and its funding defaults were set by someone who has never seen your business.

4. Singapore is already funding the consequence

A Yahoo study of 181 marketing and advertising professionals in Singapore, run in October 2024, found that 72 percent struggle to track performance and attribution across channels.

The same study found that 61 percent prioritise performance marketing over brand building. Only 30 percent intend to focus on brand.

I do not think those numbers are unrelated. I think the first one causes the second.

When you cannot prove cross-channel impact, you retreat to the channels you can prove. That retreat looks like discipline. It is closer to surrender.

Only 21 percent fully use their online first-party customer data. Only 11 percent use offline data. The identity you need is often already inside the building.

5. Start with what you can attribute, not with which model

Most attribution advice opens with the wrong question. It asks which model fits your funnel. The first question is whether you can join one person across the places they meet you.

Answer that, then choose.

• You cannot join users across platforms, and cannot fix it this quarter. Use last-click inside a channel to optimise that channel only. Size everything else with holdout or geographic tests. This starves any channel you cannot click-track, unless you run those tests.
• You cannot join users, but you hold first-party identity you are not using. Fix the join before you buy a model. Logins, order history and customer records are the join. This starves nothing, as long as you test while you build.
• You can join users on one surface only. Use position-based or data-driven attribution on that surface. Never use it to compare across surfaces. This starves any channel whose real contribution happens where you cannot see.
• You can join users broadly. Use data-driven attribution, then run incrementality tests on your two largest line items. This starves nothing systematically, as long as the testing continues.

One rule covers every branch. Write down what your choice starves. Then fund that thing on judgement, rather than waiting for a model to permit it.

Final Thoughts: your model is a funding policy, so choose it like one

Every attribution model is incomplete. That is not the problem.

The problem is that the incompleteness is not neutral. It always favours whatever sits closest to the sale. Left alone, it moves money there quietly, quarter after quarter, and nobody ever signs off on it.

You are walking into a budget review soon. Before you do, ask a different question. Not which model is most accurate. Ask what your current model cannot see, and what that blindness has already cost you.

Then decide what to fund on purpose.

If you want to pressure-test how your growth system allocates budget, book a discovery call or connect with me on LinkedIn.

A note before you close this tab. If your top of funnel has been thinning for reasons nobody can quite explain, the cause may not sit in your strategy. It may sit in your reporting defaults. That is fixable, and the fix starts with naming what your model cannot see.

Mervyn Chua is a growth-transformation consultant helping founders and CEOs build the strategic clarity and systems to grow in an AI-first world. If this raises questions worth exploring for your brand, let’s talk.