---
title: "Attribution Models Explained | Gudu"
description: "Why Google, Meta and GA4 all claim the same sale, what each attribution model actually does, and how to pick the one number you run the business on."
canonical: https://gudu.com.au/attribution-models-explained/
---

# Attribution models, explained without the jargon

By Jordan Parrello · September 8, 2026

Six sales last month. Google claims nine, Meta claims eight, your email platform claims five, and your accountant, who is right, says six.

Nobody is cheating. Each platform is answering a slightly different question, and none of them is answering the one you actually asked.

## Why the numbers never add up

Each platform only sees its own touch. Google knows about the search ad and not about the Instagram video the customer watched a fortnight earlier. Meta knows the reverse. Both then take full credit, because from where each of them sits, they caused it.

They count different events. Google mostly counts a conversion that follows a click. Meta will also count one that follows a view, meaning somebody saw the ad, didn’t touch it, and bought inside the window. That’s a defensible thing to measure. It just isn’t the same measurement.

And the windows differ. Whether a conversion counts within seven days of a click, or thirty, or ninety, is a setting rather than a fact. Two accounts configured differently will disagree about the same month, and the disagreement means nothing.

## What an attribution model actually is

A rule for splitting credit between the things a customer touched before they bought. That’s the entire concept.

Someone reads a blog post in March, sees a Meta ad in April, searches your brand name in June and buys. Four touches, one sale. The model decides how much of that sale each touch gets, and there is no true answer, only the rule you picked.

Worth saying plainly, because attribution gets discussed as though one model is correct and the rest are broken. They’re all wrong. Some are usefully wrong.

## The models, plainly

| Model | How it splits the credit | What it makes look good |
| --- | --- | --- |
| Last click | All of it to the final touch | Brand search, retargeting, anything close to the sale |
| First click | All of it to the first touch | Awareness campaigns and the blog |
| Linear | Evenly across every touch | Whatever produced the most touches, useful or not |
| Time decay | More credit the closer a touch is to the sale | The same channels as last click, less brutally |
| Position based | 40% each to first and last, the rest shared | The ends of the journey, at the middle’s expense |
| Data driven | Modelled from your own converting and non-converting paths | Nothing consistently, but you can’t see the working |

Google’s data-driven model is the default in most accounts now and it’s generally the best of them, with a catch worth knowing. You can’t audit it, and it needs a decent volume of conversions before it does anything clever. On a small account it behaves much like last click while telling you it doesn’t.

## What your choice does to your budget

Last click makes brand search look like your best channel. Naturally it does: the customer typed your name, and something earlier taught them the name. Cut that something and you find out what it was doing, usually a quarter too late.

First click does the reverse and makes awareness spending look untouchable, which is convenient for whoever owns the awareness budget.

The practical point is that the model quietly reallocates money. If someone changes the attribution setting and then reports a channel improving, they’ve reported a settings change.

## Pick one scoreboard

The most useful thing you can do here has nothing to do with models. Decide which system is the truth for the business, and report against it every month regardless of what the platforms say.

For most companies that’s GA4, or better, the CRM where deals get marked won. You want the system furthest downstream that still knows where the customer came from, because that’s the one that knows the difference between a lead and a customer.

Then stop summing platform numbers. Adding Google’s nine to Meta’s eight to arrive at seventeen is the most common reporting error in digital marketing, and it persists because the answer makes everybody look good.

## What the platform numbers are still for

Optimising inside the platform. Which ad, which audience, which keyword, which creative to switch off. For those calls the platform’s own count is fine, because you’re comparing like with like inside one system, and the bidding algorithms need conversions fed back to them or they learn nothing at all.

Just don’t take those numbers into a board meeting. Different job, different audience.

## What no model can see any more

The picture has got blurrier every year and it isn’t coming back.

Tracking prevention in browsers and on iOS removes part of the path outright. People research on a phone and buy on a laptop, and stitching those together only works where somebody is logged in. Anything that happens in an email client, a group chat or a phone call is invisible by definition. And a growing share of journeys now begin inside an AI answer, which arrives as a direct visit or an odd referral if it’s labelled at all.

The healthy response is to stop expecting attribution to behave like an audit and start treating it as an estimate with error bars. Directionally useful, never exact.

## The tests that beat the models

When a decision genuinely matters, don’t model it. Test it.

Switch a channel off for a fortnight and watch total revenue rather than that channel’s reported conversions. Or run it in some states and not others for a month. Or take the crudest version of all and plot total spend against total revenue over a long enough window to see whether the line moved when you moved the money. None of that is elegant, and all of it survives the next browser update.

Add one manual habit worth more than most of the tooling. Ask new customers how they heard about you, in the form or on the call, and write the answers down. It’s messy self-reported data and it will occasionally be the only evidence you have that a channel exists.

## Start by fixing what you count

Most attribution arguments turn out to be measurement problems in a costume. Before you touch a model, check that a conversion means the same thing everywhere: same definition, same window, same deduplication, and no form submission counted twice because somebody hit refresh.

Southern Phone came to us unable to see where users were arriving from or how they moved through the site, with campaigns that had been set up and then left untended on top of that. The rebuild fixed the tracking and the buying together, and account cost per acquisition came down by $22 while the business chose to reinvest 68% more spend the following year. The order matters: the efficiency was only provable because the measurement got sorted first. The [Southern Phone case study](https://gudu.com.au/southernphone/) has the detail.

If your platforms are arguing over the same conversions and you want somebody to referee, that work sits across our [technology](https://gudu.com.au/services/technology/) and [advertising](https://gudu.com.au/services/advertising/) teams, because it’s never only one of the two.