Your data-driven attribution is probably last click

GA4 will show you data-driven attribution in the interface while quietly serving you last click underneath. It happens whenever you sit below the volume thresholds, which most mid-market Australian businesses do, and there is no warning label.

The thresholds are specific. Data-driven attribution requires at least 400 conversions for that particular conversion action, and at least 20,000 total conversions across all actions in the property, within the lookback window. Miss either one and the model falls back. The reports still say data-driven.

That is worth sitting with, because those reports are being used in this country every week to decide where six-figure budgets go.

Why the thresholds exist

Data-driven attribution is a model, and models need training data. GA4 compares converting paths against non-converting paths and works out how much each touchpoint moved the probability. With 60 conversions a month there is nothing to compare. The model would be fitting noise, so Google does not let it run.

That is a defensible engineering decision. The reporting decision, not to flag the fallback clearly in the interface, is the problem.

How to check your own account

Look at conversions per action for the last 90 days, not total conversions. This is where most people get it wrong. Twenty thousand across the property does not help you if the action you are optimising sits at 180.

Then count how many conversion actions you have marked as key events. If you marked newsletter signups, PDF downloads, video plays, phone clicks and form submissions all as conversions, you have spread your volume across five actions and none of them clears 400 while the property total looks healthy.

What to do at low volume

There are only two attribution models left in the Google stack. First click, linear, time decay and position based were removed: unavailable for new conversion actions from June 2023 and force-migrated in September 2023. So the choice is last click or data-driven, and if you do not qualify for data-driven, you have one model.

Given that, here is the sequence that actually helps.

1. 1.Consolidate your conversion actions. Pick one primary action that represents real commercial intent, usually a qualified lead or a purchase. Mark it as the primary. Demote the rest to secondary. You will concentrate volume on one action and get closer to 400, and you will stop bidding towards signups from people who wanted a PDF.

1. 1.Lengthen what you can. The threshold is measured within the lookback window, so a longer window includes more conversions. Check your acquisition lookback setting; the default dropped to 30 days and a 90 day window both reflects reality better and gives the model more to work with.

1. 1.Turn on the BigQuery export. Event-level data lands raw and unmodelled. Whatever GA4 does to your reporting, the underlying paths are yours and you can compute your own path analysis when you have accumulated enough.

1. 1.Stop treating the attribution report as the answer. This is the important one.

The thing attribution cannot do

Even at full volume with a properly trained model, data-driven attribution answers one question: given that a conversion happened, how should credit be distributed across the touchpoints that preceded it. It does not answer whether the conversion would have happened without them.

Those are completely different questions and only the second one decides budget.

Brand search is the clearest example. It converts beautifully in every attribution model because it sits closest to the purchase. Turn it off in a region for four weeks and you find out how much of it was incremental and how much was people who were coming anyway. Businesses routinely discover that a meaningful share of branded paid clicks were cannibalising their own organic result.

You find that with a holdout, never with a report.

The test that beats the model

Geo holdouts work at mid-market scale and cost nothing but nerve. Split Australia into matched regions by revenue and population, turn a channel off in one set, leave it running in the other, run it for at least four weeks and compare total revenue rather than platform-reported conversions. You do not need statistical software to read a large effect, and if the effect is not large enough to read, that is itself the finding.

At larger spend, marketing mix modelling has become far more accessible. Google's Meridian is now embedded inside Analytics 360, with scenario planning built in. It has real requirements, roughly two years of weekly data and genuine variation in spend, and it produces ranges rather than numbers. Ranges are more honest than the four-decimal figures in an attribution report.

Google also announced Qualified Future Conversions at Marketing Live in May 2026, a predictive metric extending attribution out to 180 days after an interaction, on the basis that only 40% of Demand Gen conversions land within the first 30 days. It is in restricted pilot with broader beta expected late in 2026. Worth watching, not worth planning around yet.

What to do this week

Open your GA4 property, count conversions on your primary action over the last 90 days, and write the number down. If it is under 400, every attribution comparison anyone has shown you from that property was last click wearing a different name.

Then decide what you will actually turn off first to find out what is real.