Data-Driven Attribution

Your Google Ads conversion counts can shift in a week when nothing in the account changed. No new campaign, no budget move, no new creative. The report simply reassigned credit for the same sales.

That reassignment is the attribution model working. For nearly every advertiser it is now data-driven attribution, and it arrived as the default rather than as a choice.

Data-driven attribution (DDA) uses machine learning to spread credit for a conversion across the ad interactions that preceded it. It compares the paths of people who converted against the paths of people who didn’t, then estimates what each touchpoint added. Last click gives everything to the final interaction; DDA divides it.

The model you’re on unless you changed it

DDA is the default for new conversion actions, and last click is the only other option left in the dropdown, per Google’s attribution documentation. The older rule-based models — first click, linear, time decay, position-based — were retired, so a menu that once held six choices now holds two.

A brand that never opened the attribution settings is still on DDA. The switch also reshapes historical reporting: because credit is recalculated, the same past week can read differently after a model change than it did before, as practitioner guides describe.

Nothing about the campaigns changed. The accounting did.

This carries downstream, because Smart Bidding optimizes toward whatever the attribution model reports. Automated bidding chases the conversions DDA credits, so the model quietly sets the target the bidder aims at.

When the pieces line up, the upside is documented. T-Mobile’s Austrian account recorded 22% more conversions, a 23% higher conversion rate and a 27% lower cost per acquisition after moving from manual, last-click-era bidding to data-driven attribution paired with Smart Bidding, per a Think with Google case study. The attribution change and the bidding change are credited together there, not separately.

The 200-conversion floor, and what happens below it

DDA is not available to every account. To become eligible, a conversion action needs 3,000 ad interactions and 300 conversions over 30 days; to stay eligible, it must keep generating 2,000 ad interactions and 200 conversions every 30 days, per Google’s own requirements.

Fall below either number and the consequence is specific. The conversion action reverts to last-click attribution. The setting may still read data-driven in places, but the credit is now being assigned by the model DDA was meant to replace.

This is the failure case the “more accurate by default” framing leaves out. A small or seasonal D2C account, or a single conversion action that doesn’t clear the floor, can be running last click under a data-driven banner with nothing flagging the difference.

What the model can’t see, and won’t show

Even when DDA runs, its arithmetic is closed. You can read the outputs — this channel got 30% of the credit, that one 70% — but the weights and logic behind them are never exposed to the advertiser, per an analysis of the model’s limits. There is no view that shows why a touchpoint earned what it earned.

The scope is also narrower than the name suggests. DDA scores Google touchpoints only, so branded search tends to absorb credit created by channels the model never observes, per the same analysis. It leans toward interactions with cleaner click data, which credits lower-funnel search more readily than upper-funnel demand generation, per a review of where DDA falls short.

For a brand that runs Google and Meta as separate machines, that is a real blind spot. Little Sleepies runs 524 Google ads, about 60% of them Shopping, against a Meta account built on drops and scarcity that shares almost none of its vocabulary. Google’s DDA can score the first account and is structurally unable to see the second.

I think that is the part worth sitting with. The model reports a customer journey while missing whichever half of it happens off Google.

Google both writes the attribution algorithm and owns the platform that receives the credit it assigns, which critics call a structural conflict of interest that the closed model makes impossible to audit, per one measurement teardown.