What data-driven attribution needs before it works

The first is about eligibility. Google states that all conversion actions are eligible for data-driven attribution regardless of conversion or interaction volume, and that selecting it for a conversion action makes that action use it. There is no gate. Nothing stops you.

The second is about performance. The same page recommends at least 200 conversions and 2,000 ad interactions in supported networks within a 30-day period so the model can analyse the data accurately, and says the model will still function with less data while noting that sufficient volume lets it identify patterns and assign credit more precisely.

So the honest reading is: you can always turn it on, and the recommendation tells you when the output is worth arguing about. Neither sentence promises anything about a small account, and neither says the model is wrong below the line. It says the credit assignment is less precise, which is a different and much more modest claim than the one usually made in a pitch.

It also means nobody can tell you what data-driven attribution will do to your reported campaign performance in advance. Google describes the model as specific to each advertiser and computed from that advertiser's own data. A number from someone else's account is not a forecast for yours.

Notice what the two recommended numbers say when you divide them. Two hundred conversions against two thousand ad interactions is a ten percent conversion rate. That is a high rate for most lead generation and a very high one for ecommerce.

Which means for most accounts the interaction figure is not the binding constraint. If your account converts at three percent, then reaching 200 conversions takes roughly 6,667 interactions, more than three times the interaction recommendation. You will clear the second number long before you clear the first.

That is worth knowing because it tells you which lever moves you. Buying more clicks at the same conversion rate is the expensive route. Raising the conversion rate, or consolidating several thin conversion actions into one that reflects the same business outcome, is usually the cheaper one.

The arithmetic below is arithmetic, not data. The cost per click figures are inputs I have chosen to make the shape visible, and you should substitute your own before quoting any of it.

Read down the third column and the problem states itself. A business converting at three percent on eight dollar clicks needs a monthly search budget in the tens of thousands to reach the recommended volume on one conversion action.

Most of the mid-market accounts I am asked to look at are nowhere near that on a single action. They are often near it in aggregate, across four or five conversion actions that each count something slightly different, which is not the same thing and does not help the model.

That gap is the useful finding, and it is usually fixable without more budget. If a quote request, a phone call and a booking are three conversion actions representing one business outcome, consolidating them raises the volume on the action that matters instead of raising the spend.

One caution on consolidation. Merging conversion actions changes the signal that automated bidding runs on, because Smart Bidding optimises toward the conversion actions you marked primary. Sequence it deliberately, do it once, and do not do it in the same week as a budget change.

The model still runs. Google's page is explicit that data-driven attribution functions with less data, and it also notes that depending on data availability, last click and data-driven can produce the same results in certain situations.

That last line is the one worth quoting to anyone selling a switch as an upgrade. In a thin account you may get output that is indistinguishable from the model you replaced, and you will have spent a meeting explaining it.

It does not mean do not switch. It means do not build an expectation of a performance change around it. The reason to run data-driven in a small account is that it is the default and it will improve as the account grows, not that it will find hidden value this quarter.

And do not go looking for the change in a report you took too early. Attributed conversions land against the interaction date, so recent periods are still filling in, which is the lag that governs every read of the account and will happily disguise or invent an effect if you measure a week after the change.

Google states that data-driven attribution looks at website, store visit and Google Analytics conversions from Search including Shopping, YouTube, Display and Demand Gen ads. Read that list against your own conversion actions before you assume it covers the thing you care about.

The boundary implied by the list is the important part. This is a model of interactions with Google ads. It has no view of an organic visit, an email click or a referral, so it is not a cross-channel answer and was never offered as one.

If your question is which channel deserves the budget, no Google Ads attribution model answers it, at any volume. That question needs a source that observes all channels, and even then the answer is a division of credit rather than a measurement of cause. The distinction between those two is covered in the plain-language reference on what each model is defined to do.

There is a quieter prerequisite underneath all of this, which is that the conversions being modelled are real and unduplicated. Google's own conversion measurement documentation is the place to check what each action is set to count, and what conversion tracking is defined to record. A model fed by an action that fires twice will distribute credit for events that did not happen twice.

Open the conversion actions list and write down four things per action: what user action fires it, whether it is primary or secondary, the counting option, and the conversion window. Four columns, one page.

Then count last month's conversions per action, not per account. That single number tells you where you sit against the 200 recommendation, and it is the number almost nobody has to hand when the conversation starts.

Then look for the same business outcome appearing in more than one row. That is where volume is hiding, and consolidating it is free.

Then check whether anything on the list is a soft engagement someone marked primary years ago. A newsletter signup counted as a conversion inflates the volume figure and degrades every model that runs on it, which is worse than being below the line honestly.

In a multi-location business, the constraint is usually not volume at all. It is that conversion events fire without a location parameter, so per-site cost per acquisition cannot be produced under any attribution model. Raising volume does not fix an unmeasurable breakdown.

It also produces a specific trap. The instinct is to create one conversion action per location, which divides a single well-populated action into many thin ones and moves every one of them further below the recommendation. You buy a report and pay for it in the quality of the signal bidding runs on.

The version that works is one action carrying a location parameter, which keeps the volume together and still lets you segment. That was the shape of the fix at a clinic group with more than fifteen sites, where correcting it at the event layer is what made per-location quarterly reporting possible at all.

And where several systems each report their own conversion count and nobody can say which one the model should be trusting, the model is the last thing to change. That is a conversion attribution audit: quantify the gaps, separate structural variance from defects, and declare one authoritative source per decision before anybody argues about attribution again.