What is data-driven attribution

For years, splitting the credit for a conversion was a decision taken before anyone looked at the data. Last click gave everything to the final interaction, first click to the first, the linear model split it in equal parts and the position-based model assigned fixed percentages to the ends. They all had one thing in common: the rule was chosen first and the data fitted itself around it afterwards.

Data-driven attribution reverses that order. Instead of applying a split decided in advance, it calculates how much each interaction actually contributes within the paths of that particular account. If in your data brand search almost always appears at the end of paths that were going to convert anyway, that interaction will receive less credit than last click would give it.

The change is no longer optional. In Google Analytics 4, the first click, linear, time decay and position-based models have been unavailable since November 2023, leaving data-driven attribution, paid and organic last click, and Google paid channels last click. In Google Ads, data-driven attribution and last click coexist, and conversion actions that used the removed models were migrated to data-driven attribution.

The raw material is the complete paths, and not only the ones that end in a conversion. The model also looks at paths that did not convert, because comparing the two groups is what makes it possible to tell an interaction that pushes from one that was merely present.

The calculation runs in two stages. First a probability model is built from the recorded paths, then credit is assigned according to how much the estimated conversion probability changes when each interaction is added to the path. An interaction that barely moves that probability receives little credit, however close to the end it sits.

Among the signals feeding the calculation are the time elapsed since the key event, the device type, the number of ad interactions, the order of exposure and the type of creative asset. In Google Ads, clicks and video engagements across Search, YouTube, Display and Demand Gen are taken into account.

The model works better with volume. Google recommends at least 200 conversions and 2,000 ad interactions in supported networks within a 30 day period, although conversion actions are eligible even with less data. It is also worth knowing that figures get readjusted: a conversion can be reattributed for up to seven days afterwards, and direct visits are excluded from the split unless the whole path is direct.

Here is the detail that breaks half of all comparative reports: each model is specific to each advertiser and to each key event. There is no general data-driven attribution model that Google applies to everyone. There is yours, trained on your paths.

The consequence is uncomfortable for agencies. Two accounts in the same sector, with the same campaigns, the same ads and similar budgets, will receive different credit splits, and neither of them is wrong. Each one correctly describes the behaviour of its own traffic. That is why it makes no sense to bring the sentence “in this sector, brand takes such and such a share of the credit” into a meeting, or to carry an observed split from one client to another.

Within a single account the comparison is genuinely useful, with one caveat: the model retrains on new data. A changed split between quarters may reflect a real change in buying behaviour or simply mean that the model has learned from a different campaign mix. Before reallocating budget, check which of the two happened.

• Define the key event properly before looking at the split. The model trains per event, so measuring five different conversions of unequal value produces five readings that cannot be added together.
• Gather enough volume before making fine-grained decisions, with 200 conversions and 2,000 interactions in 30 days as a reasonable floor.
• Look after path quality: clean campaign tagging, a correctly implemented consent setup and first-party data reduce the gaps the model has to bridge.
• Compare periods within the same account and document the structural changes (new campaigns, swapped creatives) that could explain a different split.
• Let the seven day reattribution window pass before closing the report for a period.
• Explain the model to the client before showing the numbers, especially if they used to read their results on last click.

• Comparing the credit split of two accounts and concluding that one of them measures badly, when both models are correct for their respective data.
• Publishing sector benchmarks about how much credit each channel receives under data-driven attribution, a figure that does not transfer from one account to another.
• Treating the split as causal truth and cutting a channel's budget because of its percentage, with no controlled test to back it.
• Comparing a period measured on last click against one measured on data-driven attribution and calling the difference an improvement or a drop.
• Looking in the interface for the first click, linear, time decay or position-based models, which have been unavailable in Google Analytics 4 since November 2023.