How data-driven attribution works
Instead of handing all the credit for a conversion to one click, data-driven attribution distributes it across the clicks in the journey according to how much each appears to have contributed. Google builds that judgement by comparing paths that converted with paths that did not, then assigning fractional credit accordingly. A keyword that consistently appears in journeys which end in a sale earns more; one that appears just as often in journeys that go nowhere earns less.
The important word is modelled. Nobody observed a customer giving a third of their decision to one advert. The split is an estimate produced by a system, refreshed as more data arrives, and it is not something you can audit line by line.
Why it matters
It fixes a real distortion. Under single-click models, the terms people search right before buying — usually brand names and precise product phrases — collect all the credit, while the broader research terms that introduced the business collect none. Cut those upper-funnel terms because they look unprofitable, and enquiries fall in a way the report never predicted. Distributing credit makes that contribution visible.
It also feeds bidding. Automated bidding acts on the credit it is given, so a model that values assisting clicks will bid differently from one that ignores them. In practice that usually means more willingness to pay for earlier-stage searches, which is often correct and occasionally expensive.
Where it goes wrong
The trap is treating modelled figures as counted ones. Because the model is refreshed, credit for past periods can be restated after you have reported it. A campaign figure you screenshotted for a client last month may not be the figure showing today, and there is no error to find — that is the system working as designed. Any report that presents these numbers as a fixed record will eventually embarrass whoever wrote it.
The second problem is fractions in the wrong places. Credit is split, so campaign conversion counts arrive with decimal parts, and they do not add up the way a sales ledger does. The third is cross-platform comparison: Google’s modelled credit and Meta’s own reported figures are produced by different systems with different rules, so adding them together produces a total no business ever earned.
What to do about it
Use it for direction and let something else be your record. Bank statements, invoices and CRM records are counted facts; attribution reporting is an estimate of contribution. When the two disagree, the ledger wins and the model tells you where to look.
In reports, date-stamp any modelled figure and say plainly that it may be restated — an honest footnote costs nothing and prevents an awkward conversation later. Compare like with like by keeping one model for a full reporting period rather than switching mid-quarter, and read the split alongside a simple single-click view such as last click attribution so you can see how much of the picture is model rather than click. Understanding the wider family of attribution models makes those conversations with clients far shorter.
