What data-driven attribution does differently
Every rule-based model answers the credit question before looking at the data: last click says the closing touch did the work, position-based says the ends did, linear refuses to choose. The rule is a belief about how customers decide, applied identically to a business selling insurance and one selling socks. Data-driven attribution turns the question round. It asks the paths themselves which touchpoints tend to be present when people convert and absent when they do not, and it shares the credit in proportion to that estimated contribution. The attribution models comparison sets it beside the rules it replaced.
Three consequences follow. First, the output is specific to the property: the same channel can earn a large share of credit in one business and little in another, because the paths differ. Second, the credit is fractional. A channel is no longer awarded whole conversions but shares of them, which is why its conversion column shows decimals and why its conversion rate (CVRShare of visits or clicks that convert.) has to be read as a fractional numerator over whole sessions. Third, the model is not fixed. It is re-estimated as data accrues, so the split for a given month can move after the month has closed. None of this makes the model wrong; it makes it a statistical estimate rather than a rule, and estimates need to be read as such.
What it does not change is the scope of observation. The model only ever learns from touchpoints that were recorded, and it learns associations, not causes. Those two limits recur throughout this guide because they decide where the model helps and where it quietly misleads.
How credit is estimated
Google describes the mechanism in its help documentation without publishing the algorithm, and this section stays within that description. The model is trained on both converting and non-converting paths, which is the part that distinguishes it from every rule: a rule only ever looks at the paths that ended in a conversion, while the data-driven model also learns from the far larger set of paths that did not. From these it estimates how the presence of each touchpoint changes the probability of a conversion.
Google characterises this as a counterfactual approach: the model contrasts what actually happened with what could have happened, in effect comparing paths that contain a touchpoint with otherwise similar paths that do not, and treating the difference in conversion probability as that touchpoint's contribution. Google also documents that the model takes account of factors such as the time between an interaction and the conversion, the device type, the number of ad interactions, the order in which exposures occurred and the type of creative involved. Credit for each conversion is then distributed across the touchpoints on its path in proportion to their estimated contribution.
Two further documented behaviours matter when reading reports. GA4's attribution treats direct arrivals specially: the paid and organic last click model ignores direct traffic outright, and in practice data-driven credit for direct is confined to paths that contain nothing else, so a returning customer who types the URL after a campaign touch does not dilute the campaign's credit. And the model is applied to key events within the property's lookback window, so an interaction older than the window is not part of the path at all. Which windows apply, and where the model setting does and does not take effect in GA4's reports, is covered in the GA4 attribution settings guide.
In Google Ads the attribution model is a setting on each conversion action and the data-driven model is the default for most of them. Google documents that it evaluates interactions across its own ad formats, including search, shopping, video, display and demand generation placements, and that the resulting fractional conversions appear in the conversions columns and are used by automated bidding. The Google Ads measurement section covers conversion actions, counting and windows on that side.
Data requirements
The model is only as good as the paths it is trained on, and three properties of those paths matter more than their count.
Completeness of channel tagging. Every touchpoint must be recognisable as a channel. Untagged email links, campaign parameters that fall into a catch-all channel group, and paid social traffic recorded as a plain referral all distort the estimate, because the model attributes the behaviour of a mislabelled touch to the wrong channel. The tagging discipline that makes last-click reports trustworthy is a precondition here too.
Stable identity across the path. A path is only a path if the sessions are connected to one user. Where identity breaks, across devices, after consent is declined, or when a cookie is cleared, the model sees two short paths instead of one long one, and the touches at the start of the real journey are separated from the conversion at the end. Google's identity options for a property, and the share of users they cover, shape the model more than any setting on the attribution page.
Enough history, and enough variety. Google offers the model to every GA4 property without a documented volume requirement, and in Google Ads it removed its earlier volume conditions for most conversion actions; the current conditions are on the help pages in the sources. Availability is therefore rarely the problem. Stability is. With few conversions, each refresh can move credit noticeably, and with little variety in the mix, for example a property where nearly every path is brand search alone, the model has little to learn from. Google does not document how much data makes the estimate stable, and no one outside Google can, so the practical test is whether the split holds steady from one month to the next for the channels that matter.
What it cannot observe
The model's blind spots are the same as the property's, and listing them is the fastest way to decide whether its output can carry a decision.
| Not observed | Why it is missing | Effect on credit |
|---|---|---|
| Impressions without a click | Analytics records arrivals, not exposure | Video and display earn little or nothing, however much they shape demand |
| Offline touchpoints | Television, print, out-of-home, events and word of mouth leave no session | Their effect lands on the next recorded touch, often brand search or direct |
| Consent-declined visits | No identifier, so no path | The channel mix of the observed population may differ from the real one |
| Cross-device journeys without a shared identity | Sessions cannot be joined | Early touches become orphan paths that never convert |
| Touches outside the lookback window | The path is cut at the window boundary | Long consideration cycles lose their beginnings |
| Interactions on other platforms' ads, in a platform's own model | Each platform sees only its own ads | Credit is complete within the platform and silent about everything else |
| The counterfactual | No path shows what the person would have done without the touch | Association is reported as contribution |
The last row is the one that matters most for budgets. The model estimates contribution by comparing paths with and without a touchpoint, but those paths were not assigned at random. A person who clicks a brand search ad was already looking for the brand; a person who sees a retargeting ad had already visited. Paths containing those touches convert more often for reasons that have nothing to do with the ad, and the model, which can only see the paths, credits the ad. The only way to observe what a channel caused is to withhold it from a comparable group, which is what the incrementality testing guide describes.
Reading the output
Read data-driven credit as a comparison, not as a verdict. GA4's model comparison report places the data-driven split beside last click for the same key event and period, and the movement between the two is the information: a channel that gains credit when the model changes is one that tends to appear earlier in paths, an assisting channel; one that loses credit is one that tends to close. That reading answers the question "what role does each channel play?", which is useful for creative, sequencing and tagging decisions even when the absolute figures are uncertain.
Three habits keep the reading honest. Compare periods under the same model, and when a model changes for reporting, restate the history rather than splicing the two series. Snapshot the figures you publish, with the date, because the model refreshes and the snapshot is the only record of what was reported. And keep a last-click view alongside, because it is the one figure every platform can also produce and therefore the only basis for a reconciliation table.
A subscription box reads its model comparison
A meal-kit subscription service compares the two models for the sign-up key event across one quarter in GA4. Under paid and organic last click, brand search is credited with 1,420 sign-ups, paid social with 610, organic search with 980 and email with 390. Under data-driven, brand search falls to 1,060, paid social rises to 940, organic search to 1,050 and email slips to 350. Paid social has gained more than half again, and brand search has lost about a quarter. The team reads this as a role, not a result: paid social introduces people who later search the brand, so its last-click figure understated it. They do not move budget on the strength of the shift. Instead they run a geo holdout on paid social the following quarter, and find that withholding it reduces sign-ups in held-out regions by about 180 per month, roughly 60 per cent of the monthly credit the data-driven model gave it. The model was directionally right and quantitatively generous, which is the normal outcome.
When not to use it
Data-driven attribution is the best default reporting model available in the Google tools, and there are still situations where it is the wrong instrument.
- When the decision is a budget move between channels. The model cannot see the counterfactual; use it to form the hypothesis and an experiment to test it.
- When most of the media is unobserved. A brand built on television, out-of-home or podcasts is giving the model paths that start at the wrong place, and the credit lands on brand search. Mix modelling, introduced in the attribution hub, is the tool for aggregated, unobserved media.
- When volume is small or the mix has just changed. The estimate will be unstable, and a new channel has no history to be learned from. A fixed rule, clearly labelled, is more honest until the data settles.
- When the number must reconcile across systems. Finance, platforms and analytics can all produce last click; only each tool's own model can produce that tool's data-driven split.
- When the lookback window is much shorter than the buying cycle. The model is learning from truncated paths, and its confident-looking split describes the last part of the journey only.
Which instrument fits which decision, and how a team moves from last click to a fuller measurement stack, is the subject of the guide to choosing an attribution approach.
Common mistakes
- Treating fractional credit as proof of causation. It is an estimate from observed paths.
- Comparing this month's data-driven figures with last month's as reported then. The model moved.
- Running it on untagged or badly grouped traffic. The model learns the labelling errors.
- Dropping the last-click view entirely. Reconciliation becomes impossible.
- Moving budget on a small shift in credit. Noise in a low-volume property looks like insight.
- Assuming Google Ads and GA4 data-driven credit agree. Different scopes, different training data, different windows.
- Reading a short-window split as the whole journey. The model only learned from the part of the path the window kept.
The attribution models comparison places the model beside the rules it replaced, and the attribution hub sets out where experiments and mix modelling take over from any model.
