What an attribution model decides
A conversion path is the sequence of recorded sessions or ad interactions for one converting person within the lookback window. The model answers one narrow question: given that this conversion happened and these touchpoints were observed, how should the credit be divided among them? Everything else is decided elsewhere. Tracking, consent and cross-device identity decide which touchpoints are recorded at all. The lookback window, a setting rather than a model, decides how far back a path can reach. The key event definition decides what counts as a conversion in the first place.
Three things people expect from a model that no model can deliver. It cannot tell you whether a touchpoint caused the conversion, because it only ever sees paths that happened and never the path that would have happened without the touch; that is the job of IncrementalityConversions caused by marketing. testing. It cannot reconcile numbers across ad platforms, because each platform runs its own AttributionAssigning conversion credit to touchpoints. over the touchpoints it alone can see. And it cannot credit a channel it never recorded, which is why offline media, word of mouth and consent-declined visits are silently absent from every model's output.
In GA4Google's event-based analytics platform. the model setting is narrower than most people assume. It changes the event-scoped attribution dimensions used in the Advertising reports and in key-event credit; the user acquisition report keeps a first-session view and the traffic acquisition report keeps a session-level view whatever the setting. The GA4 attribution settings guide covers the setting, the windows and the reports; this page covers the models themselves.
The models side by side
Rule-based attribution models vs Data-driven attribution
| Model | Credit logic | Favours | Misleads when | Available in |
|---|---|---|---|---|
| Last click | All credit to the final recorded click before the conversion | Brand search, retargeting, email, affiliate codes | Journeys are long and demand was created upstream | Nearly every ad platform within its own scope; most analytics tools |
| Paid and organic last click (GA4) | Ignores direct arrivals and gives all credit to the last channel the person clicked through | The same closing channels, with direct removed from the picture | Returning customers arrive direct after an earlier campaign touch | GA4 reporting attribution setting |
| First click | All credit to the first recorded touch in the window | Prospecting display, paid social, content, referral | The first touch was incidental and the real work happened later | Retired from GA4 and Google Ads reporting; custom analysis on exported paths; some third-party tools |
| Linear | An equal share to every touch on the path | Long paths and whichever channel appears most often | Repeated low-value touches, such as many retargeting sessions, swamp the path | Retired from GA4 and Google Ads reporting; third-party tools; custom analysis |
| Time decay | More credit the closer a touch sits to the conversion | Closing channels, more gently than last click | Consideration is long and early touches did the persuading | Retired from GA4 and Google Ads reporting; third-party tools; custom analysis |
| Position-based | Larger fixed shares to the first and last touches, the remainder spread evenly across the middle | Both ends of the path | The middle is where the decision was made, or paths are so short that it collapses into last click | Retired from GA4 and Google Ads reporting; third-party tools; custom analysis |
| Data-driven | Credit estimated from converting and non-converting paths by comparing paths with and without each touch | Whatever the data shows shifts the chance of converting | Volume is low, history is short, or a channel's touches are not recorded | Default in GA4 and Google Ads; other platforms offer their own versions |
| Platform-scoped last click | Last click or qualifying view within one platform's own ads only | The platform doing the counting | Figures from several platforms are added together | Every ad platform's own conversion reporting |
The rule-based models
Rule-based models are easy to explain, cheap to compute and stable over time, which is exactly why they survive even where data-driven attribution is available. Their weakness is the same in every case: the rule is an assumption about how people decide, applied to everyone.
Single-touch models
Last click credits the final click. It is the oldest model, the one most ad platforms use inside their own reporting, and the one finance teams understand without explanation. It rewards channels that sit at the end of a journey, which is why brand search, retargeting and email look excellent under it and prospecting looks poor. Its advantage is consistency: because nearly every tool can produce a last-click figure, it is the natural baseline for reconciling platforms and analytics, and the platform versus analytics comparison leans on it for that reason.
GA4's version, paid and organic last click, removes direct arrivals from consideration: a person who clicked a paid social ad last week and typed the URL today is credited to paid social, not to direct. That is a sensible correction for a model that would otherwise reward the absence of tracking, but it also means that GA4 last click and a platform's last click are not the same rule.
First click is last click reversed. It rewards whatever introduced the person, which makes it a useful lens on demand creation and a poor one on closing. It is rarely the right reporting model but is a good question to ask of path data: which channels appear first, and does the mix of first touches change when prospecting spend changes?
Multi-touch rules
Linear splits credit equally. It sounds fair and is the least informative of the set, because it treats a single, decisive click and the eighth retargeting session identically. In practice a linear view mostly rewards channels that generate many sessions per converter.
Time decay gives more credit to touches closer to the conversion and less to earlier ones, which fits businesses with short consideration cycles and promotional calendars. For a long business-to-business evaluation it behaves much like last click with extra steps, since the decisive early demonstration is a long way from the signed contract.
Position-based gives the two ends a larger fixed share each and spreads the rest across the middle. It is the rule that most closely matches how marketers already think about introduction and closing, which is both its appeal and its risk: it confirms a story about the funnel rather than testing it. Google has retired these three rules from GA4 and Google Ads reporting, so where they still appear it is in third-party tools or in your own analysis of exported paths.
The tutorial on building an attribution model in SQL does that analysis on the GA4 BigQuery export, rebuilding last non-direct, first, linear and position-based credit with the weights and lookback window as parameters.
Data-driven in brief
Data-driven attribution replaces the fixed rule with an estimate learned from the property's data. Google documents that its model evaluates converting and non-converting paths, uses a counterfactual comparison of what happened against what could have happened to judge how each touchpoint changes the chance of a conversion, and takes account of factors such as the time between an interaction and the conversion, device type, the number of ad interactions, the order of exposure and the creative involved. Credit is then shared in proportion to the estimated contribution, which is why a channel's conversions appear as fractions rather than whole numbers.
Its strengths are real: the split reflects observed behaviour rather than a belief, and it adapts as the mix changes. Its limits are equally real. It still sees only recorded touchpoints, so an unseen channel cannot be credited. It is estimated from observational data, so a channel that reaches people who were already on their way to converting can still earn credit. And its output moves as the model is refreshed, so last month's credit may not match what the report showed last month. The data-driven attribution guide covers the mechanism, the data it needs and the cases where it is not the answer.
One path, six models
A furniture retailer exports a single converting path for a £900 sofa order: a paid social click on day 1, an organic search visit on day 4, a comparison-site referral on day 9, and a brand search click on day 12 that leads to the purchase. Last click gives the full £900 to brand search. First click gives all of it to paid social. Linear gives each of the four touches £225. Time decay, with a seven-day half-life, gives roughly £410 to brand search, £255 to the referral, £160 to organic and £75 to paid social. Position-based, in its common 40/20/40 configuration, gives £360 each to paid social and brand search and £90 each to the two middle touches. The property's data-driven model, trained on thousands of paths like this one, assigns £470 to the referral, £200 to paid social, £150 to brand search and £80 to organic, because in this property paths that include the comparison site convert far more often than paths that do not. Same order, same touches, six different reports: the model is a lens, not a measurement.
Choosing for a decision
The right model depends on what the number is for. The table is a starting point; the choosing an attribution approach guide turns it into a full decision framework that also covers experiments and mix modelling.
| Decision | Use | Because |
|---|---|---|
| Comparing campaigns inside one ad platform | That platform's own model, usually data-driven | Same scope and rules for every campaign; it is also what automated bidding optimises towards |
| Reporting channel performance to the business | One analytics model, held constant, with data-driven preferred where volume supports it | Every conversion is counted once and the series is comparable month to month |
| Reconciling platform figures with analytics | Last click on both sides | It is the only rule every tool can produce in the same shape |
| Understanding which channels assist rather than close | Model comparison: data-driven or first click against last click | The shift in credit between models is the assist signal |
| Moving budget between channels | An incrementality test, informed by the model | Attribution cannot say what the conversions would have been without the channel |
| Judging conversion rate (CVRShare of visits or clicks that convert.) by channel | A fixed model with the denominator stated | A fractional numerator over a session count only makes sense when the rule is known |
Two practical rules follow. Keep one model constant for any series you report over time, and when you change it, restate history under the new model rather than splicing two rules together. And treat model differences as questions, not answers: a channel that gains credit under data-driven relative to last click is probably assisting, and the way to confirm that is to test it, as the budget allocation guide recommends before moving money. The attribution hub has an interactive comparator for trying paths against each rule.
Common mistakes
- Searching for the true model. There is none; there is only the rule that fits the decision.
- Reading data-driven credit as causation. It is a better description of observed paths, not a test.
- Changing the model and keeping the old history. The series breaks at the change date.
- Comparing a platform's last click with analytics last click. Different scopes, different direct handling.
- Judging prospecting on last click and retargeting on first click. Each channel measured by the rule that flatters it.
- Letting the model carry a budget decision alone. Confirm with an experiment first.
The data-driven attribution guide goes deeper on the one model that is not a rule, and the GA4 attribution settings guide shows where the model choice takes effect in the reports and where it does not.
