Ask GA4 which channel produced a lead and you can get three different answers depending on which report you open. That is not a bug. It is attribution, and once you understand the three models GA4 actually uses and where each one shows up, the reports stop contradicting each other. This guide explains the models with one campaign followed all the way through, and ends with a practical rule for which number to trust for which decision.
What attribution is deciding
A buyer rarely arrives once. A typical B2B path looks like this: a LinkedIn ad, a week later a Google search for your brand, three days after that a click from your newsletter, and then a demo request. Four touches, one conversion. Attribution is the rule for dividing credit for that one conversion among the four touches. Every model is a different rule, and no rule is “true.” Each answers a different question.
The three models GA4 uses
Google retired first click, linear, time decay and position based models from GA4 in late 2023. What remains in 2026 is three models, and they are applied in different places without always saying so.
| Model | Rule | Where GA4 uses it | Question it answers |
|---|---|---|---|
| Data-driven | Machine learning distributes credit across touches based on how much each one raised the probability of converting, using your property’s own data | Default for conversion attribution in most reports, adjustable in Admin | “Which touches actually moved people toward converting?” |
| Last click, paid and organic | All credit to the final non-direct touch | Available as an alternative reporting model; the basis for the Traffic acquisition report’s session dimensions | “What was the last thing they did before converting?” |
| First click, user-scoped | All credit to the first recorded touch for that user | User acquisition report, via First user source and medium | “How did we first find this person?” |
Two details trip people up. First, “direct” traffic is never given credit when any other touch exists in the lookback window, in any model. A user who first came from an ad and later typed your URL is attributed to the ad. Second, the lookback window is a setting, not a constant: by default 30 days for acquisition events and 90 days for everything else, and Admin lets you shorten both. A shorter window pushes more credit to later touches.
One campaign, three reports
Follow a single buyer to see how the numbers diverge. Say you are running the campaign from our UTM naming convention example: a LinkedIn sponsored post promoting a pricing calculator, tagged linkedin / paid_social / leadgen_pricing_calculator_202609.
Here is the buyer’s path:
- Day 1: clicks the LinkedIn ad, reads the calculator page, leaves.
- Day 6: searches your brand name on Google, clicks the organic result, reads two case studies.
- Day 9: clicks the pricing link in your newsletter, tagged
newsletter / email. - Day 9, same session: submits a demo request. That is the conversion event.
Now open three reports:
| Report | Dimension used | Model applied | Who gets the demo request |
|---|---|---|---|
| Reports, Acquisition, User acquisition | First user source / medium | First click | linkedin / paid_social gets 1.0 |
| Reports, Acquisition, Traffic acquisition | Session source / medium | Last click, session scoped | newsletter / email gets 1.0 |
| Advertising, Attribution, Conversion paths, or any report using Source / medium without a scope prefix | Event scoped source / medium | Data-driven (the property default) | Fractional: for example linkedin 0.45, google organic 0.15, newsletter 0.40 |
All three reports are correct. The user acquisition report is telling you LinkedIn found this person. The traffic acquisition report is telling you the newsletter closed them. The data-driven report is telling you the ad and the email both mattered and the brand search barely did. If you export all three and add them up, you will count the same demo three times, which is the most common attribution mistake there is.
Where each model is right and wrong
Data-driven
Best for deciding how to split budget across channels, because it is the only model that tries to measure incremental contribution rather than position. Its weakness is opacity: you cannot see the weights, they change over time, and a property with fewer than a few hundred conversions a month gets a model trained on thin data. For small properties Google quietly falls back toward last click without saying so in the interface.
Last click
Best for judging what converts, meaning which landing pages, offers and channels are present at the moment of decision. It is also what most ad platforms report, so it is the model to use when reconciling GA4 against Google Ads or LinkedIn. Its weakness is that it starves awareness channels: paid social and content almost never get last-click credit in B2B, so a last-click budget review cuts them first and then wonders why the pipeline dries up two quarters later.
First click
Best for measuring top-of-funnel efficiency and for answering the boardroom question “where do our customers come from.” Its weakness is the mirror image of last click: it ignores everything that happened after the introduction, so a channel that introduces many people who never engage again looks excellent.
A rule for which number to use
| Decision | Use | Why |
|---|---|---|
| Which landing page or offer converts best | Last click, session scoped | You want the page present at conversion |
| Reconciling with an ad platform | Last click | Same model on both sides, differences become explainable |
| Splitting next quarter’s budget between channels | Data-driven | Only model that rewards assists |
| Reporting to leadership on where customers originate | First click | Answers the question as asked |
| Calculating cost per qualified lead by channel | Data-driven for the split, last click as the sanity check | See the CPQL method in our benchmarks article |
Pick one model per decision, write it down, and use it consistently. Switching models between quarters makes trends meaningless.
Setting it up correctly
- Check the property model. Admin, Attribution settings. Data-driven is the default; leave it unless you have a specific reason. This setting affects the event-scoped dimensions in all standard reports, not just the Advertising section.
- Check the lookback windows. Same screen. For B2B with cycles over a month, set both to the maximum, 30 days for acquisition conversion events and 90 days for all other conversion events. Shorter windows silently delete early touches.
- Mark the right events as key events. Attribution only applies to key events, formerly conversions. If your demo request is not a key event, none of this is being calculated.
- Use scoped dimensions deliberately in explorations. When you build a free-form exploration, the dimension picker offers First user source, Session source and plain Source. Each carries a different model. Label your reports with which one you used.
- Keep tagging clean. Every model is only as good as the source and medium values it sees. Untagged email and uppercase UTM values scatter credit across duplicate rows. The UTM tracking guide and the GA4 vs Universal Analytics comparison cover what changed in how GA4 reads tags.
Two things GA4 attribution cannot do
It cannot see offline. A lead that came from a conference and later searched your brand is credited to organic search unless you import the offline touch, which is an advanced setup most teams skip. And it cannot see across devices reliably without Google signals or a logged-in user ID; a LinkedIn tap on a phone followed by a demo request on a laptop looks like two people. For B2B, both gaps push credit toward late, branded, desktop touches. Discount those slightly when you read the reports, and if you have a CRM, compare against the self-reported “how did you hear about us” field once a quarter. When the two disagree badly, the CRM is usually closer to the truth about the first touch and GA4 is closer about the last.
Summary
- GA4 runs three models: data-driven, last click and first click. They appear in different reports through scoped dimensions, often without labeling.
- The same conversion can be credited to three different channels across three reports. Never add reports built on different models.
- Use last click for what converts, data-driven for budget allocation, first click for origin reporting.
- Set lookback windows to the maximum for B2B, mark the right key events, and keep UTM values clean or every model degrades.