Article
Multichannel attribution: changing the model won't make the numbers agree
Moving from last click to data-driven is the right thing to do and takes five minutes. It won't close the gap between Meta and GA4, because almost none of that gap's causes depend on the model.
Moving from last click to data-driven is the right thing to do and takes five minutes. It won’t close the gap between Meta and GA4, because almost none of that gap’s causes depend on the model. The problem isn’t that the numbers differ. It’s that nobody tells you which part of the difference is normal.
Category: attribution · 10 minute read · Mattia Minafò · updated 18 August 2026
What this is about
You look at Meta Ads Manager and see 50 purchases. You look at GA4 for the same month and see 14. Someone tells you the pixel is broken, someone else that you need data-driven attribution, and just in case you cut the budget of the channel that looks most expensive.
The question this article answers: does changing attribution model reduce the gap between platforms? The short answer is almost never, and the useful part is understanding what actually does reduce it.
In order: why the two numbers differ by construction, what happens to budget when you decide on last click, why changing the model moves little, and in what order the checks should be run.
Method and limits. Statements about how the platforms work come from official documentation, verified on 18 August 2026 and linked at the end. Project numbers come from audits we ran ourselves, with the time window stated case by case. This is not an implementation guide and it does not replace a check on your own setup.
You’ll find three labels in the text. Documentation when the source is official, measured by us when the number comes from one of our audits, recommendation when it is a professional judgement rather than a fact.
Contents
- Why the two numbers will never be identical
- What happens to budget when you decide on last click
- Why changing the model moves less than you think
- What happens if you ask every source the same question
- In what order to separate the normal part from the lost part
- When last click is still fine
- The most common mistakes
Why the two numbers will never be identical
Last click gives 100% of the credit to the last channel clicked. A user arrives from an organic post, a search, a newsletter and a retargeting ad, then types the domain by hand and buys. As far as last click is concerned, zero out of four channels contributed.
This is not an implementation flaw. It is the model doing what it was written to do, back when journeys were short and on a single device.
Documentation. GA4 has three models left: data-driven, paid and organic last click, Google paid channels last click. First click, linear, time decay and position-based have not been available since November 2023.
Documentation. In Google Ads, data-driven is the default for most conversion actions, announced on 27 September 2021 with rollout from October that year. Last click was never removed and remains selectable.
Then there are three things the model does not touch, and they explain most of the gap.
The windows. Meta also counts people who saw the ad without clicking it, on a one-day window. GA4 never counts views. Two systems counting different events over different periods do not reach the same total, and that is not a fault.
What counts as a click. Since 3 March 2026 Meta counts as click-through only clicks on the link, with rollout continuing through late March and April. Before that, likes, comments and post expansions counted too. People who measured the effect across multiple accounts report an average 20-25% drop in click-through conversions, higher on campaigns built around engagement. Those aren’t lost conversions, they are attributions that used to be counted and shouldn’t have been.
Consent in Europe. For users in the European area who have not consented, Google cannot use the click identifier for attribution. The result is that a Google Ads campaign can appear in GA4 as Direct or unassigned even though the user really did click the ad.
What happens to budget when you decide on last click
The damage from last click is not the wrong row in a report. It is the decisions that follow from that row.
On a store with a purchase cycle longer than three days, the channels that close the last click are almost always Direct, Organic Search or Email. Paid channels work earlier, and on last click they look expensive. The same journey, read through two models, produces two incompatible budget allocations.
The numbers in this table are made up for illustration. They exist to show the mechanism, they do not describe a real case.
Journey: prospecting ad, organic search, email, retargeting ad, a €100 purchase.
| Model | Prospecting | Organic | Retargeting | |
|---|---|---|---|---|
| Last click | €0 | €0 | €0 | €100 |
| Data-driven (example) | €20 | €15 | €15 | €50 |
With the first model prospecting returns zero and the rational decision is to shut it down. With the second it is worth a fifth of the sale.
There is a second effect that makes this hard to see. When you cut an entry channel, conversions don’t fall the next day. They fall when the audience that channel had already brought in runs out, that is weeks later, when nobody connects effect to cause any more.
Recommendation. Before cutting a channel’s budget because cost per acquisition looks high, look at how often that channel appears as the first touchpoint in GA4, under Advertising, Attribution, Conversion Paths.
Why changing the model moves less than you think
This is the part that stops the article ending with “switch to data-driven and you’re done”.
An attribution model distributes credit among the conversions you recorded. It says nothing about the conversions you recorded badly, or didn’t record at all.
Three measurements from our audits, showing where the model ends and everything else begins.
Measured by us. On a Shopify store, comparing the order ledger with GA4 over 90 days, 62% of online revenue was attributed to no source at all. It wasn’t a recent fault. Extending the window to twelve months, only 55% of real orders had the user journey populated. The cause was neither the model nor GA4: it was Shopify failing to associate the order with a browsing session. No attribution model recovers an order that was never linked to a visit.
On the same store, across 92 orders in twelve months, a single order carried UTM parameters. No model, however sophisticated, distributes credit among channels that were never labelled.
Measured by us. On another property, two channel groupings defined slightly differently returned results 97.8% apart over the same period and on the same data. It was a match condition written with “exactly matches” instead of “contains”, on campaign names carrying a prefix.
Measured by us. On the same property, 17% of sessions arrived with no source and no medium, worsening after a release. The interesting part is why. Some of it was traffic that used to be lost entirely and that server-side tracking had started to see, arriving however without a provenance label. Another part was a genuine fault, a consent code snippet removed from the theme.
That 17% is the clearest illustration of the argument. The same number contained something getting better and something getting worse. The correct actions were opposite, and no report flagged it. We separated the two causes, but we did not measure how much each weighs, and we wrote that in the document instead of estimating it.
Changing attribution model touches none of these four cases.
What happens if you ask every source the same question
Measured by us. A client reported 50 purchases in Meta Ads Manager against 14 in GA4 for the same month, a ratio of almost four to one. The natural reading is that three conversions out of four had been lost.
We redid the comparison changing one thing only: the same window and the same question for every source. How many real orders came in between 31 May and 27 June, and how many each source sees.
| Source | Purchases | Distance from the order ledger |
|---|---|---|
| Store order ledger | 41 | reference baseline |
| Meta Events Manager, deduplicated | 37 | 90% |
| GA4 | 36 | 88% |
The real gap was ten or twelve points, not four times. The 50 and the 14 were not two measurements of the same thing: they covered different periods and answered different questions.
The remainder has three causes and only one is fixable. View-through conversions and the window difference are structural. Missing UTM parameters on the Meta links are not, those are configuration, and they were the reason GA4 saw less.
Recommendation. Before opening a tracking ticket, check that the two figures you are comparing cover the same period and answer the same question. In our experience it is the check that closes more reports than any technical intervention.
In what order to separate the normal part from the lost part
Order matters more than the individual checks, because each step changes the baseline of the next.
| Situation | What the documentation says | What we recommend |
|---|---|---|
| Total sales in the platforms don’t match the back office | No advertising platform certifies revenue | Start here. Order ledger against platforms over 90 days, before looking at any other number |
| The two numbers you compare come from different reports | Each surface has its own scope; Ads Manager only shows attributed events | Realign window and question before hunting for a technical cause |
| Ad links carry no UTM parameters | Tracking parameters have to be set by hand in the platform | Fix this before touching the model. It is the most frequent cause and the cheapest to solve |
| Meta shows more conversions than GA4 | Meta includes view-through conversions, GA4 doesn’t | Exclude views and compare clicks only. What remains is cross-device and browser restrictions |
| The active model in GA4 is still last click | GA4 offers three models and the switch is not retroactive | Change it. Five minutes, no risk, historical data stays as it is |
| Monthly purchase volume is low | Google recommends at least 200 conversions and 2,000 interactions in 30 days for data-driven, and notes it works with fewer | Below that threshold, read conversion paths by hand and don’t trust the algorithmic weighting |
| A share of sessions arrives with no source | GA4 classifies it as unassigned | Don’t treat it as a single value. Compare it with the period before the last release, because it can contain both recovery and fault |
| Channels don’t match between two reports on the same property | Custom grouping definitions are independent | Check the match condition before drawing any conclusion about traffic |
Recommendation. The first three rows come before the other five. If revenue isn’t reconciled and links aren’t labelled, every percentage calculated on top will be formally correct and substantially false.
When last click is still fine
Three cases where it remains a defensible choice.
Purchase cycle under 24 hours. If nearly all purchases happen within the same session, last click really does capture the journey. Impulse products, low ticket, purchase straight from the ad.
Bottom-of-funnel retargeting only. If you work exclusively on people who already put something in the cart, the last click reasonably reflects the channel’s contribution.
Very low volume. Below twenty purchases a month, any multi-touch model has too much variance. Last click is at least deterministic and adds no noise.
The limit that doesn’t change in any of the three: last click is never sufficient grounds for shutting down an entire channel.
The most common mistakes
Changing the model and considering the job done. The model redistributes credit among recorded conversions. Check first how many you are recording.
Comparing two numbers covering different periods. It is the mistake that produces the most spectacular gaps and the ones that close fastest. Same window, same question, then you can reason.
Leaving ads without UTM parameters. The platform’s click identifier isn’t enough for GA4 to build the channel. Without UTMs, paid traffic ends up in Direct or among unassigned sessions.
Reading unassigned traffic as a single fault. It can grow because the system has started seeing traffic it used to lose. Compare it with the period before the last release.
Reacting immediately to the March 2026 drop on Meta. The change in what counts as a click lowered everyone’s click-through conversions. Compare your drop with the average before rewriting campaigns.
Trusting data-driven on low volumes. Google recommends 200 conversions and 2,000 interactions in 30 days. Below that the model still runs, but the weights it assigns are worth little.
Changing the model and comparing with previous months. The switch is not retroactive, so the time series breaks exactly where you changed it.
In short
Moving to data-driven should be done, costs five minutes and breaks nothing. Don’t expect the platforms’ numbers to converge, because the model acts on one piece of the problem only.
The difference between two platforms always has a normal component. The useful work is knowing how large that component is, and that number appears in no report.
Sources
- Google Analytics Help, About attribution and attribution modeling, retrieved 18 August 2026
- Google Ads Help, About attribution models, retrieved 18 August 2026
- Google Ads Help, About data-driven attribution, retrieved 18 August 2026
- Google Blog, The future of attribution is data-driven, 27 September 2021, retrieved 18 August 2026
- Google Analytics Help, Attribution settings, retrieved 18 August 2026
- Meta Business Help Center, About attribution, retrieved 18 August 2026
- Matteo Tobia, analysis of the Meta attribution change of 17 March 2026, published 31 March 2026
- Internal measurements from audits of Shopify and WooCommerce stores, windows stated in the text
The first step
Take your store’s order ledger for the last 90 days and put it next to the total transactions in GA4 for exactly the same period. If the distance exceeds 20%, the attribution model is not your main problem.
It is a check you can run on your own in half an hour. If the number you get doesn’t add up, get in touch and we’ll look at it together.
Content verified against official documentation on 18 August 2026. Platforms change often; if you are reading this much later, re-check the links.