The scene repeats in almost every review meeting. The client opens three tabs side by side: Meta Ads claims 180 conversions, Google Ads claims 90, GA4 counts 70, and the e-commerce back office logs 100 actual sales. Four numbers, one reality, and the question lands: "So, who do we believe?" The honest answer is uncomfortable: none is wrong, and none is right. Each looks at the same customer journey through a different lens. Attribution, in 2026, is no longer a tool to configure better; it is a belief to give up, the belief that somewhere out there sits one exact number.
This article takes stock of what GA4 can no longer see, why the platforms will never reconcile, and what concretely replaces the quest for perfect attribution when you are steering a real budget.
Attribution never measured the truth; it distributed credit according to a rule. The real shift of 2026 is that we can no longer pretend that rule describes reality.
GA4 has lost half its toolbox
Let's start with a fact many advertisers missed: GA4 now offers only one configurable attribution model besides last-click. Between May and September 2023, Google removed the four rule-based models (first-click, linear, time-decay and position-based) from Analytics and Google Ads, as Search Engine Land documented when the sunset was announced. Official reason: those models accounted for under 3% of conversions. Since April 2024, whatever remained switched automatically to data-driven, now the default model, as Google's documentation on attribution models confirms.
Data-driven isn't a flaw in itself: it distributes credit through machine learning rather than an arbitrary rule. But it is a black box: you don't know why a given touchpoint gets a given weight, and the model needs a substantial conversion volume to run properly. For most of the SMBs I work with, that volume isn't there: below the threshold, the algorithm falls back on approximations. So we traded a transparent-but-rigid model for a flexible-but-opaque one, and, in both cases, it can only attribute what it sees.
What the tag no longer captures
And what it sees shrinks every year. Three holes have opened up beneath measurement. The first is consent: since Consent Mode v2 rolled out, a share of visitors refuse cookies, and Google fills the gap with modeled conversions, statistical estimates, not real events. Useful so you don't steer blind, but they are probabilities added up as if they were facts.
The second hole is cross-device and cookieless. A prospect sees a Meta ad on mobile on Monday, searches your brand on their work computer on Wednesday, and converts on Saturday through a direct link. Without a stable identifier, GA4 sees three strangers where there is one person, and credits the sale to the last visible link, often "direct" traffic. The third hole is the dark funnel: the entire journey that leaves no clickable trace. A recommendation in a private chat, a podcast, an answer from ChatGPT or Perplexity that cites your brand, a LinkedIn post read without a click. That influence is real, measurable in revenue, yet invisible to a tool that can only read clicks. In GA4, it ends up in the catch-all "Direct" or "Unassigned", the line that keeps growing and that no one can explain.
Why the platforms will never reconcile
Then there's the great misunderstanding: hoping Meta, Google Ads and GA4 will eventually agree. They won't, and it's structural. Each counts by its own rules (different attribution windows, view-through counted or not, last non-direct channel in GA4), and above all, each ad network has an interest in claiming the conversion it "touched". Summing their numbers means counting the same sale several times. That's exactly what I see in review meetings, and what analyses of the gaps in conversions across Meta, Google and GA4 confirm: the totals never match, by design.
The table below sums up the main measurement families and what you can, and cannot, ask of them.
| Approach | What it measures | Its blind spot |
|---|---|---|
| Last-click (GA4 / Ads) | The last channel clicked before conversion | Ignores everything before it; over-credits bottom of funnel (brand, retargeting) |
| Data-driven (GA4) | Credit distributed by machine learning | Black box, volume-hungry, blind to the non-clicked |
| Platform self-attribution (Meta, Ads) | What the platform thinks it influenced | Judge and jury: over-reports its own impact, double counts |
| Incrementality (geo / lift) | The real causal effect of a channel | Requires cutting spend to test; not real time |
| Marketing mix modeling | Each lever's aggregate contribution | Macro view, no keyword- or creative-level detail |
The most expensive mistake
Steering budget on the ad network's own numbers. An advertising platform optimizes what it measures, and measures what makes it look good: trusting its self-reported ROAS to decide where the next euro goes is letting the seller grade their own performance. The healthy reflex is the opposite: anchor the decision on a neutral source (the CRM, the e-commerce back office, the actual number of sales) and use the platforms' figures only to compare trends, never as accounting truth. A channel that "performs" inside Meta Ads but doesn't move real revenue when you switch it off is not performing: it is claiming credit.
Triangulate instead of chasing THE number
If perfect attribution no longer exists, what do you actually do? You triangulate. You cross three imperfect but independent readings, and you act on what they say together, not on the decimal of a single tool. This is precisely the approach that, because individual data is drying up, brings marketing mix modeling back to the fore after years of neglect.
The first pillar is a reliable first-party base: server-side tracking to recover the conversions the browser loses, disciplined and consistent UTMs, and above all the source logged in the CRM or back office at the moment of sale, the one data point no one can dispute. The second pillar is incrementality testing: cutting a channel in one geographic area or over a period, and measuring what actually moves. It answers the question no attribution model can ("what happens if I stop?"), and it's often the most useful one. The third pillar is a lightweight MMM: reasoning at aggregate level (spend per channel against total revenue, week by week) to capture the effect of levers the click can't see, including the dark funnel and brand awareness.
In the field, this triangulation settles debates no dashboard could. For an event ticketing campaign I ran on Meta Ads, it was matching the actual number of tickets sold against spend, not the platform's ROAS, that validated a cost per ticket around €2.66. For a training offer, it was tracking signups tagged in the CRM, not the raw leads the platforms claimed, that revealed a real acquisition cost of €3.73. In both cases, the truth wasn't inside a tool: it was in the cross-check.
In the crucible: distilling signal from noise
Attribution in 2026 isn't broken: it has simply stopped promising what it could never deliver. GA4 remains an excellent behavior sensor; it is no longer a final arbiter. The lead, here, is the single reassuring number you wave in a meeting because it comes from a familiar tool; the gold is the cross-reading, less precise on the surface but infinitely truer, that tells you where the next euro genuinely creates effect. Giving up false precision is where real measurement begins.
If your platforms contradict each other and your "Direct" line keeps swelling with no explanation, the problem isn't your GA4: it's the frame. I make measurement reliable and build this triangulation (server-side collection, source logged in the CRM, incrementality tests and reporting that reconciles channels) as part of my analytics expertise, working hand in hand with day-to-day SEA and SEO steering.
Your numbers contradict each other? Let's reconcile them
I audit your measurement (GA4, server-side, Consent Mode), log the source in your CRM and build a triangulation that tells you where your budget truly creates effect, not what each ad network claims.