Why Google Ads Attribution Fails Ecommerce Brands

You open GA4. Google Search accounts for 54% of your attributed conversions. Meta is at 11%. You shift spend toward Search, cut Meta, and watch blended ROAS improve in-platform for two weeks — then fall apart.

What happened? You made a budget decision based on a model Google built to measure Google.

The Quiet Attribution Model Migration

In 2023, Google retired four attribution models from Google Ads and GA4: first click, linear, time decay, and position-based. Campaigns using those models were automatically migrated to data-driven attribution (DDA) — or to last click if conversion volume was too low.

Most brands didn't choose this migration. They woke up one day and their conversion credit was distributed by an algorithm they can't inspect.

This wouldn't be a problem if DDA were neutral. But every attribution model needs data to work — and the data it has access to shapes what it can see.

Why Google Naturally Appears at the Bottom of Every Funnel

Here's the journey most ecommerce purchases actually take. A customer sees your brand mentioned in a YouTube video they weren't looking for. Two days later, they click an Instagram Story. A week after that, a friend mentions your product. The customer searches your brand name on Google, clicks a Shopping ad, and buys.

In last-click attribution: 100% of the sale goes to Google Shopping.

In data-driven attribution: Google distributes credit, but it can only weight touchpoints it can observe. Meta's ad impression, the YouTube organic mention, and the friend referral aren't in Google's dataset. The channel that built the intent gets zero; the channel that captured the final click gets the attribution.

This structural bias isn't unique to Google. But Google is where most purchase intent gets expressed as a search query — and that makes it the last recorded touch for most conversions, almost by definition.

Branded search is almost always the final step before purchase. When you run paid brand campaigns alongside organic demand you've already created through social, email, or influencers, Google Ads captures credit for intent those other channels generated.

Data-Driven Attribution: Useful Model, Unavoidable Blind Spots

DDA does something last-click doesn't: it distributes credit fractionally across multiple touchpoints. That's genuinely more sophisticated.

The limitation is what it can't see. Google's DDA draws on data from Google-owned surfaces — Search, Shopping, Display, YouTube, Gmail, Maps. It has no visibility into a paid Meta campaign, a podcast placement, or an offline event. When it "credits" a non-Google channel, it's partly inferring from what it knows, not fully measuring what it can't observe.

There's also a volume threshold. Google's own documentation states that DDA requires at least 400 conversions within 28 days to function properly. Brands below that volume default to last click — often without realizing it. The lower your conversion volume, the more your attribution looks like "whoever got the final click."

Performance Max: A Black Box Inside a Black Box

Performance Max campaigns package Search, Shopping, Display, YouTube, and Gmail into a single automated unit. Everything is optimized together and reported as a combined ROAS.

Practical Ecommerce has documented what practitioners have been saying for years: a PMax campaign's reported return can bundle brand keywords — searches your loyal customers make regardless of paid ads — with genuinely incremental prospecting, and you can't separate them by default. As the publication noted, "a campaign with a 500% return on ad spend could include brand and nonbrand terms."

Without the ability to isolate brand from non-brand spend, and one placement from another, you cannot tell whether Performance Max is creating demand or simply capturing it.

Google has added incremental controls — brand exclusions, placement reports, search term insights — but the fundamental opacity remains. You're bidding on a combined portfolio that cannot be fully disaggregated, and the reported ROAS reflects that combined result.

The Data Loss That Biases the Whole Picture

Even if the attribution model were neutral, GA4 tracking has structural gaps that systematically favor Google's own reported contribution.

Ad blockers. Plausible Analytics ran a controlled experiment comparing standard Google Analytics against a proxy-routed tracker on viral content. In tech-forward audiences, 58% of visits went untracked by Google Analytics — 68% on desktop. The users GA4 can't see don't disappear from reality; they just disappear from your attribution model.

iOS App Tracking Transparency. Since Apple's ATT framework launched in 2021, users must explicitly opt in to cross-app tracking. As of Q1 2024, AppsFlyer reports a 44% opt-in rate in the US — meaning roughly half of iPhone customers aren't fully trackable by browser-side pixels. Google's server-side infrastructure survives this better than third-party pixels do; brand search clicks remain measurable. Meta, TikTok, and display impressions often don't.

When GA4 loses a touchpoint to ad blockers or ATT, that touchpoint disappears from the attribution pool. The clicks GA4 can reliably log tend to be on Google's own properties. The resulting channel mix reflects what GA4 can measure — not what actually happened.

What Post-Purchase Surveys Reveal Instead

Ask customers "How did you first hear about us?" and you bypass the entire pixel stack. No cookies, no consent prompts, no data gaps. Just the customer's memory of what put your brand on their radar.

This isn't a perfect signal — recall fades, survey response rates vary, and some channels are easier to remember than others. But it has one property no platform can match: it is independent of any ad platform's interests.

When you compare survey responses to GA4's channel breakdown at scale, a consistent pattern emerges:

Word-of-mouth and organic sources are systematically underreported in GA4. Customers who discovered you through a friend, a podcast, or an organic post often show up in GA4 as a Google Search conversion — because that's what they clicked last.
Upper-funnel paid channels get less GA4 credit than they deserve. A customer whose discovery was driven by a Meta video, but who purchased a week later via brand search, is "a Google conversion" in last-click reporting.
Brand search is often a symptom, not a cause. The survey tells you how the intent was created. GA4 tells you how it was expressed. Both are true — but only one should drive your prospecting budget decisions.

"The post-purchase survey doesn't tell you everything about attribution. It tells you the one thing no platform can: where the customer's journey actually started."

This matters concretely. If survey data shows 35% of customers first heard about you through Instagram, but GA4 attributes 9% to Meta, you're likely over-investing in branded search to capture demand that Meta already created. Cutting Meta in that scenario won't reduce waste — it will eventually reduce the branded search volume you were crediting to Google.

Building an Attribution Baseline That Doesn't Depend on Google

The goal isn't to replace GA4 with survey data. It's to stop using any single platform to judge its own performance.

A practical baseline combines three signals:

Post-purchase survey data captures the discovery channel — the customer's honest answer to where their journey started.
GA4 + UTM tracking shows the click path — useful for flow analysis and catching technical gaps.
Blended ROAS (total revenue ÷ total ad spend across all channels) is the gut check that can't be gamed by any single platform's reporting.

When survey data and GA4 agree, act with confidence. When they diverge significantly — Google Search claiming 4× more attribution than the survey supports — that gap is where the bias lives, and where budget decisions go wrong.

For a complete framework on combining these signals, see Multi-Signal Attribution: Survey, Pixel, and UTM Together. If you're also working through Meta attribution gaps, Why Your Meta ROAS Is Lying to You covers the parallel problem in detail.

You wouldn't let Google score its own exam. Build your independent attribution baseline with Rauxdata — and see what the surveys say.