Reading Your Shopify Attribution Data | PMD

You log into Shopify on Monday morning and last week's ROAS was 3.4x. You open Triple Whale and it's 2.2x. GA4 says 4.1x. Meta Ads Manager claims 6.8x. Your CFO emails asking which one to trust for the board deck. You stare at the screen.

This is the modern Shopify attribution problem, and it isn't going away. Every brand we work with at PM Digital Design runs at least three reporting tools simultaneously, and they always disagree. The mistake most teams make is trying to reconcile them. You cannot reconcile them. They measure different things using different rules over different windows. The job isn't to make them agree — it's to know which one to listen to for which decision.

Below is the framework we run internally and roll out for clients. Seven principles. Opinionated. Designed to stop attribution noise from ruining your weekly decisions.

1. Pick ONE source of truth for Shopify attribution data — never mix

The single biggest cause of bad decisions is teams quoting blended ROAS from Triple Whale on Tuesday and from GA4 on Thursday. The numbers will differ by 30-60% depending on the week, and the second you mix them you've lost the plot.

Pick one tool as your source of truth for blended ROAS. Pick another (or the same) for unit economics — contribution margin per order, LTV, payback period. Write it down. Tell the team. If anyone quotes a different source in a Monday meeting, the answer is "that's interesting, but the source of truth says X."

The choice matters less than the commitment. Triple Whale, Northbeam, and Polar all have credible methodologies. What kills you is switching between them based on which one tells the nicer story this week. Our view on which tool fits which stage is in the profit optimisation framework.

2. Watch incremental, not platform-attributed

Meta will happily tell you it drove £180k of revenue last week. Google says it drove £140k. TikTok claims £90k. You spent £85k total. The platforms have collectively claimed 4.8x of your actual revenue. They are all telling the truth from their own perspective and they are all lying to you in aggregate.

The number that matters is incremental, what wouldn't have happened without that channel running. The brutal truth is that for most mature brands, branded search, retargeting, and even a meaningful chunk of prospecting Meta would have converted anyway. Platform-attributed ROAS overstates incremental ROAS by 30-70% depending on the channel.

You don't need a fancy MMM to act on this. You need to be sceptical when platforms self-report and aggressive about running geo holdouts on the channels you suspect of double-counting.

3. Cohort-level revenue beats session-level revenue

Session-level reporting tells you a £40 ad acquired a £55 first order. Looks bad. Cohort-level reporting tells you that same £40 ad acquired a customer who spent £218 in their first 90 days. Looks great.

Subscription brands and high-LTV verticals get destroyed by session-level thinking. You will kill profitable channels because they look unprofitable on day one. The decision rule should always be: "what did this acquisition cohort do over the payback window?" not "what did the first session do?"

This is the same lens we apply to retention work in our subscription optimisation guide. First-order revenue is a vanity metric for subscription businesses. Cohort revenue at day 60 or 90 is the real one.

4. Triangulate big claims across two or more sources before acting

The cost of being wrong on a small claim is small. The cost of being wrong on a big claim, like "Meta is dead, shift the budget to TikTok", can be catastrophic. The rule we use with clients: if you're about to move more than 15% of monthly spend based on a finding from one tool, force yourself to verify it in a second tool before pulling the trigger.

Triangulation doesn't mean the numbers have to match. They won't. It means the directional story has to match. If Triple Whale says Meta is up week-on-week and GA4 also shows Meta-tagged sessions converting up, you've got two independent signals pointing the same way. If only one source shows it, treat it as a hypothesis, not a fact.

5. The five attribution tools, when to trust them, when to ignore them

Here's the cheat sheet we hand clients on day one. It's deliberately opinionated.

Source	What it measures	Trust it for	Ignore it for
Shopify reports	First-party orders, AOV, repeat rate, cohort revenue	Ground-truth revenue, AOV trends, cohort retention	Channel attribution (Shopify's UTM stitching is rough)
GA4	Session-level, last-non-direct click by default	Traffic shape, on-site behaviour, conversion paths	Absolute revenue (it always undercounts vs Shopify)
Triple Whale	Pixel + post-purchase survey + click-tracking blend	Daily blended ROAS, channel-level directional trends	Absolute incremental claims, sub-channel granularity
Northbeam	Multi-touch + MTA model with view-through	Upper-funnel credit, journey-stage attribution	Daily decisions (model is noisy on small windows)
Ad platforms	Self-reported, biased upward, click + view windows	In-platform optimisation, creative-level CTR/CVR	Cross-channel decisions, blended ROAS, board reporting

Nobody gets this perfectly right. The brands we see making the best decisions tend to use Shopify as ground truth for revenue, one third-party tool as the daily blended ROAS read, and ad-platform numbers strictly for in-platform decisions.

6. Distrust any model that won't show its assumptions

If a vendor cannot tell you their attribution window, their view-through rules, how they handle direct traffic, and how they treat returning customers, walk away. Every model has assumptions. The credible vendors publish them. The ones who hand-wave through it are hiding something, usually the fact that their "AI" is a glorified last-click with extra steps.

"If your attribution vendor can't tell you their assumptions in one sentence, they're not measuring your business. They're selling you a vibe."

The same scepticism applies to in-house dashboards. Whoever built the dashboard made decisions. Those decisions are baked into every number it spits out. If the person who built it has left, you don't have a source of truth, you have a haunted spreadsheet.

7. Build a weekly "what moved" report, never a daily one

Daily attribution data is mostly noise. Day-of-week effects, payday cycles, weather, a Reddit thread you didn't see, they all swamp the actual signal. Brands that check ROAS at 9am every morning end up reactive, not strategic. They cut channels on a bad Tuesday and bid them back up after a good Saturday.

Run a weekly Monday review instead. Three questions: what moved week-on-week, what's the four-week trend, what are we doing differently this week. Anything that didn't move 10%+ is noise. Anything that did move 10%+ gets one of three labels (random, seasonal, or causal) before any decision is made.

This discipline alone has stopped more bad decisions than any tool upgrade we've ever recommended. The framework is similar to the one we run in our broader CRO audit framework. Measure cadence matters as much as measure source.

8. Hold-out tests beat MMM beats platform attribution

The hierarchy of attribution credibility, ranked from most to least trustworthy: geo holdout, then a properly-built MMM, then any multi-touch model, then platform self-reporting. Most brands do this hierarchy backwards.

A two-week geo holdout on Meta will tell you more about your incrementality than six months of Northbeam dashboards. The catch is that it requires real discipline, real cell sizes, and the willingness to learn that you've been paying for traffic that would have converted anyway. We saw exactly this lesson in the analysis behind the £25,535/month cart-threshold split test. Controlled tests reveal what dashboards cannot.

If you're a mid-market Shopify brand spending £150k+ per month on paid, you should be running at least one holdout test per quarter. Not because the result will be fun, but because it's the only number in your stack that isn't compromised by self-interest.

Where to go from here

Most attribution misery is self-inflicted. Brands buy three tools, never pick a source of truth, never document assumptions, and then wonder why the team makes contradictory decisions. The fix is process, not software. Pick one tool. Document the rules. Run a weekly review. Test incrementality once a quarter.

If you want a second pair of eyes on how your stack is structured, the PMD profit optimisation team can audit it. We look at the gap between what your dashboards say and what your bank account is actually doing. The CRO learning hub has more on how we approach measurement. And if you'd rather just talk it through, grab a 30-minute slot with Paddy directly.

FAQs

Which Shopify attribution tool is best in 2026?

There is no "best", there's the one that matches your channel mix and your decision cadence. Triple Whale tends to suit daily-decision brands with heavy Meta spend. Northbeam suits brands that need upper-funnel credit. Polar suits finance-led teams. Shopify's own reporting is the only one that won't lie to you about revenue. Pick based on what decisions you actually make, not which logo looks prettiest.

Why does GA4 always show less revenue than Shopify?

GA4 relies on client-side tracking, which loses 10-30% of conversions to ad-blockers, iOS privacy settings, and tracking gaps. Shopify sees every order on the backend. Trust Shopify for revenue totals, use GA4 for behaviour and journey shape. Don't try to reconcile them. You'll lose your weekend and find no answer.

How do I explain attribution discrepancies to my CFO?

Pick one source of truth for blended ROAS and put it in the board deck. Footnote that platform numbers will be higher and GA4 will be lower because they measure differently. CFOs respond well to "we picked X because Y". They respond badly to three different numbers with no commentary. Pre-empt the question.

Are MTA models like Northbeam worth the money?

For brands spending £100k+ per month across three or more channels, usually yes. They surface upper-funnel credit that last-click misses. Below that spend level, the model noise tends to exceed the signal. And no MTA model is a substitute for actual incrementality testing. Treat MTA as a useful hypothesis generator, not a verdict.

How often should I run holdout tests?

Once per quarter at minimum on your largest channel. Once per month if you have the spend and the geographic spread to do it cleanly. The biggest mistake we see is teams running one holdout, getting a result they didn't like, and never doing it again. The result you don't like is usually the most valuable one.

What's the single biggest attribution mistake brands make?

Mixing sources. Quoting Triple Whale on Monday, Meta on Wednesday, GA4 on Friday. The numbers will never reconcile and the team will make contradictory decisions. Pick one. Stick with it. Note the others' biases. Move on.