If you run a store with real sales and real ad spend, you have already lived the problem. Meta claims one return on ad spend, Google claims another, your analytics shows a third, and your checkout survey names a channel where you run no ads. Each number is "right" in its own frame. The question is which frame deserves your next budget decision.
This guide skips the "it depends" non-answer that most ranking pages stop at. It names the models, states each one's built-in bias, and gives you a five-step playbook for deciding what to actually do when the numbers disagree.
There is no "correct" attribution model — only lenses with known bias
Every model answers the same question — which touchpoint gets credit for this order? — with a different rule. None is correct. Each is a lens with a predictable distortion, and the distortion is the part that moves money.
The hard truth top articles gloss over: changing your model changes credit allocation, not revenue. If switching to a model "improves" a channel, nothing real happened in your business. That is why model choice is the wrong obsession and triangulation is the right one. For the full vocabulary and how credit flows between touchpoints, see the attribution and tracking guide.
The models, ranked by the one job each is good for
- Last-touch (last-click). All credit to the final interaction. It overweights closing channels — branded search, retargeting, email — and has driven more bad budget calls than any other model because it crowns the channel that harvested demand others built.
- First-touch (first-click). All credit to the first interaction. Useful for one question only: which channels introduce new customers. Useless for judging what closes.
- Last non-direct click. Last-click that skips "direct" and walks back to the most recent real referrer. This is Shopify's default order model over a thirty-day window (WeltPixel).
- Linear. Equal credit to every touch. Fair-sounding and rarely actionable — it flatters channels that appear often but move nothing.
- Position-based (U-shaped). Roughly forty percent to first, forty to last, the rest spread through the middle, so discovery and closing stay visible at once (ReferralCandy).
- Time-decay. Weighted toward touches near the purchase. It quietly talks you out of upper-funnel spend, because awareness always sits early.
- Data-driven attribution (DDA). A machine-learning model — and GA4's default — that learns each touchpoint's contribution from your own conversion paths and splits one order fractionally across channels. It is less arbitrary than rule-based models, but it is a black box you cannot audit and it needs volume to be stable. A commonly cited threshold is around three hundred conversions a month before its splits settle down (ReferralCandy); below that, the fractions are noise.
One nuance that decides how much any of this matters: model choice only matters in proportion to how long your customer journey is. For a one-session impulse buy, first-touch and last-touch are usually the same event. For a considered purchase with many touches, different models give materially different answers. For a deeper breakdown of each rule, read attribution model in marketing, explained.
Where your tools already sit (and why they will never agree)
You do not pick most of these models from scratch — your tools come with defaults.
| Tool | Default model | Lookback |
|---|---|---|
| Shopify Analytics | Last non-direct click | Thirty days |
| GA4 | Data-driven (DDA) | Key-event based |
| Meta Ads Manager | Its own window-based self-credit | Seven-day click / one-day view |
| Google Ads | Data-driven, in-platform conversions | Platform-defined |
GA4 removed first-click, linear, time-decay, and position-based models in November of 2023 and made data-driven the default — you literally cannot choose those in GA4 anymore (Google Analytics Help). Shopify, meanwhile, does not ingest your ad spend on any plan, so native Shopify reports cannot compute ROAS at all (Polar Analytics).
The structural consequence: Shopify's last-non-direct-click, GA4's DDA, and Meta's and Google's self-credit are three different questions asked of the same order. Their channel rows will never match, and no setting exists to make them.
Why the gap keeps widening: Apple's privacy timeline
A big chunk of your "mystery direct" traffic is tracking decay, not brand love. Apple broke click-based tracking in layers. iOS 14.5's App Tracking Transparency prompted roughly ninety-six percent of US iPhone users to opt out at launch (Flurry); the industry-wide opt-in rate had only climbed to about thirty-five percent by the second quarter of 2025 (Adjust).
Then iOS 17's Link Tracking Protection began stripping user-level click IDs like fbclid and gclid from links opened in Apple Mail, Messages, and Safari Private Browsing — while leaving your UTM parameters intact (Klaviyo). This matters because Safari is roughly fifty-five percent of US mobile sessions (Backlinko), so for a US store this is half your mobile traffic, not an edge case. The takeaway: click IDs are the fragile layer, UTMs are the durable one, and disciplined UTM tagging got more important after 2023, not less.
The best attribution setup for marketing: a five-step playbook
This is the part that beats the SERP. Defining models is easy; knowing what to do when they disagree is the real query.
Step 1 — Fix the self-inflicted layer first. Audit your UTMs: lowercase everything, use a documented source and medium vocabulary, and never tag internal links, which resets the session's real source mid-journey. Most "Unassigned" and mystery "Direct" traffic is a tagging problem, not a measurement mystery — and it is free to fix. The same discipline keeps your email attribution clean; see Klaviyo UTM tracking.
Step 2 — Give each tool its one honest job. Shopify tells you how many orders and how much revenue (server-side truth). Platform dashboards are good for comparing campaign A to campaign B inside the same platform, because the same bias applies to both. GA4 shows cross-channel journey shape. Never use one platform's self-credited number to compare against another platform — each grades its own homework.
Step 3 — Add the lens no pixel has: a post-purchase survey. A one-question "How did you hear about us?" on the confirmation page catches zero-click channels — podcasts, word of mouth, organic social — that click tracking structurally misses. These single-question surveys routinely hit forty to sixty percent response rates (KnoCommerce), which makes them a real dataset, not anecdote. Treat them as directional evidence of influence, not an order-level replacement.
Step 4 — Anchor everything to blended profit. Total ad spend divided by total new-customer orders (blended CAC), and total revenue divided by total ad spend (blended MER), computed from Shopify plus your ad invoices, cannot be gamed by any attribution model. If platform numbers improve while blended numbers stall, credit is being reshuffled, not created.
Step 5 — Reserve causal tests for the big claims. When one channel's claimed contribution is large enough that being wrong changes your budget, run an incrementality test. Meta's Conversion Lift needs at least a seven-day window and ten percent of the audience held out (Haus). These are periodic calibration, not an always-on dashboard — and they genuinely cost you the paused sales.
A worked example: turning ROAS into profit
Say you run 340 orders a month at a $31 average order value, spending $2,800 on Meta and $1,400 on Google — $4,200 total ad spend.
Revenue is 340 × $31 = $10,540. Your blended MER is $10,540 ÷ $4,200 = 2.5x. But Meta's dashboard proudly reports a 4.2x return on its own spend. Which do you trust for the budget call? The blended number, because Meta is grading itself.
Now add profit. Say each shirt costs $9 from your supplier and carries $3 in fees, so your contribution before ads is $31 − $12 = $19. Across 340 orders that is $6,460 — and after $4,200 in ad spend you keep $2,260 before overhead. A model that "improves" to 5x on paper does not change that $2,260 unless real orders or real margin moved.
Incrementality closes the loop. If a lift test shows only seventy percent of Meta's claimed conversions were truly caused by the ads — an incrementality factor of 0.7 (Haus) — then Meta's 4.2x is really a 2.9x incremental return, far closer to your blended reality than its dashboard admits.
Where Victor fits
PodVector AI's Victor is a senior AI employee who connects to Shopify, Meta Ads, Google Ads, Printify, Printful, Gelato, and Klaviyo. He sees every order, every ad dollar, and every supplier cost together, so he computes the true per-order profit on each product — and because product costs from Printify, Printful, and Gelato are known before the first sale, that profit math is right from day one.
Victor acts on what the numbers say. He pauses the losing ad set in Meta, shifts budget and bids in Google Ads, reprices the thin-margin SKUs in Shopify, and builds the Klaviyo flow that re-engages past buyers — then delivers the profit report to your Google Drive and drafts the customer-support email. Victor is not a dashboard you read; he is the operator who does the work attribution is supposed to inform.
Put Victor to work on your store's numbers.
FAQs
What is the single best attribution model for marketing?
For an operating store, data-driven attribution is the best default single model, because it learns from your own conversion paths instead of a fixed rule — but only above roughly three hundred conversions a month, where its splits stabilize (ReferralCandy). Below that volume, or for any channel with offline influence, pair it with a survey and a blended guardrail rather than trusting one model alone.
Why do Meta, Google, Shopify, and GA4 all report different sales?
Because they answer different questions. Shopify uses last non-direct click over thirty days (WeltPixel), GA4 uses data-driven attribution (Google Analytics Help), and Meta and Google each self-credit on their own windows. No setting makes them match — the fix is to assign each tool one job, not to force reconciliation.
Is data-driven attribution objective because it uses machine learning?
It is less arbitrary than rule-based models, but it is trained on observable clicks, so it inherits every blind spot of click tracking — zero-click channels, Safari's cookie clipping, and stripped click IDs (Klaviyo) — and adds opacity you cannot audit. Objective-looking is not the same as complete.
Does attribution tell me what my ads actually caused?
No. Attribution assigns credit among touchpoints it can see; it cannot see the counterfactual — the sale that would have happened anyway. Only holdout-style incrementality tests measure causation (Haus). A channel can win every attribution model and still be non-incremental, which is the classic case with retargeting.
Did iOS 17 kill UTM tracking?
No — it is the opposite. Link Tracking Protection strips user-level click IDs like fbclid and gclid in Apple contexts, but leaves UTM parameters untouched (Klaviyo). That makes disciplined manual UTM tagging more valuable, because UTMs are now your durable attribution layer.
Why is so much of my traffic showing as "Direct"?
Rising direct share is usually tracking decay, not people typing your URL. Stripped click IDs, expired Safari cookies, and untagged links all collapse into direct — and Safari alone is around fifty-five percent of US mobile sessions (Backlinko). Clean UTMs and a post-purchase survey recover most of what direct is hiding.