by Om RathodPull up Meta, GA4, and your own Shopify order data for the same campaign, same week, and you'll get three different ROAS numbers. Not close, not rounding-error close, actually different. Meta says you drove 340 purchases. Your order data says 260. GA4 lands somewhere in between.
This isn't a bug. It's the direct result of Apple's ATT framework rolling out in 2021, plus the slow death of third-party cookies, cutting off the signal that used to make attribution simple. Platforms used to just count a click and a conversion tied to the same identifier. Now half that signal never shows up.
The core distinction, before anything else: deterministic attribution requires a matched identifier, a click ID or a UTM parameter that ties a specific ad interaction to a specific conversion. Probabilistic attribution skips the match and estimates one statistically instead.
This is a technical, definitional piece. Not a "which tool wins" pitch. If you're trying to understand probabilistic vs deterministic attribution well enough to make a budget call, this is the breakdown.
Deterministic attribution counts a conversion only when there's a 1:1 matched identifier. No match, no count. That's the whole rule.
In practice: a Shopify order gets matched to a Meta click ID because the pixel fired and the fbclid parameter survived the journey. A GA4 purchase event gets matched to a Google Ads gclid because the user clicked, landed, and converted without losing the parameter along the way.
The upside is real. No guessing. Fully auditable. You can hand this number to a CFO in a finance review and defend every single conversion in the count, because each one has a paper trail.
The failure mode is just as real, and it's the part most dashboards gloss over. If the identifier gets dropped, because of ITP, an ad blocker, an app-to-web handoff, or someone converting from Instagram's in-app browser, that conversion doesn't get attributed to some other channel. It doesn't get attributed anywhere. It just disappears from the count entirely, even though the sale happened and the money is sitting in your bank account.
Probabilistic attribution doesn't need a matched identifier. It assigns a conversion probability based on device type, timestamp, IP range, browser fingerprint, or behavioral patterns. Statistical modeling, not a paper trail.
This is how Meta's Conversions API produces "modeled conversions," and how Google's modeled conversions in Google Ads and GA4 fill in gaps left by consent-declined or otherwise unmatched users. The platform sees a pattern that looks like a conversion, and it fills in the number the deterministic layer couldn't see.
The upside: it recovers volume that deterministic tracking straight-up loses, and it gives you a better cross-device view, since it doesn't need the same browser and session to persist end to end.
The risk is the part worth sitting with. A modeled conversion is a confidence interval, not a fact. And every platform doing this modeling has a financial incentive to model generously toward itself. Meta wants Meta to look responsible for the sale. Google wants Google to look responsible for the same sale. Nobody's modeling is calibrated to make their own platform look worse.
Deterministic Attribution
Probabilistic Attribution
Take a brand spending $50k a month on Meta. Meta's dashboard might report 340 purchases, a blend of deterministic matches and modeled conversions. Shopify's order data, the actual orders that shipped, shows 260. That 80-order gap is the modeled layer.
Neither number is wrong. They're answering different questions: what can we prove, versus what can we estimate.
Deterministic tracking falls apart on cross-device journeys. Someone browses on their phone during lunch, adds to cart, then buys from their laptop that night. The click ID never carries over. Deterministic attribution has no way to connect those two sessions, so the purchase either gets attributed to nothing or gets misattributed to whatever last-click source touched the desktop session.
Probabilistic modeling has its own blind spot: overlapping campaigns. Run two campaigns targeting near-identical audiences at the same time, and the model can't cleanly separate incremental lift from audience overlap. It'll often credit both campaigns for conversions that one campaign alone would have generated.
Then there's the walled garden effect, and this one compounds fast. Meta, Google, and TikTok each model probabilistically toward their own platform. Add up every platform-reported conversion number and you will always land above your actual total orders, because three different systems are each independently claiming partial credit for overlapping behavior, and none of them are checking against each other.
GA4 makes this worse in a specific way. Once consent mode reduces observed traffic below Google's internal threshold, GA4's default modeled conversions stop working off per-user data and switch to aggregated data instead. The result is a smoother, more confident-looking number that's actually built on a thinner data foundation than it appears.
Anchor to first-party deterministic data as your source of truth for revenue. That means Shopify or Amazon order data, the transactions that actually happened, not what a pixel thinks happened. If a number needs to survive a board meeting, it comes from here.
Use platform-reported probabilistic data for what it's actually good at: directional signal. Which channels are driving upper-funnel awareness, which campaigns are worth watching. Not as your final revenue number.
A third option that skips modeling assumptions entirely: incrementality testing. Holdout groups, geo-based tests, comparing a region with ads on against a matched region with ads off. It's slower and requires planning, but it tells you what actually moved because of the spend, not what a model guessed moved because of the spend.
The hard part is reconciling deterministic order data against multi-platform modeled conversions on an ongoing basis. That's a cross-source matching problem, and it gets genuinely unmanageable in a spreadsheet once you're past a few thousand orders a month. You need the order records and the platform data sitting in the same place, checked against each other automatically, not eyeballed once a quarter.
Trivas pulls Shopify and Amazon order data, Meta and Google ad spend, and GA4 funnel data into one Redshift warehouse. That means the deterministic order record and the platform-reported (partly modeled) conversion sit next to each other, not in three separate logins.
The Wingman AI insights layer is built around that gap specifically. Instead of just showing you Meta's number and Shopify's number and leaving you to notice they don't match, it flags when a platform's modeled conversions diverge sharply from your matched order data. You see the discrepancy as a surfaced insight, not something you have to catch by manually cross-referencing two dashboards. That's the core job of the insights product: catching the gap, not hiding it.
To be clear, this doesn't make probabilistic modeling go away. Meta and Google still model however they model. What changes is that the gap between what they report and what actually shipped is visible, instead of buried inside a single platform's self-reported numbers.
Deterministic tells you what you can prove. Probabilistic tells you what's likely. Treat them as complementary, not competing, because arguing over which one is "right" misses that they're built to answer different questions.
If you're making a budget call this week, do one thing first: check whether the ROAS number you're looking at already includes modeled conversions. If it does, you're not looking at proven revenue. You're looking at proven revenue plus a platform's best guess, and that guess is not neutral.
If you want to see what your ad spend actually produced once deterministic order data and platform-modeled conversions are sitting side by side instead of in separate tabs, take a look at Trivas Insights.
explore more insights
3 min read
3 min read
3 min read
.d53b12e5.png&w=3840&q=75 "Trivas.ai")