Ad attribution has been structurally broken since Apple introduced App Tracking Transparency (ATT) in iOS 14.5 in 2021. Google's gradual deprecation of third-party cookies finished the job by 2025. The cross-platform attribution stack that growth teams used in 2018-2020 — deterministic user-level matching across networks — no longer exists for most users. By 2026, the operating model has shifted to modeled attribution plus periodic incrementality testing plus media mix modeling for the most spend-heavy teams.
This guide is the operator's view of cross-platform attribution in 2026 — what each platform is actually measuring, where the gaps are, and what to do when Meta's reported ROAS disagrees with what your finance team sees in the bank.
What each platform is actually measuring
Each platform now uses a hybrid of deterministic matching (when consent allows) and modeled estimation (when it doesn't). Meta's Aggregated Event Measurement (AEM) and Conversions API (CAPI) replace user-level tracking with cohort-level modeling. Google's Enhanced Conversions + consent mode does the same. TikTok's modeled conversions use Smart+ AI to estimate post-click conversion likelihood.
The structural consequence: per-platform ROAS is now closer to a directional indicator than a precise measurement. Each platform's number is right within its own methodology but systematically biased toward over-counting (since each one claims credit for conversions other platforms also touched).
The blended-ROAS workaround
The accounting-grade fix is blended ROAS: total revenue from all sources divided by total ad spend across all platforms. The blended number can't be inflated by attribution overlap, since it sums both sides. Blended ROAS is what your CFO can verify against the P&L.
Tools that compute blended ROAS automatically (via each platform's Marketing API for spend + the shop's analytics for revenue) shortcut the manual reconciliation that used to be a Monday-morning spreadsheet task. The trade-off is that blended ROAS gives you the portfolio view, not the channel-level diagnosis — for that you still need per-platform numbers, knowing they're directional.
Incrementality testing as the truth source
The accounting-grade measurement of channel impact is incrementality testing — geo-holdouts, audience-holdouts, or ghost-bids experiments that compare 'spend on this channel' against 'no spend on this channel' for matched cohorts. The numbers that come back almost always tell you that platforms over-claim attribution (typical adjustment: Meta's reported ROAS is real but 20-40% of the credited conversions would have happened anyway).
Incrementality testing is operationally heavy — you have to design the experiment, hold out spend, wait 4-6 weeks for statistical significance, and re-run periodically as the channel mix changes. Most teams run incrementality tests quarterly on their top 2-3 channels and use the results to calibrate the attribution they trust day-to-day.
Media mix modeling (MMM) for the heavyweights
Teams spending $500k+/month in blended ad spend increasingly add media mix modeling on top of attribution — typically through tools like Northbeam, Triple Whale, or Recast (or in-house if engineering capacity exists). MMM uses statistical regression on spend + revenue time-series data to estimate each channel's contribution, with the advantage that it doesn't depend on user-level tracking at all.
MMM is mathematically heavy but the practical takeaway is straightforward: when MMM, incrementality tests, and platform-reported attribution disagree (and they will), trust the incrementality test first, MMM second, and platform-reported numbers last.
What this looks like in Gapscout
Gapscout's reporting surfaces both per-platform reported attribution (using each network's native methodology) and blended ROAS computed across all connected networks. The audit agents flag campaigns whose ROAS regression is platform-internal (driven by attribution methodology change) versus account-internal (driven by actual performance change) — a distinction that matters for whether to act on the signal.
For incrementality testing and MMM, Gapscout integrates with the dedicated tools rather than rebuilding them — Northbeam and Triple Whale connections export Gapscout's spend data into their attribution models, and the unified reporting view surfaces both Gapscout's blended ROAS and the MMM-adjusted version side-by-side.