Your ad dashboard shows 5x ROAS. Your backend shows 2x. That gap isn’t usually fraud or a math error — it’s two measurement systems answering different questions. And that mismatch steers real budget choices, fuels automated bidding, and quietly kills the upper‑funnel channels that seed future sales.

Two systematic bias directions, one shared cause

Ad platforms (Google, Meta, Microsoft, TikTok) count conversions with defaults that tend to inflate their own performance: view‑through credits for impressions, modeled conversions when tracking is partial, and long conversion windows that assign revenue to clicks from weeks earlier. Those choices push platform‑reported ROAS upward.

Most backends and CRMs do the opposite: they use last‑click logic or a close variant. The last observed touch — often branded search or a direct visit — receives the sale. The ad that initiated the customer’s journey can be left uncredited. That pulls backend ROAS downward.

Both systems force a single touch to own the outcome. Platforms resolve ambiguity in their favor; backends resolve it toward the most recent click. Averaging the two doesn’t find the truth — it creates a number that belongs to neither instrument and conceals causal impact.

Which channels get hurt — and why search looks less broken

Channels that work by influence rather than clicks — social, display, video, connected TV — are the ones last‑click measurement sees least. An impression that creates awareness, prompts later branded search and then a purchase is invisible to a CRM that only records the final click. To the backend, an impression‑led campaign can look like it did nothing.

Search is different because it usually requires a click and that click often happens near the purchase. Last‑click systems therefore capture more of search’s contribution. They still undercredit upstream research queries and general awareness work, but the gap is smaller. That structural difference explains why search often appears “less broken” in backend reconciliations.

Why reconciliation and fancy attribution aren’t the final answer

Common fixes — reconciling dashboards, applying platform data‑driven attribution (DDA) or moving to multi‑touch models — still answer one question: which touch gets credit? They do not answer the causal question: would the sale have happened without the ad?

DDA and reconciled views are built from the same inputs that produced inconsistent numbers. They can redistribute credit plausibly but cannot prove causation. Worse, automated bidding that optimizes to a reconciled ROAS amplifies whatever bias exists: a sophisticated optimizer compounds a biased signal faster, at scale.

Regional consent regimes and tracking changes (for example, the EU’s consent mode and post‑iOS‑14 privacy constraints) increase the share of modeled versus observed conversions in platform reports. That widens the numeric gap with last‑click backends rather than narrowing it.

What actually measures impact: incrementality and how to run it

Incrementality — the additional revenue caused by running the ad — is the only reliable answer to “would this have happened anyway?” The direct route to incrementality is experimental or quasi‑experimental testing, not more attribution algebra.

Practical tests marketers can run:

  • Geo holdouts. Hold the channel out in comparable regions and run tests across a full purchase cycle. Four weeks is a practical minimum; longer is better for longer purchase windows.
  • Measure backend revenue deltas between live and held‑out regions. Don’t rely on platform‑reported conversions for this — the platforms can’t attest to their own absence.
  • If multiple platforms run concurrently, hold each platform out in its own window rather than trying to apportion one sale across several platforms that each claimed the whole of it.
  • Use conversion‑lift tests or randomized controlled trials when geo segmentation isn’t feasible; expect higher complexity and larger sample requirements.

Use these causal checks together with top‑down tools. Marketing‑mix modeling (MMM) gives an aggregate, cross‑channel view using spend and outcomes; holdouts provide channel‑level causal checks. Attribution tools remain useful for operational decisions — creative testing and pacing — but treat them as diagnostic, not definitive.

Build or buy attribution — and for whom each makes sense

Building an in‑house attribution stack demands continuous engineering work: deduplicating platform conversions, resolving identity across devices, handling modeled conversions, and updating pipelines whenever platforms change APIs or windows. It’s an ongoing cost, not a one‑time project.

Packaged multi‑touch vendors can speed deployment and bundle MMM and incrementality features. Their value generally appears at scale: vendors pay off when an account handles seven‑to‑eight‑figure revenue and has a dedicated analyst to interpret outputs. Smaller accounts often get clearer insight from a single clean holdout at far lower cost than a half‑configured enterprise dashboard.

Whatever you choose, remember this distinction: attribution clarifies which touch receives credit; incrementality proves whether the ad created incremental value.

Practical next step: before you reallocate meaningful spend or hand decisions to automated bidding, run a controlled holdout for the channels you contest most. Use holdouts to validate models and MMM to inform cross‑channel strategy. Treat platform ROAS as a directional signal, not a final verdict.

What to watch next: consent frameworks, modeling techniques and platform defaults will continue to change the numbers you see. Those shifts will alter reported gaps, but they won’t remove the core requirement — causal tests remain the only reliable foundation for major budget decisions.