Marketers often see inflated ROAS figures from ad platforms, while backend data tells a different story. Attribution errors and platform self-interest distort results, making true campaign impact hard to measure. Incrementality testing offers a clearer answer.
Every digital marketer has faced the confusion: Google Ads, Meta Ads, or TikTok Ads report a 5x ROAS, but your backend system shows only 2x. It’s tempting to trust the backend and dismiss the campaign as underwhelming. Yet, neither number tells the full story. Ad platforms are designed to count generously, including view-throughs, modeled conversions, and long conversion windows that inflate results. Meanwhile, backend systems often rely on last-click attribution, crediting only the final touchpoint-usually a branded search or direct visit-while ignoring the initial paid click that started the journey weeks earlier.
This mismatch isn’t about fraud. It’s about how each system resolves ambiguity in its own favor. Marketing platforms claim every possible assist, while backend systems strip credit from anything but the last click. The gap between 5x and 2x is mostly the result of these opposing biases, not a sign of deception. Trying to reconcile the two often produces a third, equally unreliable number-like averaging two miscalibrated thermometers and expecting an accurate temperature.
Invisible Impact
The backend’s blind spot is especially severe for impression-based channels like social, display, video, and connected TV. These formats influence buyers without generating clicks, so their impact is invisible to last-click systems. For example, a Meta campaign might drive a surge in branded search, but the backend credits only the final search or direct visit, not the ad that sparked interest days earlier. Since iOS 14, even social clicks are harder to match to purchases, further weakening backend visibility into these channels.
Search campaigns fare better because clicks are closer to the purchase, but even here, upper-funnel activity-like generic queries that lead to branded searches-gets under-credited. The real casualty of last-click logic is social and display, not search. Judging impression-based campaigns solely by backend revenue risks cutting off the demand generation that fuels downstream conversions.
Double Counting Across Platforms
The problem compounds when running multiple platforms. If you add up reported conversion revenue from Google and Meta for the same period, the total often exceeds actual sales. That’s because both platforms may claim the same sale: a customer sees a Meta ad, later clicks a Google ad, and both platforms log the conversion. Neither discounts for the other’s influence, so you end up with inflated totals. Adding more channels only worsens the overcounting, making reconciliation impossible-each platform has already claimed the full value of the sale.
This dynamic is not unique to social or search. Meta’s view-through conversions are a major source of over-reporting, but Google also uses engaged-view and modeled conversions to boost its numbers. The result is a reporting environment where no single figure reflects reality.
Why Reconciliation Fails
Many marketers try to fix the problem by reconciling data sources or adopting data-driven attribution (DDA) models. While DDA reallocates credit among paid touches, it still relies on the same flawed platform data and never answers the key question: would the conversion have happened without the ad? Automated bidding systems that optimize toward these reconciled numbers only amplify the problem, pouring budget into campaigns based on unreliable inputs. In regions like the EU, consent mode and modeled conversions widen the gap between platform and backend numbers even further.
Building a custom attribution system or buying a multi-touch attribution suite like Triple Whale or Northbeam can add sophistication, but these solutions are only justified for large accounts with dedicated analysts. For most, the complexity outweighs the benefit, and the core issue remains: attribution models answer "which touch gets credit," not "did the ad make a difference?"
Incrementality: The Real Test
The only way to answer whether ads truly drive incremental revenue is to run holdout tests-turning off a channel in a specific region or segment and measuring the change in backend revenue. This approach, known as incrementality testing, reveals the actual lift caused by advertising. Marketers should hold out channels for a full purchase cycle, compare backend revenue between test and control groups, and use these results to validate or challenge attribution models. When multiple platforms are involved, hold out each one separately rather than trying to split credit after the fact.
For broader budget decisions, marketing mix modeling can provide a top-down view of channel contributions, but no single method is definitive. Attribution, incrementality, and mix modeling each offer a piece of the puzzle and should be used together to check one another.
Instead of debating which number to trust, marketers should focus on what happens to backend revenue when a campaign goes dark. That’s the only test that truly measures impact. For more on how AI-driven search is changing conversion strategies and measurement, see this analysis of how LLM referral traffic is reshaping the conversion funnel.
