The distance between what platforms report and what actually happened has a name now. It is called the double count, and it has been hiding inside your ROAS dashboard for years.
On March 3, 2026, Meta published an announcement titled "Simplifying Ad Measurement for a Social-First World" and changed the way it counts conversions. Hitherto, click-through attribution had included not only actual link clicks but also likes, shares, saves, comments, profile taps, and any non-link interaction with an ad. These social gestures — the machinery of passive engagement — had been counted as if they were the same as someone visiting your product page and adding something to a cart. They were not the same. Meta knew this. The advertisers bidding on Meta's auction knew this. The discrepancy between Meta's reported conversions and Google Analytics's last-click data had reached 46 percent, according to Meta's own published figures. Nobody had fixed it.
Now Meta has fixed it — in a particular direction.
The change works like this. Click-through attribution now means only link clicks: someone actually tapping through to a website. Everything else — likes, shares, saves, comments, video views at five seconds or longer — moves into a new category called engage-through attribution, with an attribution window of one day. Previously, these signals lived in the seven-day click-through window. The engaged-view video threshold also dropped from ten seconds to five seconds, capturing faster interactions on Reels.
The effect on reported numbers is immediate and significant. A conversion that previously counted because someone saved an ad on Monday and purchased on Wednesday now counts for nothing — it is neither a click-through (no link click) nor an engage-through (the one-day window expired). For advertisers running retargeting campaigns with longer consideration cycles, reported conversion counts drop by 25 to 50 percent. The actual behavior of consumers has not changed. The ruler has.
Meta's stated rationale is to close the measurement gap with third-party tools like Google Analytics, which have always counted only link clicks. This is accurate. It is also incomplete. The reclassification also reveals — retroactively — how much of Meta's historical conversion volume was phantom: conversions that other channels caused, but which Meta captured through generous view-through and social-engagement windows. The 46 percent gap Meta cited was not a bug in the measurement. It was the feature. It was the number that made Meta campaigns look effective enough to justify continued spend.
Here is the structure of phantom performance as it existed before March 2026. A consumer encounters a brand for the first time through a podcast mention, a word-of-mouth recommendation, or a Google search. She does not click. She is not counted anywhere. She encounters the brand again through a Meta retargeting campaign. She engages — a like, a save, a comment. Seven days later, she purchases. Meta recorded the conversion. Google Analytics recorded nothing. The advertiser's dashboard showed Meta performance. The reality was that the podcast created the purchase, and Meta's engagement window simply arrived last and claimed it.
This is not an edge case. This is the documented norm. Celeri's 2025 analysis of ecommerce advertisers found that platform-reported ROAS exceeds true ROAS by an average of 38 percent across all channels. For high-COGS businesses the figure reaches 52 percent; for fashion, 45 percent. In one documented case, a platform dashboard showed 5.0x ROAS while incrementality testing revealed 0.95x — a 426 percent overstatement. The gap is not noise. It is structural. Every platform has economic incentives to report favorable numbers, and every platform has designed attribution windows and conversion definitions that serve those incentives.
Cassandra's analysis of 194 advertisers across 792 marketing mix models found that 35 percent of budgets — in one case, EUR 5.9 million of a EUR 16.7 million budget — were allocated to channels showing positive ROAS in attribution dashboards but zero measured incrementality. Every dashboard was green. The channels were not working. The home improvement retailer discovered that search claimed 70.2 percent of conversions in attribution data, while MMM found only 3.9 percent truly incremental. Video told the opposite story: 0.6 percent in attribution, 6.8 percent in incrementality. When the retailer reallocated 25 percent of budget from search to video, conversions increased 18 percent with the same total spend.
The mechanism is last-touch attribution's systematic over-crediting of the channel nearest the conversion. A channel that only fired because someone was already going to convert — retargeting someone who had already decided — claims credit for demand that existed before the retargeting ever appeared. The channels that actually moved someone from unaware to ready-to-buy — upper-funnel, slow-burn, difficult to track — get no credit, because they do not appear at the bottom of the funnel.
Meta's March 2026 reclassification did not improve measurement. It changed the scoreboard while the industry kept benchmarking against it as if the rules had not changed.
The problem is not merely that Meta uses different rules than Google Analytics. It is that every platform now uses different rules, and the industry has not developed the infrastructure to normalize across them.
Google Ads defaults to last-click within its ecosystem — a 30-day click window. TikTok counts seven-day click-through plus one-day view-through. Amazon's Sponsored Brands and DSP moved in January 2026 to a shopping-signal enhanced attribution model that measures brand discovery moments rather than last-touch. Meta's engage-through, post-March, counts one-day only for engagement signals. These are not comparable denominators. They are different units being used to measure the same phenomenon — whether advertising produces revenue — without conversion.
An advertiser comparing Meta's ROAS to Google's ROAS is not comparing Meta to Google. They are comparing Meta's ROAS under Meta's rules to Google's ROAS under Google's rules, in a world where neither platform discloses its methodology with enough specificity to understand what the numbers actually represent. Search Engine Land's March 3 coverage of the Meta reclassification noted that Meta named Northbeam and Triple Whale as integration partners specifically to incorporate both click and engagement signals — which is another way of saying that Meta itself acknowledges that neither signal alone is a reliable picture.
Cassandra's data on cross-platform overstatement is stark: Meta reports a median multiple 2.34 times higher than incrementality-tested truth. Google reports 1.18 times. TikTok, which a 2025 Measurable analysis found to be the worst offender for attribution inflation through aggressive view-through crediting on fast-scrolling immersive content, is somewhere between. A direct comparison of any two of these platforms requires a shared ruler. None exists.
The industry has recognized the problem. IAB and MRC published the Attention Measurement Guidelines in November 2025, establishing a framework for measuring attention across media and formats. IAB Tech Lab released the ECAPI specification in January 2026, standardizing advertiser-to-platform event communication. The U.S. Joint Industry Committee launched a Streaming Data Service in late 2024 enabling privacy-first cross-platform measurement. These are real efforts. They have not solved the problem.
The guidelines address attention as a quality signal. The ECAPI specification addresses event-level data standardization. The JIC's SDS addresses deduplication in streaming TV. None of them provide a unified, transactable currency for performance advertising — a single unit that an advertiser can use to compare the return on a Meta campaign against a Google campaign against a TikTok campaign on the same baseline. That infrastructure does not exist. The industry has spent five years building around the problem rather than through it.
Advertisers who want true cross-platform visibility must run their own marketing mix modeling with incrementality testing, which is expensive, slow, and operationally demanding. Most do not. Most rely on platform-reported metrics because that is what the dashboards show, because the platforms have built the dashboards, because the platforms benefit from the dashboards showing favorable numbers. The incentive structure produces exactly the measurement system the incentive structure rewards.
Ron Berman's 2018 Wharton research established that last-touch attribution systematically lowers advertiser profits because it misallocates credit. An et al. demonstrated at NeurIPS 2025 that the last-click mechanism is not incentive-compatible and can perform arbitrarily poorly, with strategic timestamp manipulation feasible and undetectable. Verma (2025) found that attribution model selection alters perceived channel effectiveness by up to 143 percent. The academic literature has been describing this problem for years. The marketing industry has continued to run on platform-reported last-click because it is legible, it is cheap, and it is what the vendors sell.
Meta's March 2026 reclassification exposed the crossing — the moment when a platform changes its own measurement rules and the industry has no way to know whether what follows is better measurement or simply different counting that happens to flatter the platform's historical record.
There is a version of this story in which Meta is being honest: the prior definitions were misleading, the new definitions are more accurate, and advertisers now have clearer visibility into two distinct conversion paths. That version is available. It is not the version the data supports. The data supports a version in which the prior definitions generated campaign performance numbers that justified continued spend, the gap between those numbers and incrementality-tested truth had become large enough to be embarrassing, and the reclassification both narrows the gap and transfers the overcounted conversions into a window that will count fewer of them going forward — while the historical record now shows lower performance retroactively, which is the platform's problem, not the advertiser's.
The advertiser comparing Meta ROAS to Google ROAS today is in the same position as the investor comparing two funds that use different accounting standards, different risk adjustments, and different definition of what constitutes a return. They are not comparing the funds. They are comparing the marketing materials.
The double count has not been eliminated. It has been reclassified. The phantom conversions did not disappear. They moved from one column to another, and the advertisers whose budgets depend on those numbers are still looking at the wrong column, in a dashboard built by the vendor with the strongest economic interest in what the dashboard shows.