In 2025, a growth team optimized budget off a three-stage funnel — open, register, pay — and the ROAS got worse the more they spent. On audit, the funnel was lying at both ends, and the team had been making half a year of decisions off it. This pattern is especially common inside super-app ecosystems, because a mini-program's runtime lives on someone else's floor, and analytics tools default to treating "what happened on someone else's floor" as "what we measured accurately."
Preloading makes "open" the first lie
To feel instant, hosts pre-create the WebView and sometimes prefetch the first route, so the tap seems immediate. The developer's onLaunch fires during that preload. Counted as a session, the top of the funnel is inflated by every user who skimmed the entry and never intended to enter. The honest first step is not "launched" but "seen" — fire the entry event when the first frame actually renders, not during runtime warm-up.
// wrong: counts preloads users never saw
runtime.onLaunch(() => track('session_start'));
// right: counts a moment someone actually saw
runtime.onFirstPaint(() => track('session_seen'));
Do not underestimate that one-word change. If an entry is preloaded a hundred thousand times a day and only twenty thousand actually enter, the funnel top is watered fivefold, and every downstream conversion rate becomes unexplainable noise.
The attribution layer leaked by design in the privacy era
The second corruption is in attribution. AppsFlyer found that under SKAdNetwork, 32% of non-organic installs were misattributed as organic (games "best" at 21.8%), and their fraud data shows roughly 15% of installs are outright fraudulent; one social iOS channel hit a 275% real-users lift in a quarter — three of every four installs fake. SKAdNetwork itself limits precision: it returns no per-install channel, only an aggregated conversion value, with postback delay and crowd-anonymity thresholds. It was built for "a rough direction under privacy," not "exact per-install accounting." Using it as an exact ledger is the second systematic misjudgment.
Cross-host identity breaks into two rows
A user who finds the mini-program on WeChat and later arrives via Telegram is two people in the analyst's table — one openid, one tg user id, with no key to join them. When one person splits into two rows, real cross-host LTV cannot be computed. Worse, the install that brought them was itself suspect: that 32% misattribution means nearly a third of "paid" installs were actually organic, so CAC and ROAS both err in the expensive direction. A team watching only single-host attribution keeps pouring budget into the channel that looks cheap but is actually expensive.
| Metric | Client sees | Actual truth |
|---|---|---|
| Sessions | Includes preloads | Inflated several times |
| Paid installs | Attribution credit | ~32% are organic misattrib. |
| CAC / ROAS | Built on above | Inflated by fraud plus misattrib. |
Put the conversion where the host cannot fake it
The only number that holds is the one written at the moment value actually moved, on the server. A client purchase event can be forged, sampled, or lost in preload; a server record of a settled payment cannot. Record the funnel top honestly (first frame, not launch) and anchor the bottom to the server. Everything in between — channel attribution, session stitching — is a model, and should be labeled a model, not treated as fact. Teams that finally stopped trusting the client funnel and moved conversion to the backend were the first to see which host actually pays for them.
Privacy law structurally ended client attribution
Not just Apple. Under GDPR and similar rules, the stable client identifier depends on a consent gate the user can refuse, and decays even when granted. Cross-host identity was already broken; regulation makes client attribution structurally unreliable. One team that moved conversion events to the server discovered one host's real CAC was 40% lower than the dashboard showed, and reallocated budget — half of that "expensiveness" had been a phantom manufactured by preload inflation and misattribution.
The bottom line
The device was never the source of truth; the server is. Record a moment someone saw at the top, anchor a settled record at the bottom, and label the middle as a model — that is closer to reality than any fancy client instrumentation. Teams still deciding off a host-polluted funnel are optimizing a guess dressed as a metric.