Hey 👋 We're hitting an SRM on a URL Redirect experiment and want a sanity check on whether our

trackingCallback

is wrong or we're hitting a known limitation. Setup: Astro static site, GA4 loaded via GTM (deferred),

@growthbook/growthbook

client-side, bucketing on

deviceId

. 3 variations (control + 2 redirects). Prod (N=2048, p<0.001): Control 47.66% / Var1 25.34% / Var2 27%. Variations losing ~22% of expected exposures. Already ruled out: local simulation (~2000 fresh sessions, capturing every GA4 request) shows balanced bucketing, exactly 1 event per session, no double-fire, no destination re-evaluation. SDK looks clean in isolation. Our hypothesis: because GA4 is GTM-deferred,

gtag.js

sometimes isn't ready before GrowthBook fires the redirect (within 2s

maxNavigateDelay

). The queued

experiment_viewed

sits in

dataLayer

, page unloads, event lost. Control has no deadline, so it always delivers. Implementation (per docs' async-callback +

transport_type: 'beacon'

pattern):

trackingCallback: (experiment, result) => {
  const attrs = gb.getAttributes() as Record<string, string>
  const isRedirectVariation = !!(result.value as { urlRedirect?: string } | undefined)?.urlRedirect

  if (isRedirectVariation) {
    return new Promise<void>((resolve) => {
      gtag('event', 'experiment_viewed', {
        transport_type: 'beacon',
        event_category: 'experiment',
        experiment_id: experiment.key,
        variation_id: result.variationId,
        gb_user_id: attrs.userId,
        gb_device_id: attrs.deviceId,
        event_callback: resolve,
      })
    })
  }

  return new Promise<void>((resolve) => {
    gtag('event', 'experiment_viewed', {
      event_category: 'experiment',
      experiment_id: experiment.key,
      variation_id: result.variationId,
      gb_user_id: attrs.userId,
      gb_device_id: attrs.deviceId,
      event_callback: resolve,
    })
  })
},

Question: is something wrong here, or is

transport_type: 'beacon'

known to be unreliable with GTM-deferred GA4? If the latter — what do you recommend for client-side redirect experiments when Edge Worker SDK isn't an option? Thanks!

Hi Dalius.

thanks thankyou 🙏