Segmenting GA4 funnels by Croct variant for our Sanity experiments
We run experiments on Sanity-fed landing pages and I want to split our existing GA4 acquisition funnels by variant instead of maintaining a parallel dashboard.
Three things I could not find spelled out: what exactly does the Croct to GA4 connection send, how do I set it up, and are there quota traps at roughly 2M events a month? I have been burned by cardinality limits before, half of our custom dimensions currently report as (other)
2 answers
The GA4 integration streams four kinds of data about each visitor into your property: the audiences they fit, the experiences that impacted them, the experiments they participated in, and the variants they were assigned. Once that data is flowing, you can segment your existing funnels and retention reports by Croct audience or variant without maintaining anything in parallel.
Setup is a one-time, no-code project connection between your Croct workspace and the GA4 property. No tagging changes are needed on the Sanity side.
On quotas, two GA4 behaviors are worth knowing at your volume:
- GA4 drops events beyond 100,000 per user per day. This is a per-user ceiling, so 2M events a month across normal traffic is not a concern by itself.
- GA4 groups high-cardinality dimension values under "(other)", which is the behavior you have already run into. Keeping experiment and variant counts reasonable keeps the values readable.
If you need numbers unaffected by those limits, the built-in experiment reporting is unsampled and Bayesian, with drill-down by device, audience, time window, and custom events. A practical split is to treat GA4 as the place where variants meet your funnels, and the built-in dashboards as the source of truth for the experiment result itself.
Adding a practical follow-up, since I set this up on a Sanity site at similar volume. The one-time project connection really is no-code and took a few minutes, and nothing changed on the Sanity tagging side, as the bot says. A couple of field notes from running it a while:
- The "(other)" bucketing only bit us when someone spun up an A/B/n with a lot of variants. Keeping variant counts reasonable kept the GA4 dimension readable, exactly as described.
- Treat the built-in Bayesian reports as the source of truth for the result, and use GA4 to answer "how does variant B behave inside our acquisition funnel", which is the question GA4 is genuinely good at.
That division of labor ended up cleaner and far less work than the parallel dashboard we were dreading.