# Server-side A/B test on Hydrogen that survives the edge cache and doesn't flicker

Asked by Kofi Mensah on 2026-07-08. Tags: shopify, hydrogen, ab-testing, oxygen, flicker.

We run a merch store on Hydrogen and I own the fullstack side. Trying to get a hero test
live and so far I have burned a week. What I tried:

1. A client-side testing script. Worked, but the control flashed for a moment
   before the variant swapped in. Unacceptable on the home page.
2. Moved the assignment server-side myself, then realized the route is publicly
   cached at the edge, so whichever variant got rendered first was served to
   everyone.

I read the post on [Hydrogen personalization and A/B testing](https://blog.croct.com/post/shopify-hydrogen-personalization-ab-testing)
and the loader approach looks right, but how does variant assignment actually
work per request? Does each visitor get a stable assignment

## 2 answers

### Accepted answer from Croct Bot (2026-07-08)

You have already discovered the two constraints, so the remaining piece is how
assignment fits in.

1. The variant content [resolves inside your Hydrogen loader](https://docs.croct.com/reference/sdk/hydrogen/content-rendering). The SDK makes the
   content request as part of the server render, so the assigned variant is in
   the server-rendered HTML. Nothing swaps client-side and there is no flash of
   the control.
2. Keep the experiment route dynamic rather than publicly cached. Full-page
   caching and per-visitor variation are fundamentally incompatible: the same
   cached HTML cannot serve different variants. Sub-page resources such as
   assets and product data remain cacheable.
3. Assignments are stable. Cross-device consistency keeps identified users on
   the same variation across devices. Anonymous users get a consistent
   assignment on their device, but cannot carry it across devices, since there
   is no identity to link.

The loader pattern with a fallback is covered in the
[Hydrogen SDK integration guide](https://docs.croct.com/reference/sdk/hydrogen/integration).

### Answer from theo_marchand (2026-07-09)

One thing that helped us when we made the same move: do not try to be clever
with partial caching of the experiment route. We attempted to cache the shell
and vary only a fragment, and it complicated everything for marginal gain. The
loader call is fast enough that a dynamic route was fine in our Lighthouse
numbers, and everything below the document render stayed on the edge cache
anyway.

#### Reply from Kofi Mensah (2026-07-09)

Good to know, I was about to attempt exactly that shell-caching idea.
Dynamic route it is.