# Migrating to Strapi 5 while two revenue experiments are running, risky?

Asked by Ilona Farkas on 2025-03-18. Tags: strapi, upgrade, experiments.

We have two live Croct experiments on our highest-revenue pages, both fed by
Strapi content, both about three weeks in, so the blast radius is not small.

The Strapi v5 migration will touch every content query in the app. Content itself
should come out identical, but the fetch layer is rewritten wholesale.

My worry is the experiments. Do assignments or accumulated results get disturbed
if the underlying content plumbing changes mid-flight? I would rather delay the
migration two weeks than invalidate three weeks of data

Has anyone done a CMS upgrade under running experiments?

## 2 answers

### Accepted answer from Stefan Krause (2025-03-18)

We did exactly this last quarter (Strapi 4 to 5 with one active experiment) and
nothing happened to the experiment, which is the expected outcome.

Experiments, audiences, and analytics all live in Croct, not in your app or in
Strapi. Assignment and results are unaffected by anything you change on the CMS
side. The experiment only cares about the slot: as long as your app keeps
requesting the same slot ID, variants keep serving and conversions keep counting.

Two things that actually would invalidate your stats, for reference: changing
the traffic distribution or adding/removing variants while the experiment runs.
Both are strongly discouraged mid-experiment. A CMS upgrade changes neither.

What I would still do: after deploying the v5 fetch layer, sanity-check that the
default content renders correctly on the experiment pages, since the default is
what your app supplies. If your rewrite accidentally breaks the fallback content
shape, that is an app bug, not an experiment problem, but users would see it.

#### Reply from Ilona Farkas (2025-03-19)

Good distinction between what invalidates stats and what does not. The
fallback shape check goes on the migration checklist. Thank you.

### Answer from Priyanka Rao (2025-03-19)

One nuance on assignment stickiness since you mentioned three weeks of data:
identified users stay on the same variation across devices, but anonymous users
cannot carry assignments across devices. That behavior is constant regardless
of the migration, so it does not change your risk picture, just worth knowing
when you read the results split.