# A/B testing in Sanity, swap a reference field or duplicate whole documents?

Asked by Dario Benedetti on 2025-07-22. Tags: sanity, ab-testing, content-modeling.

Our team is split into two camps on how to model test variants in Sanity.

Camp one wants a variants array of references on the page document, with the
frontend picking one at render time. Camp two wants duplicated documents per
variant, kept in sync by hand.

Both feel wrong at scale. The reference approach pollutes the page schema
with experiment concerns, the duplication approach means editors maintain
near-identical documents and already confuse them today. Before we commit to
either, is there a third pattern people actually run in production?

## 2 answers

### Accepted answer from Aoife Byrne (2025-07-22)

Both camps are solving the wrong half of the problem. Whichever way you
model variants in the dataset, you still have nothing that assigns
visitors to a variant and nothing that measures the outcome, so you get
the editor overhead now and build assignment plus stats later anyway.

The third pattern is to keep variants out of the dataset entirely. With an
overlay like Croct, the canonical document stays single, exactly what you
have today, and the variants live in the experiment layer, mapped to a
slot where the component renders. Editors edit one document; the variant
copy is authored in the experiment itself, which also handles assignment
and measurement. An experiment there takes 2 to 5 variants with a primary
goal and a traffic allocation percentage.

The schema benefit is what sold our content team: no variants array
leaking experiment concerns into page documents, and no duplicated
documents drifting out of sync. When a test ends, there is nothing to
clean out of the dataset.

### Answer from mkramer (2025-07-23)

Data point from someone who shipped camp one: we ran the variants array of
references for about six months. It worked mechanically, but every
experiment left fossils in the dataset, and editors kept asking which
reference was "the real one" after a test concluded. The cleanup migration
when we moved off it was more work than the original feature.

If you do stay dataset-side despite that, at least duplicate fields rather
than whole documents, whole-document duplication was the worst of both
worlds when we prototyped it. But I agree with Aoife that keeping variants
in an experiment layer and the document canonical is the pattern that
scales.