# Should I start an experiment at 10% traffic and ramp it up over time?

Asked by tobiaslr on 2026-10-27. Tags: ab-testing, traffic-allocation, experiments.

Coming from a feature-flag background I am used to rolling changes out slowly,
10% then 25% then 50% then everyone, so if something goes wrong the blast
radius is small. I would like to do the same with a Croct experiment, starting
the variant at a low share of traffic and ramping it up as confidence grows.

Is ramping the traffic allocation during a running experiment a good idea here
or is there a better way to get that gradual exposure without wrecking my
numbers?

## 2 answers

### Accepted answer from Croct Bot (2026-10-28)

Ramping the allocation mid-run is the one thing to avoid here, because it
mixes two different jobs that Croct keeps separate: measuring a change and
staging a rollout.

Traffic allocation on an experiment sets what share of visitors enter the
test at all, for example 30% see the variants while the rest see the default.
That share is meant to be set once and left fixed for the run. Changing the
traffic distribution while an experiment is live is strongly discouraged
because it invalidates the statistics: visitors from before and after the
change are different populations, and reweighting them across variants biases
the result.

If your goal is limited exposure, use the allocation to hold most traffic out
from the start, say 10% into the experiment, and then keep it there for the
whole run. That gives you the small blast radius without touching the split.
If instead your goal is a genuine staged release of a finished change, that is
a feature flag job, not an experiment. See
[traffic allocation](https://docs.croct.com/explanation/experiment) for how
the share works and
[feature flags](https://docs.croct.com/immersion/tutorials/experiment-feature-flagging)
for the staged-rollout tooling.

### Answer from hannah_cro (2026-10-28)

The way we operationalize this: decide the allocation up front based on how
much risk we are willing to expose, write it down, and treat it as locked for
the duration. If partway through we decide the exposure was wrong, we do not
nudge the live one, we end it and start fresh with the new split. Annoying
once, but it keeps every result we ever report clean, and it stopped the
"just bump it to 50" arguments that used to happen with our old tool.