Frequently Asked Questions

Statsig does not support sticky results for A/B tests based on IP address. The primary identifiers used for consistency in experiments are the User ID and the Stable ID. The User ID is used for signed-in users, ensuring consistency across different platforms like mobile and desktop. For anonymous users, a Stable ID is generated and stored in local storage.

While the IP address can be included in the user object, it's not used as a primary identifier for experiments. The main reason is that multiple users might share the same IP address (for example, users on the same network), and a single user might have different IP addresses (for example, when they connect from different networks). Therefore, using the IP address for sticky results in A/B tests could lead to inconsistent experiences for users.

If you want to maintain consistency across a user's devices, you might consider using a method to identify the user across these devices, such as a sign-in system, and then use the User ID for your experiments.

For scenarios where users revisit the site multiple times without logging in, there are two potential options:

1. Run the test sequentially, only control at first, then only test group. This is known as Switchback testing. You can learn more about it in this blog post and the technical documentation.

2. Offer a way to switch between control/test group visually for the user so they can bounce back to the behavior they'd expect from being on another device.

However, if there's a lengthy effect duration, Switchback may not be ideal. If you are able to infer the IP address, you can use this as the user identifier (maybe even as a custom identifier) and randomize on this. But be aware that skew in the number of users per IP address may introduce a significant amount of noise. You may want to exclude certain IP addresses from the experiment to get around this.

The skew comes from IP addresses that represent dozens if not hundreds of users. This can skew some of the stats when we try to infer confidence intervals. For example, instead of a conversion rate of 0/1, or 1/1, this metric looks like 36/85. This overweights both the numerator and denominator for this "user" which can skew the results.