A sample ratio mismatch is one of the clearest warning signs that an experiment is broken.
Random assignment produces groups very close to the planned split, so a large gap means something non-random happened.
The check is cheap, and it should run on every test before anyone looks at the results.
The expected counts come from the total and the planned split, such as half each for a 50/50 test.
A chi-square test of observed against expected counts gives a p-value for the gap.
Very small p-values, such as below 0.001, are a common trigger for investigation.
A frequent cause is the new version failing to load for some users, so they are never counted.
Bots and other automated traffic can also hit one version more than the other.
Redirects, stored copies of pages and changes to logging midway through a test are other common causes.
The missing users are rarely random, which is why results from a mismatched test can be badly biased.
For example, if the treatment crashes on old phones, the remaining treatment users are mostly on newer phones and spend more.
Re-weighting the groups to the planned split does not fix this, because the missing users differ from those who stayed.
The fix is to find the cause, repair it, and usually run the test again.
In practice, many experimentation platforms flag a mismatch automatically on the results page.
The usual mistake is treating the mismatch as a curiosity and launching the winner anyway.