Answers tied to variant

When a visitor answers, the response is stored with a record of which variant they were in for every experiment they were enrolled in at that moment. Not just the one you happen to be interested in later.

That matters because the question you want to ask is usually decided after the answers have arrived. Storing the whole picture means you can split responses by any concurrently running experiment afterwards, rather than only by one nominated in advance.

It also removes an entire class of error. Reconstructing which variant somebody saw after the fact means joining an answer to an assignment through time, and assignments can change when weights are edited or an experiment is deployed. Recording it at the moment of answering makes the attribution exact.

Impressions carry the same record, which is what lets a per-variant response rate be measured rather than estimated. Being shown a survey and answering it are different events, and both need to be attributable for the rate between them to mean anything.