Because almost everything around the model changes even when the model itself does not. New documents, an edited instruction, a new tool, your provider's own update; any of these changes the answers you get. Nobody sends you a notice, so the change usually arrives as a customer complaint.
People assume the model was swapped. Usually it was not. These are the common causes, roughly in order of how often we see them.
A failure is loud. Drift is not. The answers stay fluent and confident, the format looks the same, and only the substance moves. So the first signal is usually a person, not a dashboard: a customer who got told the wrong thing, or a colleague who says it used to handle that better.
That is the argument for checking on a schedule rather than when you get suspicious. A weekly run against the same questions makes a change visible as a change, rather than as an anecdote you cannot quite pin down.
Keep a fixed set of questions with the answers you expect. Run them on a schedule. Compare today against last week, not against your memory. When something moves, you will usually find the cause in the list above within a few minutes.
How often depends on how much changes around your model. If your content and instructions are stable, weekly is generally enough; if they are edited frequently, daily is better. Either is better than finding out from a customer.
Possibly, but it is the least likely explanation. Check your own documents, instructions and integrations first; those change far more often than model versions do.
No. Hallucination is inventing something that was never true. Drift is behaving differently than it used to. A system can drift without hallucinating, and hallucinate without drifting, so it is worth testing for both.
Only by having something to compare against. A fixed set of questions, the answers you expected, and a record of what came back each time.
Point it at your own model, run the tests on demand or on a schedule, and get a score and a report you can hand to someone else. Scheduled runs are the point of it; $30 a month covers up to four models and twenty runs a day.