Training data cutoff
The point beyond which a model knows nothing unless it goes and looks. It explains most cases of an engine describing a business that no longer exists in that form.
Checkable · the test is below
What it means
Models are trained on material up to a date. Anything after it is invisible to the model itself, though retrieval can supply current material on top.
Why it matters for an HVAC company
It is why an engine can recommend a company that closed, quote a price you stopped charging, or give an address you moved out of. None of that is malice or a ranking penalty, it is an old snapshot being read aloud.
What would settle it
Ask the model its cutoff, then test it with something you know postdates that. If it knows a fact from last month, retrieval is working. If it insists on an old one, you are talking to training data.
A worked example
An engine recommends a contractor at a unit they left in 2024 and gives the old phone number. Their current details are correct everywhere on the live web. The fix is not on their website, it is that the stale record still exists on several directories the model was trained on.