Knowledge Cutoff

Cutoffs matter less than they used to, because consumer assistants now search the web for anything current or local. But they still shape answers in quiet ways: a model's background sense of "who the established names are" comes from training data, and a newer business earns its way into answers through the live evidence layer rather than reputation the model absorbed in training.

The subtler effect is on what a model decides to search for in the first place. Retrieval is not neutral — the assistant forms an initial impression from memory, then looks for evidence about it. A model that absorbed a market's incumbents during training is more likely to search their names, read their pages, and confirm them. This is a mild but real incumbency advantage that grounding reduces without eliminating.

Cutoffs also age unevenly. Facts that rarely change — a long-established firm's specialty, a city's major hospitals — survive the cutoff comfortably. Facts that change constantly are stale the day training ends: staffing, pricing, hours, whether a business is still trading. Assistants are reasonably good at recognising which category a question falls into and searching accordingly, which is why local recommendation questions almost always trigger retrieval.

A related trap for anyone measuring this channel: models frequently misreport their own cutoff, and the date a provider publishes describes the bulk of training data rather than a hard boundary. Treat any specific date as approximate, and never design a measurement that depends on knowing it precisely.

Practically: if your business changed anything important since the big models last trained, the live web is your only messenger. Make sure it says the right thing — on your own pages, on your profile, and on the third-party sources retrieval is most likely to reach.