Effective cutoff
The effective cutoff is the point where an AI model's knowledge starts to thin out, which is usually earlier than its official training cutoff.
The effective cutoff is the point where an AI model's knowledge starts to thin out. Every model has a training cutoff, the date its training data ends. But the last months before that date are usually thinly covered, so in practice the model knows less about recent events than the official date suggests.
For a brand, that means a model answering from memory may describe you as you were a while ago: last year's prices, an old product range, a name you've since changed.
Why it comes earlier than the stated cutoff
The web takes time to catch up with anything new. When you launch a product, your own page is usually the first one about it. Reviews, news stories, comparisons and forum threads arrive over the following months.
A model trained soon after your launch sees only the first few of those pages. It has a lot of material about what came before and very little about what's new. So even though the launch falls inside its training data, it hardly knows about it.
An example
For example, say Harrow & Pine launched a new grinder in March, and a model's training data ends in June. By June there were a handful of pages about the grinder. There were hundreds about Harrow & Pine's older models. Asked about Harrow & Pine's range, the model is likely to lead with the older ones.
What it means for you
When a platform searches the web before answering, it can make up for this by reading current pages. When it answers from memory, it can't. So for anything recent, it helps that the facts are easy to find on pages the platforms search, starting with your own site, and that old pages say clearly what has changed.