What it means
Every model is trained on data collected up to a certain date. After that, its knowledge stops. A model with a cutoff in early 2025 doesn't know about products launched in late 2025, unless it searches.
Cutoffs are usually several months to a year before a model is released, and models stay in use long after.
Why it matters for brands
- New brands may be unknown to the model's own memory.
- Old facts stick: last year's prices, old features, a previous name.
- Old leaders may be favoured because they were written about most before the cutoff.
How search fills the gap
Most AI tools now search the web for product questions, which brings in current pages. But the model's own knowledge still shapes which brands it searches for and how it reads what it finds. In our study, ChatGPT put brand names into its web searches for 109 of 210 questions before reading anything.
Example
A startup launched in 2025 asks a model with search off "what is [startup]?". The model says it doesn't know. With search on, it finds the startup's site and two reviews, and answers correctly.
Frequently asked questions
How do I find a model's cutoff?
AI companies usually publish it in their model documentation.
Will the next model know about my brand?
If it was written about on widely read sites before the next cutoff, probably.
Can I get AI to forget old facts about me?
Not in the model itself. But clear, current pages help answers that search the web get it right.