Every language model is trained on data collected up to some date, often called its knowledge or training cutoff. Model providers usually publish it: OpenAI’s model pages, for example, list a knowledge cutoff for each model, and for GPT-5 that date is 30 September 2024.
For a brand, the cutoff explains a common surprise. A product launched, renamed or repriced after the cutoff simply isn’t in what the model learned. If the model answers without searching the web, it may describe an old range, an old price or a company name you stopped using.
Web search is how current information reaches the answer. When a model searches, it can find and cite pages published after its cutoff. That is one reason to keep your own pages current and consistent: they are what a searching model can read today.
When Citationly records answers, it notes whether each one used web search, and answers without search are flagged wherever they appear, because they can’t see anything published after the cutoff. Read why ChatGPT gives a different answer every time.