GlossaryAI search
What is a knowledge cutoff?
Also called: training cutoff, training data cutoff, knowledge cut-off date
Definition
A knowledge cutoff is the point in time after which a language model has no training data. Anything that happened or launched later is unknown to the model unless it's retrieved at answer time.
Knowledge cutoff, explained
Models are trained on a snapshot of text collected up to some date, then released months later. Providers publish this date. Anthropic's model overview, for example, lists a "reliable knowledge cutoff" for each Claude model, and points to its Transparency Hub for both reliable-knowledge and training-data cutoffs. The word "reliable" is doing work: coverage of the last few months before a cutoff is usually thinner.
For a startup this has a blunt implication. If you launched after a model's cutoff, the model has never heard of you. Even if you launched before it, a young company with little coverage may barely register in training data. When someone asks about your category, the model's built-in answer will lean on older, better-documented competitors.
Web search is how assistants close the gap. ChatGPT, Claude, Perplexity, Gemini and Google's AI features can retrieve current pages and cite them. So for recent products, being found through retrieval matters far more than being in training data. That means crawlable pages, reachable by the assistants' search crawlers, that clearly say what you do.
Training data still matters over time. Future model versions will be trained on a web that includes your launch coverage, directory listings, reviews and discussions. Consistent, accurate descriptions published now become part of what later models learn. Think of it as two horizons: retrieval for this month, training data for next year.
When you test what an assistant knows about you, note whether it searched. An answer without citations is coming from training data and tells you about the past, not about your current pages.
Why it matters for founders
Most models can't know about a product launched this year from training alone. Your visibility in AI answers depends on retrieval now and on your web footprint for future models.
Example
A founder asks an assistant about their tool with web search turned off and gets "I'm not familiar with it". With search on, the assistant finds the launch page and a directory listing and describes the product correctly.
Common mistakes
- Judging AI visibility from answers that didn't use web search.
- Assuming a new model release will automatically know your product.
- Blocking AI search crawlers while your product is newer than every model's cutoff.
Sources
Checked
Related terms
- Large language model (LLM)A large language model (LLM) is an AI model trained on very large amounts of text to predict the next piece of text. That one skill lets it answer questions, summarize, write and reason in plain language.
- GroundingGrounding is giving an AI model real source content, such as search results or documents, at the moment it answers, so the answer is based on that content instead of only what the model memorized in training.
- AI crawlersAI crawlers are bots run by AI companies that fetch web pages, either to collect training data, to build an index for AI search answers, or to read a page a user asked about. Each has its own robots.txt name.
- ChatGPT searchChatGPT search is the feature that lets ChatGPT look things up on the web and answer with current information and links to its sources, instead of relying only on what the model learned in training.