Every language model is trained on data collected up to some point in time, called its training data cutoff or knowledge cutoff. Without a search tool, the model knows nothing about events, launches or price changes after that date, and its knowledge just before the cutoff is often thin. For brands, this means the model's built-in picture of your company can lag reality by a long time.
Why it matters for brands
If you rebranded, changed pricing, launched a new product line or moved upmarket after a model's cutoff, the model's built-in knowledge reflects the old version. When an assistant answers without searching, users get the old version. When it does search, your current pages have to be retrievable and clear enough to override what the model already believes.
Models are released some time after their cutoff and stay in use for a long time after release, so the gap between cutoff and the moment someone asks can be large.
Checking and working with cutoffs
Model vendors usually state cutoff dates in their model documentation. To see what a model believes about you, ask it with search turned off where the product allows that, then compare with answers that use search.
- Keep core facts consistent and current on your own site.
- Make change announcements explicit: 'As of March 2026, the Starter plan includes...'
- Update third-party profiles that are widely copied, such as directories and review sites.
- Expect older positioning to persist in parametric answers for a while.
Example
A rebrand the model has not heard of
Mossbank, a fictional payroll company, renamed itself from an older name after its cutoff-era coverage. Assistants answering without search still use the old name and old pricing. With search on, answers improve, but only after Mossbank adds a clear 'formerly known as' line to its About page and asks major directories to update their listings.
Common mistakes
Assuming AI answers reflect your current site
Without retrieval, they reflect whatever was in the training data at the cutoff.
Announcing changes without stating the old version
A model may not connect new and old names unless a source states the relationship plainly.
Questions
Not necessarily. Coverage depends on what was in the training data, and small or new companies may be barely represented.
Check the vendor's model documentation or release notes. Asking the model itself is unreliable.
