What knowledge cutoff means
A knowledge cutoff is the edge of a model's memorized world: the date at which its training data ends. A large language model learns from a corpus assembled up to some point in time, and everything it "knows" from memory — every fact, brand, product, and price — reflects the web as of that date. Ask it about anything later and, absent live retrieval, it either admits ignorance or, worse, answers confidently from the old world.
Every major model has one, typically several months to a year behind its release date (training and evaluation take time). Providers usually disclose it, and models will state theirs when asked — though a model's sense of its own cutoff is itself only as reliable as its training.
Why the cutoff matters for brands
For a brand, a knowledge cutoff means AI engines carry a frozen snapshot of you — and speak about it in the present tense. Everything that changed after the cutoff is invisible to the model's memory:
- A rebrand or renaming — the model keeps using the old name, or treats old and new as different companies.
- Pricing changes — last year's prices quoted as current fact.
- Launched or discontinued products — new flagships unknown; killed products still recommended.
- Fixed problems — an outage, lawsuit, or bad review cycle from before the cutoff lives on in the model's associations long after resolution.
This is one of the most common sources of brand hallucinations: the claim was true once, the model just doesn't know time has passed.
How retrieval bypasses the cutoff
The cutoff bounds the model's memory, not the engine's answer. Most consumer engines — ChatGPT with browsing, Gemini, Perplexity, Claude with web search — augment memory with live retrieval: they search the web at answer time and ground the response in current pages via retrieval-augmented generation. When that works, a page you published last week can correct what the model memorized years ago. But retrieval is selective — engines skip it for questions they "think" they know — and it can surface stale pages, so cutoff-era facts still leak into answers. Two clocks run in parallel: model retrains reset the cutoff every several months, while retrieval updates continuously — the dynamic detailed in how often AI models update their knowledge.
Working with cutoffs in a GEO program
Three practices follow. First, make the current facts findable: since retrieval is the only fast path past the cutoff, keep authoritative, crawlable, up-to-date pages for exactly the facts that changed — the core of content strategy for AI search. Second, expect model releases to move your metrics: a new model version means a new cutoff and a re-learned view of your brand. Third, monitor for time-capsule claims: scheduled prompt tracking with accuracy checks catches engines still describing the pre-cutoff version of your company before customers act on it.