The short answer: two clocks run in parallel. Base models retrain infrequently — new versions ship every several months to a year, each with a fresher knowledge cutoff. But most consumer AI engines now also search the live web while answering, so what ChatGPT or Perplexity says about your brand can reflect a page published days ago. For anyone doing GEO, the second clock is the one you can act on this quarter.
The slow clock: model retraining
An LLM's core knowledge is frozen at training time. Everything it "knows by heart" about your brand — your category, your reputation, whether you exist at all — comes from the training data collected before its cutoff date. Between releases, that memory does not update. If you rebranded, repriced, or launched after the cutoff, the base model literally cannot know.
Providers ship new model versions on cycles measured in months, not weeks, and each new version trains on a fresher snapshot of the web. That makes every model release a scoring event for your accumulated footprint: the coverage, reviews, and consistent entity facts you built over the past year either did or didn't make it into what the next model memorizes.
The fast clock: live retrieval
The slow clock would be discouraging if it were the only one. It isn't. ChatGPT, Gemini, Perplexity, and Claude all ground answers in live web search when a question benefits from current information — the architecture called retrieval-augmented generation. At answer time, the engine issues searches, reads what ranks, and writes its answer from those pages, often attaching a citation to each claim.
This path updates as fast as AI crawlers and search indexes do. Publish a direct, factual page today, get it indexed, and it can be read into an answer within days. Correct a wrong fact on a page engines already cite, and the corrected version flows into answers on the next retrieval. The fast clock is why GEO produces measurable movement long before any retrain — the mechanism behind the timelines in how long GEO takes to work.
What this means in practice
- Retrieval-heavy questions update fast. Anything commercial, comparative, or current-events-shaped usually triggers a web search. Your levers: crawlable pages, updated third-party sources, structured data.
- Memory-heavy questions update slowly. Broad "tell me about X" or category-association questions often come from the model's weights. Your levers: sustained coverage and entity consistency, compounding across release cycles.
- The mix varies by engine. Perplexity retrieves for nearly everything; other engines answer more from memory. That's a core reason to track multiple engines rather than extrapolating from one.
Watch the updates land
Because both clocks tick on their own schedules — and model versions swap out underneath you without announcement — the only way to know what today's engines say is to keep asking them. Scheduled sweeps re-run your prompt panel across all four engines and timestamp every answer, so you can see retrieval-driven changes within a week and catch the step-change when a provider ships a new model. A free audit shows you the current state of all four engines in one pass.
Knowing when knowledge updates is half the picture; how AI models learn about your brand in the first place is the other half.