AI assistants miss your latest launch because the model's training data stops at a cutoff date, and anything after it only appears if the assistant searches the web and finds clear, current pages. That's why a product released last month can be unknown, and a price changed last quarter can still be quoted wrong. This post explains how cutoffs work, how to diagnose the problem, and what to do around launches, rebrands and changes.
For the two-source model behind every answer, see how AI assistants generate answers.
What a knowledge cutoff is
A knowledge cutoff is the point in time after which a model learned nothing from its training data. Providers publish these dates in their model documentation, and they differ by model.
Anthropic's model pages are a useful illustration because they list two dates for some models: a "reliable knowledge cutoff" and a later "training data cutoff." For example, Anthropic lists Claude Sonnet 4.6 with a reliable knowledge cutoff of August 2025 and a training data cutoff of January 2026. The gap matters: a model may have seen a little about recent events but not enough to describe them well.
Three practical points follow:
- Cutoffs lag release dates. A model released today was trained on data collected months earlier.
- Recent events are thin. Even before the cutoff, the last few months are underrepresented because the web hadn't written much about them yet.
- Old models stay in use. Different products, plans and API integrations may run on models with different cutoffs.
Training gap or retrieval gap?
Before fixing anything, find out which problem you have. Run the same brand question twice: once with web search off (or in a setting that answers from model knowledge), and once with search on.
| Search off | Search on | Diagnosis | Fix |
|---|---|---|---|
| Wrong or unaware | Correct | Training gap only | Nothing urgent; keep coverage consistent for future models |
| Wrong or unaware | Still wrong | Retrieval gap | Fix crawlability, indexing and outdated pages |
| Correct | Wrong | Bad sources outranking good ones | Update or replace the outdated pages being cited |
| Unaware | Unaware, no citations | Not discoverable at all | Launch basics: indexing, third-party coverage |
When search is on, check the citations. If the assistant cites an old review, an old press release or your own outdated help article, that page is the problem. See AI hallucinations about your brand.
Why search doesn't automatically fix it
Web search closes the gap only when the assistant decides to search and then finds clear pages. Both steps can fail:
- No search triggered. For questions the model thinks it can answer from memory ("what does [brand] do?"), it may not search at all, and answers from training data.
- Search finds the old version. Older pages often have more links, more coverage and more time to rank than your launch announcement.
- The new page is vague. A launch page full of slogans and no plain facts gives the model little to quote.
The launch checklist
Treat every launch as an AEO event, not just a PR event. The goal is that a search-grounded answer the week after launch is accurate, and that the next model's training data describes it consistently.
Before launch
- Write a plain fact sheet: product name, what it does, who it's for, price, availability, regions
- Decide how the name will be written everywhere, including capitalization and spacing
- Prepare updates for every page that will become outdated: pricing, comparisons, docs, FAQs
- Brief partners, resellers and marketplaces on exact wording
Launch day
- Publish a dedicated, crawlable page that states the facts in the first paragraph
- Update your pricing, product and comparison pages at the same time
- Update XML sitemaps with accurate lastmod dates; see XML sitemaps for AI search
- Submit key URLs through Bing Webmaster Tools or IndexNow; see Bing Webmaster Tools and IndexNow
- Update your Organization schema, social profiles and directory listings
First 30 days
- Earn third-party coverage: reviews, roundups, trade press, community discussion
- Run your launch prompts weekly with search on, and check citations
- Fix any outdated page that keeps getting cited
Rebrands, renames and price changes
Changes to existing facts are harder than launches, because the old version is already everywhere. The model, and the web, need an explicit bridge between old and new.
- State the change plainly. "Acme Analytics is now Northwind (formerly Acme Analytics)" on your home page, about page and in schema
alternateName. - Redirect, don't delete. Old URLs should 301 to their new equivalents. See site migrations and redesigns.
- Update the sources AI cites. Review profiles, Wikipedia and Wikidata where appropriate, directories, partner pages and major roundups.
- Keep dated facts dated. "Pricing updated October 2026" helps assistants prefer the current version.
See fixing negative or outdated information for working through third-party sources.
Example: a product launch six weeks later
This is an illustrative example of a common pattern.
A software company launches a new product tier in September. Six weeks later:
- Search off: the assistant says the company "offers two plans," because the model's training predates the change.
- Search on: the assistant mentions the new tier, but quotes an old price, citing a 2025 roundup article.
- Diagnosis: retrieval works, but an outdated third-party page outranks the company's own pricing page for that sub-question.
- Fix: add the price in plain text near the top of the pricing page, ask the roundup's editor for an update, and add the tier to the company's review platform profiles.
The training gap needs no action now. If the facts are consistent across the web, the next model trained after the change is far more likely to get it right.
What you can't control
You can't choose when providers train new models, which data they include, or when a given product switches models. You also can't remove old information from an existing model. What you control is how clear, consistent and widely available the current facts are, which serves both search today and training later.
The bottom line
Knowledge cutoffs mean AI models are always somewhat behind, and search only fills the gap when it runs and finds clear, current pages. Diagnose each wrong answer as a training gap or a retrieval gap, treat launches and rebrands as AEO events with crawlable fact pages and coordinated third-party updates, and keep your facts consistent so the next generation of models learns the current version of your brand.
