AI Reads Today's Web, Not Last Year's

There’s a common misconception that AI models only know what they were trained on, frozen at some cutoff date. Most consumer AI search today doesn’t work that way. ChatGPT, Perplexity, Google’s AI Overview, and Copilot all do live retrieval, pulling and re-reading pages at query time, not just recalling training data.

That changes the incentive completely. A page that’s accurate but three years old isn’t a historical record the model politely ignores. It’s a source the model might cite right now, today, as current fact.

What “fresh” actually means to a crawler

Freshness isn’t just a published date in your frontmatter. It’s a mix of signals:

Signal	What it tells the crawler
Last-modified date / visible update	Whether this page has been maintained
Internal links from newer pages	Whether the site itself still treats this page as relevant
Consistency with your other current pages	Whether facts on this page still match what’s true elsewhere on the site
Crawl frequency for the domain	How often the model bothers to re-check this page at all

A page can be technically live and still functionally stale in the model’s eyes if nothing about it signals recent attention.

The quiet cost of an old page staying “correct enough”

We see this most with pricing, feature lists, and team/leadership pages — content that was accurate when published and just never got flagged as outdated because nothing on the page itself is technically wrong, it’s just old. The model has no way to know that, so it cites it as current.

What actually helps

Re-touching a page (a real update, not just re-saving the same content) gives crawlers a reason to re-fetch it sooner. Removing genuinely dead or superseded pages also matters — a model citing a page you’d have taken down years ago if you’d remembered it existed is a self-inflicted accuracy problem.

beket.ai flags which of your pages are being cited with outdated information, so you know exactly which ones need a real update instead of guessing.