Your AI chatbot isn't broken. Your website is out of date.

A support lead sends the screenshot: the chatbot told a customer the Starter plan includes API access. It doesn't. It hasn't since March. The instinct is to blame the AI, tweak the prompt, maybe shop for a smarter model.

Then someone checks the citation and finds the sentence, word for word, on a pricing comparison page that was never updated after the March change. The bot wasn't hallucinating. It was reading. And it was right about what the website said.

Two kinds of wrong

There's invented-wrong, where the model fills a gap with something plausible. Grounding and citations fix that structurally, and we've written about it before. Then there's faithfully-wrong, where the agent accurately repeats content that is no longer true. No model upgrade touches the second kind, because the model did its job.

Faithfully-wrong is far more common than teams expect, and it's a good sign in disguise. It means the agent is doing what you asked, and it means the fix is a page edit, not a support escalation to your vendor.

A bot is only as fresh as its last crawl

Every grounded agent works from an index of your site taken at some point in time. If that index was built once at setup, it starts ageing the same afternoon. Prices change, a policy gets softened, a product is discontinued, and the agent keeps quoting the old world with full confidence and a valid citation.

So the boring, decisive feature is the re-crawl schedule. Your agent should re-index your site automatically, at least weekly, and immediately when you ask it to after a big change. Update the shipping policy page on Tuesday and the answer should change by Wednesday, citation and all, without anyone touching the bot.

Turn the misses into a maintenance list

The other half of staying current is knowing what the agent couldn't answer. A good one logs every question where retrieval came up empty or the visitor said 'that's not what I meant'. Read that log weekly. It's a ranked list of pages that are missing, ambiguous, or contradicting each other.

Do this for a month and something shifts. The website gets clearer for humans, because you're fixing the exact sentences that confused a machine. Support volume drops, because the answers are now where people look. And the chatbot stops being a thing you apologise for and becomes the fastest way to find out what your own site actually says.