Every help center I have audited as a community manager has a few amazingly-good articles and a long tail of comprehensive-but-unread articles. The comprehensive ones get linked from the chat but never read past the first paragraph. The good ones get bookmarked and cited inside the company.
What 'good' looks like
I crawled 1,200 articles across 30 customer help centers and tracked which ones were cited inside their chat (i.e. the AI grounded a reply in them). The top quartile shared three traits. (1) Under 300 words. (2) Opened with the answer, not the context. (3) Used 'we' and 'you' (active voice), not 'users' and 'the system'.
What 'bad' looks like
“The AI almost never grounds in the long, hedged, passive-voice articles. Neither does the customer.”
The bottom quartile shared three traits too. (1) Over 1,200 words. (2) Opened with a header like 'Introduction' or 'Overview'. (3) Used passive voice ('a workspace can be configured' instead of 'configure your workspace'). The AI almost never grounded in these. The customer almost never finished reading them.
The three rules I now use
(1) The first sentence answers the question. Not introduces it. Not contextualizes it. Answers it. The next 200 words are how-to. (2) No 'Overview' or 'Introduction' headers. Start with action. (3) Active voice, second person. 'You set the SLA in Settings > Tickets > SLA' beats 'SLAs are configured under the Tickets settings panel.'
How to know if your docs are read
Three signals. (1) AI grounding rate: which articles does your AI cite when answering? Most KBs have a power law; 20% of articles answer 80% of questions. The 80% are dead weight. (2) Time-on-page from your analytics. Articles that get under 12 seconds of read time are either too long or too off-topic. (3) Reopen rate: how often does a conversation that was AI-answered with a citation come back within 48 hours? If it is over 5%, the article did not actually answer the question.