The five metrics that matter

Project settings → Analytics shows about a dozen widgets. The temptation is to track all of them. Resist. The five below cover ~90% of what's actually actionable.

1. Top pages (last 7 / 30 / 90 days)

The most-visited pages tell you what your users actually need. Two reasons it matters:

You should write more like the top pages. They're the ones being shared, linked from search, getting referenced in support conversations. The format and tone of these pages is working.
Top pages drive disproportionate impact. A 5-minute polish pass on your top-3 pages does more than rewriting all 50 of your low-traffic pages.

Sort by views, then by unique visitors. If a page has 10,000 views and 100 visitors, it's a single power user — not a popular page.

2. Search queries with zero results

If your full-text search is on (Pro), the analytics tab has a "Search queries" section. The most important entry is queries with zero results — what people searched for and didn't find.

Each row is a writing prompt. Your visitors literally typed the topic of a page that should exist. Write it.

After a few weeks, this list dwindles to noise (typos, weird phrasings) — that's fine, the value was the first dozen rows.

3. AI chat queries (Pro)

If your AI chat widget is on, the chat queries dashboard shows what people are asking the widget. Two cohorts to watch:

High-confidence queries: the widget answered well. These are FAQs that visitors are asking the widget instead of searching. If a query shows up here repeatedly, the corresponding doc page should be more prominent (sidebar position, internal links).
Low-confidence queries: the widget couldn't answer well. These are gaps — write the missing page.

Same energy as zero-result searches, but with more context (the full question, not just keywords).

4. Referrer breakdown

Where your traffic comes from. Useful for one specific decision: where to invest in writing for AI.

Search engines (Google, Bing, etc): traditional SEO. Long-tail content optimization works here.
AI assistants (chat.openai.com, claude.ai, perplexity.ai): AI-driven traffic. Cite-friendly markdown and a strong llms.txt matter.
Direct + bookmarks: existing customers. Worth ensuring your top pages don't change URLs.
Internal links from your marketing site: check the click-through is from the pages you'd expect.

If your AI-referrer share is climbing month over month, that's the signal to invest in AI discoverability.

5. Time on page (and bounce rate)

Less actionable, but useful as a second-derivative metric. Watch when:

A formerly-strong page's time-on-page drops. Something probably got harder to find on the page.
A new page has zero engagement. Either it's not linked from anywhere, or its title is misleading.

Don't chase higher time-on-page as an absolute. A clear, short page that gets visitors their answer in 30 seconds is better than a vague long page that holds them for three minutes.

Total page views. Vanity metric. It goes up because more people visit your site, not because docs improved.
Exit pages. Visitors leave from somewhere. The page they leave from isn't the cause of leaving.
Time-on-site. Same problem as time-on-page, in aggregate.
Geographic distribution. Interesting, rarely actionable. (Unless you're considering localization — see Internationalization for the trade-offs.)

How often to check

Once a week is plenty. Set a recurring 20-minute calendar block. Open the analytics, find the one thing that surprised you, and act on it. Done.

Daily checking trains you to react to noise. Weekly is signal.