← Back to Blog

2026-08-25 • 4 min read

Docs MCP Metrics

Once your Docs MCP is live, you can see the questions people ask it, the passages it returns, and the gaps in your content. A guide to what's measurable and how to act on it.

The short version

A Docs MCP answers questions about your documentation. That means it can also tell you what those questions were, which passage answered each one, and where your content came up short. This post covers what's measurable, what each measurement is good for, and how to turn it into a work queue.

Why this is different from docs analytics

Traditional documentation analytics tell you which pages were opened and how long someone stayed. You infer intent from that: a spike on the auth page probably means people are struggling with auth.

A Docs MCP removes the inference. When an AI assistant queries your content, the request contains the actual question, and the response contains the exact passage that was returned. You get the question and the answer, not a proxy for them.

That matters because most teams have been working without it. GitBook's 2025 report found 39% of documentation teams don't track metrics at all.

What you can measure

Usage

The basics, and a reasonable place to start: total requests, success rate, response time, which tools get called, which sources get referenced, and which AI clients your users connect from.

Two of these are more interesting than they look. Sources that never get referenced are either badly targeted or badly written, and either way they're worth revisiting. Client breakdown tells you where your users actually work, which is useful when you decide which setup instructions to write first.

The questions themselves

Every request carries the query text. Read as a list, this is the most direct signal you'll get about what people come to your product to find out, in their own words rather than the words you chose for your headings.

The unanswered ones are the most valuable. A question that came back with nothing useful is a specific, sourced request for content that doesn't exist yet.

The exact passage that answered

For each request you can see the page, the section, and the excerpt that was returned. This tells you which of your writing is doing real work.

It also traces code. If an AI assistant hands your user a snippet, you can see which source, page, and section it came from, which is the difference between "our docs are being used" and knowing exactly what's being copied into someone's project.

Retrieval confidence

Some pages get found and still answer poorly. They match the question well enough to be retrieved, then don't contain what the reader needed.

Confidence scoring per page surfaces those. They're usually the highest-leverage edits you can make, because the page is already ranking for the right question — it just needs to actually answer it.

Content gaps

Gaps are assembled from real queries that returned nothing confident, grouped so you see the shape of what's missing rather than a list of one-off questions.

Source grades

Each source gets an A–D grade based on how efficiently it produces answers. A source that consistently takes several requests to resolve a question is telling you something structural: usually missing examples, or headings that don't match how people ask.

Requests to a working answer

How many back-and-forth requests it takes before an agent reaches something useful. Lower is better, and the trend over time is the clearest single indicator of whether your content is improving.

Ranked fixes and forecasts

Rather than a dashboard you have to interpret, the fixes are ranked by how many requests each would affect, with a projection for what happens after each one. This turns "improve the docs" into an ordered list.

What's free and what isn't

The KPI cards, requests over time, and the breakdown charts are free on every server.

Beyond that, analytics subscriptions stack: Core at $49/mo adds the request log, Core + Action at $99/mo adds passages, code provenance, confidence, gaps, and the impact model, and Core + Action + Predictive at $149/mo adds requests-to-answer, source grades, ranked fixes, and forecasts.

One detail worth knowing: the underlying telemetry is recorded on every request regardless of which tier you're on. Turning a tier on fills the dashboard with your existing history rather than starting a new clock. New accounts also get a 30-day trial with all of it included.

Turning it into a work queue

The measurements are only useful if they change what you write next. A loop that works:

  1. Start with the unanswered questions. They're specific and they came from real users.
  2. Then the low-confidence pages. These are edits, not new pages, so they're cheaper.
  3. Check your worst-graded source. A C or D usually means a structural problem rather than a content one — missing examples, or sections that don't match how people ask.
  4. Publish, then watch requests-to-answer. If the number drops, the edit worked.

That last step is what makes this different from guessing. You made a change, and you can see whether it helped.

For a one-time check on whether your Docs MCP improves answer quality at all, we published a five-dimension scoring method and results from a real test in How to know if your docs MCP is actually accurate. That's a benchmark you run once; the metrics here are what you watch continuously.

If you're embedding this in your own product

If you offer MCP server creation to your own users through the MCP Studio SDK, the same measurements work at two levels.

Your users get the dashboard described above for the servers they create, and you choose which cards appear and how it's styled. You also get an aggregate view in your developer portal: adoption and activation rates, the questions coming in across all your users, and which integrations are growing or going quiet. It's the same telemetry, reported at the level of a business rather than a single server.

Explore the MCP Studio SDK →

Getting started

If you already run a Docs MCP, most of this is waiting in your dashboard. If you don't, the wizard takes about a minute and the free tier covers a real first server.

Build your Docs MCP for free →

Keep going