2026-09-22 • 5 min read • Measure
How to Find Documentation Gaps With MCP Analytics
How to read the questions your MCP server could not answer, drop the ones that are out of scope, and write the section people asked for.
Nobody files a ticket to say your documentation does not mention webhook overlap. They ask an agent, get nothing useful, and message a teammate. If the agent was connected to your docs, that question was a request, and the request is the gap.
Page views will not show it. The page was never opened.
What a gap is
A question in scope, not already answered under another heading, asked more than once. One question is an anecdote. A cluster is a backlog item when it is frequent, the task matters, and the current answer is thin.
What the questions people ask your docs can tell you is the longer review, including which view is free and which is paid. This post is the gap itself.
Why the list is not every miss
Out of scope is not a missing page. "How do I rotate a webhook key?" belongs in API docs. "What is the office wifi password?" does not, unless you indexed that runbook on purpose.
A heading problem is not a missing page. If the answer exists under another name, rename the heading or put the reader's words in the first lines. See Make your documentation AI-ready.
A retrieval rule cannot fill a gap. A rule points at a source you already have. See retrieval rules.
Read the questions in MCP Studio
MCP Studio is the no-code builder on Appa Tools. Open the server on the dashboard.
- Start with the free charts. Every server shows volume, success, and which tool and source ran. That tells you people called the server. It does not list what they asked.
- Open question text on Core analytics, $49 a month. Content gaps, which group the questions nothing matched, are Action analytics, Core plus Action at $99 a month. New accounts get a 30 day trial with analytics included. Per-request detail is recorded on the calls you already made. A higher tier fills its views with that history.
- Limit who can open the log. Question text is other people's data. Treat it like a support transcript.
- Pick one cluster. Note how many requests asked, so a payment-blocking question outranks a rare edge.
The view names are in Action and Predictive analytics.
Check that it works
Twelve requests ask how to rotate a key without dropping webhooks. The docs explain rotation. A different article explains webhook retries. Nothing states the 24 hour overlap next to the rotation steps. Both topics exist. The task the person asked does not.
Write that section. Refresh the source. Ask the logged wording again. A citation to the new section closes it. A different page means the heading still is not the one search lands on. The rerun belongs on the sheet in Test an MCP server for accuracy.
If the cluster is not a gap
Drop it when it is out of scope, or when the answer is already on a page people are not finding. Pages the agent does cite, including the ones you did not expect, are Measure which documentation an AI agent actually uses.
Read one cluster today
Open analytics in MCP Studio and find one repeated question with no matching passage. If it is in scope and unwritten, add the section and ask the logged wording again. The gap is closed when that question cites the new section.