Action and Predictive analytics

Use this guide to understand what the Action and Predictive tabs do, how each metric is calculated, and what plan is required to unlock each level.

Build absolutely anything...with the right context.

Audience	MCP owners, analytics teams, AI quality stakeholders
Prerequisites	Analytics enabled on your MCP server and enough request volume to generate trend data

Plan requirements

• Free charts: KPI cards, Calls Over Time, and Breakdown. Included with every account.
• Core Analytics ($49/month): Adds the request log.
• Core + Action ($99/month): Core plus the Action tab.
• Core + Action + Predictive ($149/month): Core plus Action and the Predictive tab.

New accounts receive a 30-day trial with all three paid tiers unlocked.

If you created a server inside another company's product, that company may already pay for a tier on your behalf. In that case the tabs are unlocked on that specific server without a subscription of your own, and any tier you buy separately still applies everywhere. See analytics and observability for how those grants work.

See full plan details on MCP Studio pricing and billing and pricing.

How both tabs are laid out

Action and Predictive share the same three-part structure, so you can read either one the same way.

1. A summary strip at the top. Four tiles, each a single number worth acting on, colour-coded green, amber, or orange. Click a tile and the section that explains it opens and scrolls into view.
2. A recommendation list, always open. Action calls it Do this next, Predictive calls it Fix these first. This is the part you act on, so it sits above everything else.
3. Collapsed evidence sections. Charts, passage-level detail, source grades, question chains, and the calculation method. Nothing is removed, it is just folded away until you ask for it. Each closed section still shows its own count or status in the header, and a section with a problem is outlined in orange, so you never have to open one to find out whether it needs you.

What the Action tab does

The Action tab gives source-level evidence for what the AI used and where content quality is costing time:

1. What the AI read for each request:
   • source page URL
   • section/title
   • exact text excerpt passed to the client
   • retrieval confidence score
2. Code snippet provenance:
   • language
   • source page and section
   • originating request context
3. Most-referenced pages and section pressure:
   • which pages are used most
   • which heavily-used sections have low retrieval confidence
4. Action recommendations ranked by urgency and by how many requests each one affects. The top three show by default, with the rest one click away.

The summary strip on this tab reports the share of requests answered from your content, how many came back empty, how many matched only weakly, and how many pages are carrying the load.

Every table and list here exports to CSV, including the passage detail and code provenance.

Impact model

The dashboard always reports counts measured from your own request history: questions answered from your content, time recovered, extra round trips, and code-backed answers.

Dollar figures are opt-in. A monetary result needs assumptions about what your time or your support contacts cost, so nothing is priced until you choose how. You pick one of three models:

• Internal time: value comes from time your team no longer spends looking for an answer.
• Support deflection: value comes from questions answered before they became a support contact.
• Both: counts each together, which is only accurate if an unanswered question really would have cost you both.

You can accept the published benchmark defaults or enter your own numbers, such as your blended hourly cost or your average cost per support contact. Your figures are saved once for your account and applied to every server.

Every assumption is shown with its provenance. The dashboard labels each number as published research, arithmetic derived from published research, or an assumption of ours, and links the sources so you can check them. The arithmetic behind each line of the result is printed in full.

Monthly projections appear only once there is at least two weeks of request history, because extrapolating a month from a few days is not meaningful.

What the Predictive tab does

Predictive adds source quality modeling and forward-looking operational forecasting:

1. MCP requests to a working answer (per question chain)
2. Source quality grades (A-D) based on answer efficiency
3. Projected requests-to-answer after content fixes
4. 30-day request forecast with confidence level, priced only if you have configured the impact model
5. Ranked content fixes showing the requests each one would avoid.

The summary strip on this tab reports average requests to a working answer, the share of question chains that reached one, how many were dead ends, and how many sources are graded below B. The 30-day forecast stays visible under the fix list, because it is the reason to work through it.

Requests-to-working-answer formula

requests_to_working_answer = requests in a question chain until first grounded answer

A question chain is grouped by MCP session ID when available, otherwise by client + rolling time window.
A grounded answer is detected when the top citation's retrieval confidence is high enough to indicate the question was actually answered from your content.

Source quality grading

Sources are graded by average requests-to-answer when that source is involved:

• A (Excellent): 1-4 requests
• B (Good): 5-9 requests
• C (Fair): 10-19 requests
• D (Needs improvement): 20+ requests

Each source row includes:

• grade + label
• average requests to working answer
• unresolved session rate
• weakest location/passages to improve
• recommendation text + projected requests after fixes

How to use Action and Predictive effectively

1. Enable the analytics tier that includes your target tabs.
2. Deploy with analytics enabled in the wizard.
3. Use Action to prioritize the top 1-3 content fixes each week.
4. Re-index sources after edits.
5. Use Predictive to verify that requests to answer trends move down.
6. Track grade changes by source over time and focus on C and D sources first.

Suggested operating cadence

• Weekly: review top actions and apply quick wins
• Monthly: compare forecast vs. actual usage
• Quarterly: reassess plan tier and dashboard scope

Data and privacy notes

• Action/Predictive telemetry is collected per request and displayed only within authenticated server dashboards.
• Stored fields focus on retrieval and quality signals (query text, citations, confidence, session grouping, result counts).
• Excerpts come from customer-indexed source material; no external data is injected into these metrics.
• Follow your internal retention requirements for MCP request logs and source content governance.