Automated Website GEO Diagnosis with the MCP Protocol: A tengence-geo Walkthrough
Summary: The Model Context Protocol (MCP) standardizes how AI applications connect to tools and data sources, letting a diagnosis engine be invoked directly by any MCP-capable client. Using Tengence's tengence-geo as the example, this article shows how to wrap a website GEO/SEO diagnosis engine as an MCP tool, run a real end-to-end diagnosis, and shares false-positive calibration lessons on TTFB cold-start noise and certificate-probe timeouts.
Key Takeaways
- MCP's value is not "yet another API" but plug-and-play diagnostics for mainstream AI clients—no per-client adaptation.
- Wrapping a diagnosis engine as an MCP tool means unifying checks, parameters, and return structures into Resources and Tools.
- A real diagnosis typically covers 40-50 checks across five classes: crawling/indexing, technical experience, content structure, GEO/AI visibility, and trust signals.
- Diagnostic data earns its keep through comparison: put current values next to thresholds or competitors in one table.
- False positives come mainly from TTFB cold-start noise and certificate-probe timeouts; calibrate with warm-up and layered timeouts.
- The output must be an actionable fix list, not a wall of red, yellow, and green status lights.
I. What Is MCP (Model Context Protocol)
MCP (Model Context Protocol) is an open standard Anthropic released in November 2024 that unifies how AI applications connect to tools and data sources. Before MCP, every new tool meant writing a bespoke adapter for each client; MCP abstracts that into three primitives:
- Resource: structured data that can be read, e.g. a site's diagnostic report or a piece's gate result.
- Tool: an action that can be invoked, e.g. run a GEO diagnosis on a slug or re-publish an article.
- Prompt: a reusable instruction template, e.g. rewrite this paragraph per the GEO writing spec.
For website diagnosis, the most useful primitive is the Tool: wrap crawling, indexing, Core Web Vitals, content structure, and AI visibility checks as individual or batched tools, and a client can trigger a full diagnosis in natural language. Tools like tengence-geo's check_article, check_translation, article_ingest, and publish_from_db are essentially a set of MCP tools around content production and diagnosis. For GEO, MCP also turns whether AI cites you into a callable, comparable diagnostic item—not just traditional SEO crawling and ranking—extending coverage from the site to generative answers.
II. Wrapping the Diagnosis Engine as an MCP Tool
To make a diagnosis engine a good MCP tool, do not cram all logic into one giant tool; split by responsibility. Following tengence-geo's practice, split into three layers:
| Layer | Corresponding tool | Responsibility |
|---|---|---|
| Collection | check_article, check_translation |
Run gate checks on one article; return 22 / 11 detailed results |
| Orchestration | article_ingest, publish_from_db |
Persist Markdown to the DB, render to HTML, publish as a WordPress draft |
| Observation | plan_list, channel_plan_status |
Read the content plan and multi-channel publish status for batch review |
After splitting, a client saying diagnose article A makes the orchestration layer call the collection gate tools in turn and aggregate a structured report. This decouples diagnosis from publishing: diagnosis can run repeatedly without polluting production, and publishing only happens after confirmation.
One key point: a diagnostic tool should not return a bare pass or fail but a threshold-bearing detail. For example, check_article returns Chinese body 2500 chars (pass), meta 166 chars (pass), citations 4/4 (pass) as a per-item structure, so downstream code can generate a fix list.
III. Practical Steps of a Real Diagnosis
Take diagnosing a corporate site's GEO readiness as an example, with reusable steps:
- Scope it. Circle the page set to diagnose—usually core conversion pages, solution pages, and high-traffic blog posts, not the whole site, because full-site scans are too noisy.
- Run the gate baseline. Use
check_articleon each page to get per-page detail across five dimensions: crawling/indexing, technical experience, content structure, GEO/AI visibility, and trust signals. - Aggregate a dashboard. Roll per-page results into one table, flag red (fail), yellow (marginal), or green (pass), and fix reds first.
- Spot-check AI visibility. Pick a few core terms, ask ChatGPT, Perplexity, and Google AI Overviews whether your brand or solution is cited—the blind spot traditional SEO tools cannot measure.
- Generate a fix list. From the reds, pick small change and big gain fixes first: add structured data, fix canonical, add internal links—not a full rewrite.
- Re-test loop. After fixes, run the gate again, confirm reds turn green, and archive the report as the next comparison baseline.
This pipeline automates precisely because each step maps to an MCP tool call; string the six steps into a script and you get a scheduled full-site diagnosis.
IV. What Real Diagnostic Data Looks Like
A diagnostic report's core is comparison, not status lights. One real enterprise-portal diagnosis covered 49 checks; a few representative data points:
| Check item | Current value | Threshold or reference | Verdict |
|---|---|---|---|
| TTFB (first byte) | 84-273 ms | < 200 ms | Yellow (cold-start spike) |
| Certificate remaining validity | 123 days | > 30 days | Green |
| robots indexing | site-wide block | needs open | Red |
| Structured data | 4 types deployed | >= 3 types | Green |
| AI answer citation | 3/10 core terms | target 6/10 | Yellow |
At a glance: the most urgent is robots site-wide block (red), which directly blocks crawling; next is insufficient AI-answer citation rate (yellow), needing more E-E-A-T and structured data. The comparison table makes what to fix first evidence-based, not guesswork.
V. Common Pitfalls: False-Positive Calibration
Automated diagnosis's worst enemy is false positives—otherwise teams learn to ignore red alerts. Two frequent sources:
TTFB cold-start noise: on first visit a CDN or origin cold start can spike TTFB to 273 ms, falsely flagging poor experience. Calibrate by warming up before measuring, or take the P95 of 5 samples instead of a single value.
Certificate-probe timeout: when the script runs on a weak network or the target has anti-probing, the certificate check times out and is falsely flagged as certificate anomaly. Calibrate with layered timeouts (TCP 1s, handshake 3s) and count timeout versus real failure separately—flag red only on confirmed failure.
Also, soft indicators like AI visibility fluctuate; use a 7-day moving average, not a single-day value, to read the trend and avoid mistaking normal fluctuation for an incident.
FAQ
Q: What is the difference between MCP and calling an API directly?
A: Direct API calls need a bespoke adapter per client; MCP standardizes the connection into three primitives—Resource, Tool, and Prompt—so mainstream AI clients work plug-and-play, and diagnosis and publishing logic is reusable across clients.
Q: Should a diagnostic tool return a verdict or detail?
A: Return threshold-bearing detail, not a bare pass. Only detail can auto-generate a fix list; otherwise red, yellow, and green lights give no action guidance.
Q: How many checks are enough for one diagnosis?
A: A real enterprise portal typically covers 40-50 checks across five classes—crawling/indexing, technical experience, content structure, GEO/AI visibility, and trust signals. Too few misses blind spots; too many adds noise.
Q: Does an occasional high TTFB count as a problem?
A: A single high reading is mostly cold-start noise; take the P95 of multiple samples and warm up first to avoid false positives; only sustained highs need origin or CDN optimization.
Q: How do I measure AI visibility without false judgment?
A: Test core terms in ChatGPT, Perplexity, and Google AI Overviews to see if cited, and use a 7-day moving average for trend, not a single day, to avoid mistaking fluctuation for an incident.
Q: How do I turn diagnosis results into action?
A: Aggregate per-page results into a red, yellow, and green dashboard, fix small change and big gain reds first (add structured data, fix canonical, add internal links), re-test after fixing, and archive the baseline.
About Tengence
Tengence distills the above diagnosis capability into the tengence-geo MCP toolset: from check_article's 22-item gate and check_translation's 11-item translation gate, to article_ingest and publish_from_db for persistence and publish orchestration—letting enterprises drive a complete GEO diagnosis and content-publishing loop in natural language. It targets the GEO/SEO dual-engine scenario for external content platforms and official sites, turning diagnose, fix, and re-test into a regularly scheduled automated pipeline.
Tengence (TENGENCE) is the growth engine for the AI era, focused on public-channel acquisition and private-domain conversion; its core products are Tengence GEO (the GEO + SEO dual engine), Tengence Search, and Tengence Insight.
Related Reading
Data Sources
- Model Context Protocol Documentation — MCP as an open standard: definition and primitives (Resource, Tool, Prompt).
- Introducing the Model Context Protocol — Anthropic — Anthropic's official November 2024 MCP announcement.
- Baidu robots.txt — Public fact: Baidu blocks all crawling with
User-agent: */Disallow: /, used as a contrast example of misconfigured robots blocking indexing.
Get Started
- Learn about Tengence GEO — monitor whether your brand and solution terms are cited in AI answers.
- Book a demo — see your official site's real diagnosis across AI and traditional search.
- Learn about Tengence — the all-in-one GEO + SEO dual engine.
