SEO Content Analysis MCP

An MCP server for Claude Desktop that gives Claude the tools to do a full SEO audit on unpublished content — HTML, Markdown, TSX/JSX, or Word (.docx) files.

It covers classic on-page SEO (meta tags, headings, keyword placement, readability) and the signals that decide whether content gets surfaced by AI answer engines — chunk-level retrievability, structured data, and E-E-A-T.

Tools included

Scoring & full audits

AI answer engines & trust signals

Keywords & topic coverage

Meta tags & SERP

Structure & readability

Input conversion

Related MCP server: gsc-mcp-server

Setup

1. Install dependencies

npm install

2. Add to Claude Desktop config

Open %APPDATA%\Claude\claude_desktop_config.json and add:

{
  "mcpServers": {
    "seo-content-analysis": {
      "command": "node",
      "args": ["C:\\absolute\\path\\to\\seo-mcp-server\\src\\index.js"]
    }
  }
}

3. Restart Claude Desktop

After saving the config, restart Claude Desktop. You should see 15 SEO tools available in the tools palette.

4. Activate SEO Skill (Recommended)

To ensure Claude follows the correct workflow (asking for keywords/meta tags before auditing), you should provide it with the instructions in SKILL.md.

1. Make a copy of SKILL.md and name it seo-content-analysis.md.

2. Upload this file to your Claude chat or add it to your Project Knowledge (if using Claude Projects).

3. If you uploaded it to a chat, ask Claude: "Read seo-content-analysis.md and follow these pre-flight instructions for all SEO audits in this thread."

Usage

Once connected, you can ask Claude things like:

• "Audit this blog post for SEO" — paste content, or provide a local file path

• "Score this draft out of 100" — give it a .docx path or paste the text

• "Will this get cited by AI Overviews?" — runs check_ai_retrievability

• "Check my schema markup" — runs check_structured_data

• "Does this have enough E-E-A-T?" — runs check_eeat_signals

• "Check keyword density for 'cohort analysis'"

• "Suggest a meta title and description for this article"

• "Is the heading structure correct in this Word doc?"

Providing content

Tools accept either:

• content — paste raw HTML, Markdown, or plain text directly

• filepath — an absolute path to a local file (.docx, .html, .md, .txt, .tsx, .jsx)

Note: Files uploaded via the Claude chat interface cannot be accessed as a filepath. Paste the text content directly instead.

Set site_domain for any site that isn't the default

calculate_seo_score, analyze_links and check_eeat_signals accept a site_domain parameter (e.g. "example.com"). It decides which links count as internal.

The default is veritly.co. If you audit content for a different site without setting it, every absolute link to that site is counted as external — which silently inflates the external link count and distorts both the Link Profile score and the E-E-A-T citation tiers.

Scoring

calculate_seo_score returns a percentage out of a 215-point maximum:

Grades: A (90–100%) · B (75–89%) · C (60–74%) · D (40–59%) · F (<40%)

The maximum scales to the inputs

Categories that cannot be assessed are removed from the denominator, so grades stay comparable:

• No expected_terms → Topical Authority (35) excluded

• No primary_keyword → Keyword Optimisation (25) excluded

• Non-HTML input (Markdown, .docx) → Structured Data (15) excluded, since JSON-LD only exists in HTML

sub_scores additionally reports AI Retrievability, E-E-A-T and Structured Data as standalone percentages.

Results carry a scoring_version. Version 2 measures title and description width in pixels rather than characters — scores are only comparable within the same version.

Why pixels, not characters

Google truncates on rendered width. Sixty characters of W is roughly 1,133px; sixty characters of i is roughly 266px — the first is cut off long before the limit, the second wastes two thirds of the space. Both pass a 60-character check. Widths are estimated from Arial metrics (no browser required) and are approximations: treat anything within ~5% of a limit as borderline.

Project structure

seo-mcp-server/
├── src/
│   ├── index.js                      # Entry point
│   ├── server.js                     # MCP server setup & request routing
│   ├── utils/
│   │   ├── content.js                # Parsing, section chunking, sentence splitting,
│   │   │                             #   link classification, intent detection
│   │   ├── docx.js                   # Word (.docx) → HTML via mammoth
│   │   └── loader.js                 # Shared file-loading utility for all tools
│   └── tools/
│       ├── index.js                  # Auto-discovers and loads all tool modules
│       ├── analyseContent.js
│       ├── analyzeLinks.js
│       ├── calculateSeoScore.js
│       ├── checkAiRetrievability.js
│       ├── checkEeatSignals.js
│       ├── checkHeadingStructure.js
│       ├── checkKeywordDensity.js
│       ├── checkMetaTags.js
│       ├── checkReadability.js
│       ├── checkSemanticCoverage.js
│       ├── checkSeoPreview.js
│       ├── checkSnippetOptimization.js
│       ├── checkStructuredData.js
│       ├── convertTsxToHtml.js
│       └── suggestMetaTags.js
├── SKILL.md                          # Master instructions for Claude (copy and rename to use)
├── package.json
└── README.md

Adding a new tool: create a new file in src/tools/ that exports { schema, handler } and call await loadContent({ content, filepath }) at the top — nothing else needs changing.

Shared helpers worth knowing about

Anything reused across tools lives in src/utils/content.js, so two tools can't drift apart on the same judgement:

• getSections($) — splits content into retrieval chunks, one per H2, walking nested markup so React/CMS output chunks the same as flat Markdown

• isQuestionHeading() / detectPreamble() — shared by check_snippet_optimization and check_ai_retrievability so they can never contradict each other on whether a paragraph answers directly

• classifyHref() — internal/external/other link classification, subdomain-aware

• extractPlainText($) — body copy only; works on a clone, strips <script>/<style>, and strips nav/header/footer/aside only when they sit outside <article>/<main>

• detectIntent() / checkIntentAlignment() — search intent classification used by the meta tag tools

Dependencies

• @modelcontextprotocol/sdk — MCP server framework

• cheerio — HTML/DOM parsing

• mammoth — Word (.docx) to HTML conversion

• marked — Markdown to HTML conversion