# AI agent for SEO: the missing piece is search data, not writing > Models write well enough. Choosing which posts to write needs search numbers, and MCP now lets coding agents fetch them. URL: https://claudegpt.si/news/ai-agent-for-seo-search-data-not-writing Published: 2026-11-07T15:59:27.000Z Updated: 2026-11-08T15:59:27.513Z Author: ClaudeGPT Si Newsroom Section: AI Tools Topics: AI agents, MCP, SEO Publisher: ClaudeGPT Si ## Key takeaways - A model alone cannot see how many people search a phrase, so AI-picked blog topics are often guesses. - Through MCP, coding agents such as Claude Code, Codex and Cursor can call SEO tools that return search volume, difficulty and trend. - Search numbers are estimates and cannot judge product fit, so a person should still approve the topic list. An AI agent for SEO can already write a decent blog post in minutes. What it usually cannot do is tell which posts are worth writing, because a model on its own has no idea how many people search for a phrase this month or how hard that search is to rank for. Give the agent that data through a tool, and the job changes from writing on a hunch to writing for searches people actually make. ## Why writing was never the hard part Ask Claude, ChatGPT or any other model for "ten blog ideas for a project management app" and the list will sound reasonable. Some of those ideas will have almost no searches. Others will repeat a post the company published last year. A few will be phrased the way the model talks, not the way people type into Google. None of that is a writing problem. It is a data problem, and it shows up before the first sentence is drafted. ## How agents get search data now Coding agents such as Claude Code, Codex and Cursor can call outside tools through the Model Context Protocol (MCP). An SEO tool provider that runs an MCP server exposes its checks as tools the agent can call by name, with the results coming back as data the agent can sort and filter. seogeoaeo.ai is one such server. Its [agents page](https://seogeoaeo.ai/agents) lists setup for Claude Code, Codex, Cursor, Windsurf, GitHub Copilot in VS Code and Claude Desktop, and the same tools are also available over a REST API. Each run costs credits, the same as in the browser. ## What the loop looks like A topic-picking session with an agent tends to follow the same order: 1. **Seeds from the brand.** The agent reads the company's own pages and names the problems the product solves and the competitors it is compared with. 2. **Numbers for each seed.** The Keyword Research tool (34 credits per run as of November 2026) returns related Google searches with monthly volume, difficulty, cost per click and trend. 3. **A ranked shortlist.** Keyword Metrics (33 credits) scores a list of up to 100 candidate phrases at once. 4. **No overlap.** Keyword Deduplicator and Keyword Clusterer (3 credits each) merge near-identical phrasings and group the rest, so each post targets one search. 5. **A human check.** The agent shows a table of topics, volumes and reasons, and a person approves it before any writing starts. After a post goes live, the agent can run the On-Page SEO Checker against the target search to see how well the title, headings and body cover it. ## What this does not fix The numbers are estimates. Search volume and difficulty from any provider are modeled, so they are good for comparing topics with each other and weak for predicting traffic. The data also cannot judge fit. A search can be popular and still be wrong for a brand if the product is not an honest answer to it. That call stays with the person reviewing the table. And the facts in the post still need sources. An agent that picks a perfect topic can still invent a price or a feature unless it is told to use only pages it actually opened. ## Why this matters for AI search too Pages that answer one clear question, with the answer near the top, are easier for search engines and AI answer engines to quote. Choosing one search per post pushes writing in that direction by default. seogeoaeo.ai's Citability Analyzer (5 credits) flags sections that lack a clear opening answer, as an editing aid rather than a ranking prediction. ## Where to start For teams already using a coding agent, the setup is one command and an API key, described on [seogeoaeo.ai's agents page](https://seogeoaeo.ai/agents). Plans start at $29 a month on the [pricing page](https://seogeoaeo.ai/pricing). Teams that already pay for a large SEO suite with its own API can connect that instead; the order of the steps is what makes the difference. ## Frequently asked questions ### Can an AI agent do SEO on its own? It can research, write and check pages when it has tools for each step, but topic fit and factual accuracy still need a person to review them. ### What is an MCP server for SEO? A service that exposes SEO checks, such as keyword research, as tools an AI agent can call through the Model Context Protocol and get results back as data. ## Sources - [SEO, AEO and GEO for AI agents: MCP and API](https://seogeoaeo.ai/agents), seogeoaeo.ai - [seogeoaeo.ai tool catalog](https://seogeoaeo.ai/tools), seogeoaeo.ai - [seogeoaeo.ai pricing](https://seogeoaeo.ai/pricing), seogeoaeo.ai - [Model Context Protocol](https://modelcontextprotocol.io), Model Context Protocol - [Keyword Research](https://seogeoaeo.ai/tools/keyword-research), seogeoaeo.ai