Agentic SEO

A classic SEO tool crawls your site or reads your data and gives you a list: missing titles, pages losing clicks, searches you could target. You do the work. An agent takes the same inputs and does the work itself: it opens the page, rewrites the title, runs the build and shows you the change.

The loop is the defining part. An agent reads the data, decides what to change, makes the change, and can read the data again later to see what happened. Model Context Protocol (MCP), introduced by Anthropic in November 2024, is the common way agents connect to data sources such as Search Console, so the agent can ask for the figures it needs instead of waiting for you to paste them.

If you build with a coding agent, you already have most of an SEO agent. Give it your Search Console export or a connection to it, access to the repository, and a rules file, and it can work through the list a tool would have given you.

• Scaled content. An agent can write a hundred pages in an afternoon. Google's scaled content abuse policy applies whatever produced the pages: many pages made mainly to rank, not to help a reader, are spam whether a person or an agent wrote them.
• Invented figures. Asked to fill a comparison or a stat, a model supplies plausible numbers. Tell it to use only the figures in the data it was given, and to say when one is missing.
• No measurement. An agent that edits pages every day, with no dated log of what changed, leaves you unable to tell what helped. Keep changes small, one page at a time, and judge them after four weeks.