Semrush writes up a keyword research workflow that pipes competitor reviews, Search Console CSVs, and its own MCP into a Claude project

Semrush published a walkthrough on Thursday of a keyword research workflow that lives inside a single Claude project, with the business context, the competitor reviews, the Search Console CSVs and the Semrush data all reachable in one place. Author Chris Hanna, with contributor Faizan Ali, describes four steps: set up a Claude project with instructions and business context, find keyword opportunities from competitor gaps and Search Console easy wins, analyze what customers actually say in reviews and sales calls, then prioritize terms by business relevance. The Semrush MCP server sits in the middle of it.

What goes into the project, on day one

The project starts with uploaded context files: a business context document, existing page URLs, Google Search Console CSVs, competitor reviews, Reddit threads, and sales transcripts. Claude reads all of it and keeps the state across every prompt in the project, so the search team stops re-pasting the same background into a fresh chat.

From there the workflow leans on a keyword gap analysis with filters for minimum search volume, difficulty range, and intent type. Hanna's post gives one worked example: the keyword "best community for new freelancers" has zero monthly searches, and the variant "best communities for freelancers" has 20 monthly US searches. That is the kind of correction a raw brainstorm inside a model misses, because the model has no idea what people actually type.

The Semrush MCP is where the numbers come from

The Semrush MCP is what closes that gap. It exposes Semrush's keyword data to Claude as tools the project can call, so the model validates every candidate term against the live volume and difficulty before the team writes anything. The workflow explicitly runs unvalidated keywords through the MCP first, then keeps only what returns real numbers.

Hanna gives a second example from the review-mining step: a G2 review noted that Trello lacks native task dependency support, which then became the topic hook for a competitor. SmartSuite wrote content about managing task dependencies and now appears cited inside Google AI Overviews for related queries. The path from a customer complaint in one tool to a citation in another is the whole point of routing every step through one project that remembers.

The consequence for a working search team

The post's output is a CSV with topic clusters ready for content planning, which sounds like the same output a spreadsheet workflow already produces. What changes is where the effort goes. The seed brainstorm, the gap analysis, the review synthesis and the volume validation all happen in one place, with the same instructions and business context, so the search team spends the hour on judging which cluster is worth writing.

Two things the post does not measure: how often the model invents a keyword that the MCP later kills, and how the workflow holds up on a large content site with thousands of pages and hundreds of competitors. The examples are small on purpose. The pattern is worth copying anyway, because it moves the "context you paste every time" out of the chat and into project files, and gives the model a real data source instead of a guess.

Source

Semrush: How to do keyword research with Claude and Semrush.