How To Do SEO With Claude Code

Claude has never run an SEO experiment. It has read an enormous amount of writing about SEO, which is a different thing, and left to its own memory it returns whatever the internet repeats most often.

Connect it to the SEOLinkMap Private MCP and the memory stops mattering. It reads your crawl, the correlation data for your keyword, and the page in front of it, then edits your source and measures the page again.

Where The Guessing Comes From

An LLM's opinions about ranking factors are a popularity ranking of sentences. "Aim for 2,000 words" and "keyword density should be 2 to 3 percent" appear in thousands of articles, so they come back strongly, and nothing in that loop ever tested them.

That is the honest case against AI and SEO, and it holds for as long as the model is answering from training. Computational analysis beats AI guessing covers the same failure with the numbers behind it.

What Changes With Measured Data

Supplying the numbers moves the model off recall and onto inference, which is the part it is actually good at. It compares your page to the pages ranking above you and reports where they differ.

The MCP hands over crawl output, audit findings, internal link counts, Search Console rows and SERP correlations for the keyword you care about. None of that came from training data. It came from a crawler and a statistics run.

Same Numbers Any Expertise Level

A correlation row is one artifact, and what you get back from it is a choice you make out loud. Tell Claude you have run SEO desks for thirty years and it stays in coefficients, p-values and variance. Tell it you run a plumbing company and it hands you three things to change this week and the reason each one is on the list.

Those are not two opinions. They are the same measured row rendered at two depths, and you can ask for the other one and check that the story survives.

That is what moves the required expertise off the reader. The measurement holds the rigor, so the person acting on it needs to be an expert in their own business and nothing else.

Start From The Scan

get_audit_summary returns your findings grouped into categories -- title, meta tags, content, media, performance, structured data -- and lists only the issues that actually fired. Clean checks are absent rather than reported as passing, so the response is the worklist.

Each issue arrives with a count and the thing it harms: rank, click-through rate, page speed, accessibility, SERP features. get_audit_issue_pages then takes one issue type and returns the URLs carrying it.

Nothing starts until you name the project. switch_project is documented to require your choice rather than let the agent pick, which matters when one account holds several sites.

Fix The Template Not The Page

Claude Code already has your source, so the fix lands in the template that generated the problem rather than in one rendered page. A title pattern that is wrong across a section of the site is usually one line in a layout file.

This is why the workflow lives in an editor instead of a chat window. The agent reading the audit is the agent holding the file.

Remeasure Before Moving On

analyze_url_content re-reads the live URL after your change and reports the same metrics again. The number moved or it did not, and Claude can see which without being told.

That loop is the whole difference between an agent doing SEO and an agent describing SEO. Skip it and you get a confident summary of edits nobody checked.

Target The SERP Not The Rule

get_serp_correlations returns per-factor coefficients, p-values and the observed range across the pages ranking for one keyword. The target is what those pages do, not a number from a best-practice list.

Keywords disagree, sometimes sharply, which is why the data is fetched per keyword instead of stored as a site-wide rule. A factor worth real work on one SERP is noise on the next.

Choose What To Ignore

Every factor lands in one of four buckets, and they are defined in terms you can argue with. CRITICAL means table stakes -- the top pages cluster tightly, so being outside the cluster is conspicuous. OPPORTUNITY means a strong correlation that is unlikely to be chance.

The other two are permission to stop. AVOID is a weak correlation, a poor p-value, or both, which is the statistical shape of an SEO myth. GAMBLE has a promising regression slope with no significance behind it, so spend there only if you like the odds.

Ignoring an issue on purpose is a decision. Ignoring it because no tool ever told you it existed is not, and the scan is what turns the second into the first.

This Loop Costs Nothing

Reading the audit, pulling correlations and re-measuring a page after an edit are data reads. They bill nothing and need no approval, which puts the agent on the same footing it has when it loads a webpage.

That is what makes the loop usable. Metering the remeasure would push you back toward assuming your change worked, and assuming is the habit this whole workflow exists to break.

The MCP Cannot Spend Money

Only a few operations create billable work -- a crawl, a research order, a new project -- and the MCP has no path to your money for any of them. It is not restrained from spending, it is incapable of it, and the strongest thing an agent can do with a paid operation is put it in a queue.

A queued item carries the operation, its cost, and the agent's reasoning for wanting to run it. That lands on the dashboard where you read what was proposed and why, then approve or reject it. An agent can queue a million dollars of research and none of it runs.

The approval code is the part an agent cannot reach. approval_code submits a code; nothing in the MCP issues one, so the only unattended path is the Agentic Approval API, which you stand up on purpose behind an agent key and an IP allowlist. That is you handing over a key, not an agent finding one.

Your repository is the part SEOLinkMap cannot see. Which fixes an agent commits outright and which ones open a pull request is a policy you write on your side, and it is worth writing before you leave the loop running.

Where This Still Guesses

The measurements are computed and repeatable. The sentence Claude writes about them is generated, and the failure mode is subtler than a wrong number.

Correlation here is measured against SERP position, where 1 is the goal, so a positive coefficient means smaller values rank better. That inversion is easy to read backwards in prose, and a model that flips it will hand you a confident recommendation pointing the wrong way. The response metadata states the convention on every call, which is worth checking against the summary you were given.

Sample sizes are small by nature -- a first page is ten results, tightened further by outlier removal -- and correlation is not causation at any sample size. The agent also has write access to your repository, so the diff still needs a human read before it ships.