Claude Code for SEO: My Keyword-to-Page Workflow

Claude Code for SEO is how this blog gets built. I run a set of SEO skills in Claude Code that pick keywords from live search results, write pages that answer the question first, grade every page before it ships, and track whether AI answer engines mention the brand at all.

Claude Code for SEO means using an AI agent for the research and production work around search: reading the real results for each keyword, choosing ones a page can win, writing answer-first pages, grading them before they ship and auditing the site's technical health.

Here is the workflow, step by step, with what Claude Code does and what I still check by hand.

Why Claude Code suits SEO work

SEO is mostly research, writing and checking, repeated across dozens of pages. That is the kind of work Claude Code handles well, because it works inside a folder and keeps everything as files: the keyword list, the cached search results, every page draft, every grade.

The key difference from a chat window is that nothing gets lost between sessions. The keyword list from last month is still there. The page grade from last week is still there. A new session picks up where the last one stopped.

The workflow at a glance

Step 1: Decide what the site should rank for

Before any keyword tool, Claude Code reads the site and writes down what the business sells, who buys, and which searches belong to which of your sites. That last part matters if you run more than one brand: I keep one owner per keyword across my sites so they never compete with each other. On this site, searches about my name, the AI Operator Kit, speaking and Claude Code tutorials belong here; software searches belong to another brand.

The output of this step is a short project file that every later step reads: what the site sells, the money action (book a call, buy the product, join the community), the audience in their own words, and the decisions log. When I come back to a project a month later, that file is how a new session picks up exactly where the last one stopped, without me re-explaining the business.

Step 2: Pick keywords by reading the results

This is where Claude Code changes the work most. Instead of trusting a difficulty score, it pulls the real top results for each candidate keyword and reads them:

• How many of the top five pages target the exact phrase in their title or address?
• Are forums, Reddit threads and directories ranking? That usually means the results are weak.
• Is one big company holding most of the page? Then it is not worth it yet.
• Did Google read the phrase the way I meant it? A keyword can look wide open when the results are about something else.

Each keyword gets a verdict: wide open, winnable, competitive or park. For this batch of articles, it checked 40 candidate keywords in one run. I still read the closest calls myself and overrode several, such as list-article results the script scored too kindly.

Before building, every keyword is also checked against my other sites. Two of my own pages competing for the same search usually means both rank worse, so a keyword gets one owner across the whole portfolio. That check caught several overlaps between this blog and my community site, and the fix was to keep each topic on the site whose readers it fits and link across instead of writing it twice.

Step 3: Write pages that answer first

Every page follows the same rules: the keyword in the title, the main heading and the first sentence; a short answer of under 50 words near the top; headings that read like a table of contents; questions people really ask, taken from the "People also ask" box, answered in a FAQ.

Claude Code drafts from real sources only: my own videos and transcripts, product files and official documentation. It never makes up a statistic. Anything it cannot confirm gets cut or marked for me to check.

A worked example: one keyword, start to finish

Here is the path one of the articles on this blog took.

1. Seed: "claude code for real estate agents", one of 40 candidates generated from my audience and offers.
2. SERP read: the top results were a Medium post, a YouTube video, a Reddit thread and a couple of small blogs. None had the exact phrase in the title or address, so it came back wide open.
3. Portfolio check: none of my other sites owned it, and it shared no more than two top-10 results with any existing page.
4. Brief: the "People also ask" questions became the FAQ: what Claude can do for real estate, whether you can build agents with it.
5. Draft: written from my AI Operator Kit's real workflows, in the second person because I am not a real estate agent, with fair housing rules up front.
6. Checks: quality gate passed, on-page grade 68 out of 70 on my local server, with the last two points only possible on the live domain.

The whole loop, from candidate list to graded draft, is a few hours of Claude Code work and an hour of my review.

Step 4: Grade every page before it ships

Two checks run on every draft:

1. An on-page grader that scores 70 points of measurable placements: keyword in title, heading, first sentence and address, title length, description length, image alt text, internal links, word count. The articles on this blog are graded this way before they go live.
2. A quality gate that flags words and patterns that make copy read like AI wrote it, plus claims that need fact-checking, such as superlatives and prices.

A page has to pass both. When it fails, the fix is usually the first sentence or a missing image description, not more keywords.

Step 5: Keep the site technically healthy

Claude Code runs a crawl of the whole site and checks what search engines see: broken links, missing titles, duplicate descriptions, redirects, structured data, sitemaps and security headers. On this site, that audit found more than 500 posts that only appeared behind a "load more" button, so they were never linked in plain HTML. The fix was topic hub pages that list every post.

Step 6: Measure, including AI search

Two kinds of measurement:

• Search Console: I export the queries and pages reports by hand, and Claude Code analyses them: pages stuck on page two, queries with high impressions and low clicks, pages that lost traffic.
• AI answer engines: a fixed set of questions goes to the major engines every month, and Claude Code logs whether the brand is mentioned or cited. Same questions every time, so the trend means something.

Mistakes to avoid

• Trusting difficulty scores over the results page. A keyword with a scary score can be wide open, and an easy-looking one can be owned by a single big brand.
• Writing to the grader. A page can pass every automated check and still be too thin to rank. Depth comes from real examples and first-hand detail.
• Publishing in bulk without review. Small batches that a person reads beat hundreds of unreviewed pages.
• Letting your own sites compete. Check every keyword against all your domains before building.
• Reworking pages too early. New pages take months to settle. Leave a page alone for about 90 days unless it is not indexed.

What I still do by hand

• Read the closest keyword calls and decide.
• Read every page before it goes live.
• Export Search Console data, since there is no automatic connection in my setup.
• Request indexing for new pages.

Claude Code does the research and the first draft of everything. Judgment on what to publish stays with me.

Claude Code for SEO: FAQ

Is Claude Code good at SEO?

Claude Code is good at the research, writing and auditing work that SEO is made of, especially when it reads real search results and real sources instead of guessing. It still needs a person to judge close calls and approve what gets published.

Can Claude Code do keyword research?

Claude Code can run keyword research when you connect it to a search results source, such as a SERP API. It then reads the top results for each keyword and judges how winnable it is. Without real results it can only suggest phrases, and those need checking.

Will AI-written pages get penalised by Google?

Google's spam policies target large numbers of low-value pages made mainly to rank, whoever writes them. Pages built from real expertise, checked by a person and published in small batches are a different thing. I publish in batches of about five and read every one.

What SEO tasks should I automate first?

Start with the technical audit and page grading, because they are repetitive and easy to check. Then move to keyword research with live results. Writing comes last, once your sources and quality checks are in place.

How long should a blog post be for SEO?

An informational article that answers one main question usually needs 1,200 to 1,500 focused words, while a bottom-of-funnel landing page can be 400 to 500. Match what the top results reward, and add depth with examples rather than padding.

Can Claude Code fix technical SEO problems?

Claude Code can find and fix many technical SEO problems in a site it can edit: missing titles and descriptions, broken links, sitemaps, structured data and security headers. Check each fix on a preview before it goes live.

Copy the writing side of this setup

The AI Operator Kit includes a local and AI search play: Google Business Profile fixes, review replies, one service page a month and a monthly log of what AI engines say about you. For building the site itself, see how to build a website with Claude Code.