How to Use Claude for SEO: A Practical Weekly Workflow (2026)

To use Claude for SEO, give it your own data and one job per day: Search Console quick wins on Monday, a technical check on Tuesday, briefs on Wednesday, on-page and schema fixes on Thursday, and an AI-answer check on Friday. Claude analyzes and drafts; your tools supply the numbers, and you approve what ships.

This is the routine as of October 2026. Each step names the Claude surface to use, the input, a copy-paste prompt and what to check before you act, with links to the deeper version.

What does a weekly Claude SEO workflow look like?

Five short sessions, one job each, all inside one Claude Project for the site. The table is the whole week; the sections after it give the prompt and the checks for each day.

Which Claude mode should you use for SEO?

Pick the surface by the job. As of October 2026, four matter for SEO work:

Projects keep instructions and files across chats, so you describe the site once instead of every Monday. Cowork takes on multi-step tasks across your local files and can return a finished spreadsheet or report. Claude Code runs in the terminal, IDEs, a desktop app and the browser, and suits teams whose site templates live in git.

How do you set up Claude for weekly SEO work?

Create one Project per site. Add the sitemap, a list of money pages, your style rules and the latest exports as knowledge files. Anthropic's prompting guide describes Claude as a brilliant but new employee who lacks context on your norms; the project instructions are where you supply it:

You are the SEO analyst for [site], a [type of business]
selling to [audience] in [markets].
Money pages: [URLs]. Main competitors: [2 or 3 domains].
Rules:
1. Work only from the files I attach, the pages I link and
   the data in this project. If a number is not in the
   data, write "not in data". Never estimate rankings,
   search volume or backlinks.
2. Tables first, then at most 5 actions ranked by expected
   impact, each with the URL and the exact change.
3. Mark anything I must verify in Search Console, a crawler
   or a validator before it ships.
4. If an input is missing, ask one question before starting.

Then add the 8 free Claude SEO skills: technical audit, keyword clustering, content brief, schema markup, internal linking, Search Console quick wins, GEO citation audit and content gap. A skill is a SKILL.md file that Claude loads only when a task needs it, so each weekly job runs the same way without a long prompt. Anthropic lists skills on every plan, with code execution switched on; in Claude Code they install into your skills folder and run as commands such as /seo-gsc-quick-wins.

Monday: which Search Console queries are closest to page one?

Task: find pages ranking just off page one where one change could move them. Where: a chat inside the site's Project. Input: the Search Console Performance report for the last 90 days, exported with queries and pages. If you would rather not export every week, connect Google Search Console to Claude through MCP so Claude queries it directly.

Attached: my Search Console Performance export for the
last 90 days, with queries and pages.
Find the 10 best quick wins: pages with an average position
between 8 and 20 and more than 50 impressions. For each,
give the page, the query with the most impressions, its
position and CTR, the one change most likely to move it
onto page one (title, H2, new section, FAQ, schema or
internal link) and your confidence (High/Medium/Low).
Rank them by expected impact.
Skip these pages, changed in the last 4 weeks: [URLs].

Check: Claude has no view of Google's live results, so every position in its answer must come from your export. Search Console data arrives about two days late, so do not read the last two days as a drop. Daily positions, alerts and competitor rankings still need a rank tracker; what Claude does well is analyze the positions Search Console already reports, as the 60-day test of Claude as a rank tracker shows task by task.

Tuesday: what changed technically since last week?

Task: catch regressions before they cost traffic: new errors, noindex tags, canonical changes, missing titles and orphan pages. Where: Project chat for a small site; Cowork for a folder of exports; Claude Code if the site is in a repo, where it can also find the template behind a problem. Input: this week's and last week's crawl exports, or the HTML of a key page. For a quick first pass on one URL, the free 31-check SEO and AI audit runs in the browser.

Attached: this week's crawl export and last week's.
List every change on indexable URLs: new 4xx/5xx errors or
redirects, pages that became noindex or canonicalized
elsewhere, titles, meta descriptions or H1s that changed or
disappeared, new orphan pages, and indexable pages missing
from the XML sitemap.
Table: URL | Change | Severity (High/Medium/Low) | Fix.
High first. List thin pages separately for me to review;
do not mark them as errors.

Check: re-crawl a sample of flagged URLs and confirm noindex and orphan findings in Search Console's page indexing report before changing anything. Do not merge or delete short pages on Claude's word alone: a short page can be the right answer to a transactional query. The full version of this step is Workflow 1 in the Claude SEO audit playbook.

Wednesday: what should you write or refresh next?

Task: choose one page to create or refresh, and brief it. Where: Project chat; Claude can read the competitor URLs you give it. Input: a keyword export from Search Console or your keyword tool, the target query, your page and the pages that rank for it in Google.

Target query: [query]. My page: [URL].
These pages rank for it in Google: [3 to 5 URLs].
Read them and tell me what they all cover that my page does
not, which questions none of them answers well, and the
format searchers seem to expect (guide, list, comparison,
tool). Then write a refresh brief: a 40-60 word answer-first
opening, H2s phrased as questions, sections to add or cut,
3 internal links from the sitemap in this project, and the
sources to cite. Use only the keyword data I attached; no
volume estimates.

Check: Claude's web search is not Google's results page, so pick the competitor URLs yourself from a Google search. Search volume and difficulty must come from your keyword tool. Before writing something new, look in Search Console for a page of yours that already earns impressions for the query, so two pages do not compete for it. Prompts 20 to 25 in 47 Claude AI SEO prompts cover the rest of the content work.

Thursday: which on-page, schema and internal-link fixes ship this week?

Task: turn Monday's and Wednesday's findings into changes for three to five pages. Where: Project chat; Claude Code if your templates live in a repo, so changes come back as diffs you can review. Input: each page's HTML from view-source (what crawlers get before JavaScript runs), its Search Console queries and your sitemap.

Attached: the HTML of [URL] and its Search Console queries.
1. Write 3 title options under 60 characters and 2 meta
   descriptions under 155, using the words in the queries.
2. Audit the JSON-LD: list errors and missing recommended
   properties, then give a corrected block that describes
   only content visible on the page.
3. From the sitemap in this project, pick 3 pages that
   should link here, with a natural anchor for each.

Check: validate every JSON-LD block in Google's Rich Results Test and the Schema.org validator, and mark up only what visitors can see. Add internal links over several weeks, not dozens at once, and only where the link reads naturally. Log the date of every title change and judge its effect on CTR in Search Console after about four weeks.

Friday: do AI engines mention you for your buyer prompts?

Task: the GEO step: see whether AI assistants recommend or cite you for the questions your buyers ask, and which pages they cite instead. Where: an incognito chat with web search, outside your SEO Project, so Claude's memory and your project instructions do not tell it who you are. Input: five to ten prompts phrased the way buyers ask, not the way you would phrase keywords.

I'm researching [category] for a [type of company] in
[country]. Which [products / consultants / tools] would you
recommend, and why? Please cite the sources you used.

Run each prompt in three fresh chats and log whether you are named, where you appear and which URLs are cited. Then paste the answers into your Project and run the seo-geo-citation-audit skill on the page that should have been cited. By hand, monthly is enough; on a weekly Friday, review a tracker's results instead.

Check: one chat is a sample, and Claude cannot see what ChatGPT, Gemini or Perplexity answer; for those you need a tracker or an API script (nine AI visibility trackers compared). My own tracker has run 10 B2B buyer prompts on ChatGPT, Claude, Gemini and Perplexity weekly since 24 September 2026, at about $1.96 to $2.14 a run. Of the latest 80 answers, 10 used no web search, so nothing could be cited in them, and 9 of those 10 came from the two prompts about using Claude for SEO. For "best X consultants" prompts, 81 to 96% of cited URLs were third-party lists. So for how-to questions you compete with what the model already knows; for "best" questions the route runs through the lists the engines cite. Measuring and acting on that for your category is what GEO consulting covers.

What runs monthly and quarterly instead of weekly?

• Monthly: compare the latest 90 days with the previous 90 in Search Console; pages where impressions grew but clicks did not are often queries that AI Overviews or AI Mode now answer. Run the manual AI-answer check and refresh the Project's knowledge files.
• Quarterly: the full five-workflow audit from the playbook, a competitor gap with fresh exports, and a review of your buyer prompts against Search Console queries and sales calls.
• Automation, last: on paid plans a recurring task can run the Friday prompts on a schedule in the cloud. Automate a step only after it works by hand, and keep a person on the review.

What should you not hand to Claude?

• Numbers it cannot see. Live rankings, search volume, keyword difficulty, backlink counts and Core Web Vitals come from Search Console, your keyword and link tools, and PageSpeed Insights. A figure Claude gives you without your data is a guess.
• Other engines' answers. Claude can show its own answer to a prompt, not ChatGPT's or Perplexity's.
• Final decisions. Claude does not know which pages make money, what you published last month or what buyers tell your sales team unless you say so. It proposes; you decide what ships.
• Experience and evidence. Drafts need your first-hand results, examples and sources. Claude can structure and edit them, but it cannot supply the experience E-E-A-T is about, and it can state wrong facts fluently.
• Credentials and unvetted code. Keep service-account keys in local config files, never in a chat. Install only MCP servers and skills you have read: Anthropic does not verify custom connectors, and a skill can direct Claude to run code.
• Bulk changes in one go. Mass rewrites and dozens of new internal links in one week are hard to measure and harder to undo. Ship in batches and log the dates.