Claude for SEO: what it does well, where it fails, and three setups that work.

1. What Claude is genuinely good at in SEO work.

The useful way to sort SEO work for Claude is by where the facts come from. When you hand it the raw material (a keyword list, a set of pages, your product notes, a template file), it transforms that material quickly and well. When it has to produce the facts itself (how many people search a phrase, where you rank, what a study found), it guesses, and the guess reads exactly like a measurement. Almost every good and bad result people report traces back to that one split. It holds for the other general-purpose assistants too, which is why our ChatGPT for SEO guide reads like a sibling of this one. If you are still choosing which of the two to draft in, our Claude vs ChatGPT for writing comparison runs the same brief through both and calls it by writing job.

On the right side of that split, the list is long:

• Keyword clustering from a list you supply. Paste 200 queries from a Search Console export or an autocomplete sweep and ask for groups one page could rank for. It clusters by intent rather than shared words, so “dog grooming booking app” and “appointment software for groomers” land together even though a spreadsheet filter would split them.
• Content briefs. Given a target query and the pages that currently rank, it produces an H2 set built from the follow-up questions, a list of entities to name, and the angle the ranking pages all miss.
• Drafting and editing. A draft written against a good brief and a style guide is a solid first version. Editing is where it earns its keep: tightening a slow intro, moving the answer to the top, cutting padding, and holding the answer-first structure that LLM SEO depends on.
• Title and meta rewrites. Give it each page's current title, meta description, and the query the page actually gets impressions for, and it rewrites them in bulk to length. For one page at a time, our free SEO title generator and meta description generator do the same job with character and pixel limits built in.
• Schema markup. JSON-LD for Article, FAQPage, Product, or Organization, written from facts already on the page, is quick and usually right. Validate every block before it ships.
• Internal-link maps. Feed it your sitemap URLs with titles plus a new article, and it proposes which pages should link to it, from which sentence, with which anchor. Our internal linking tool builds that map into every draft automatically.
• Technical audits of your own code. Pointed at a Next.js app or a WordPress theme through Claude Code, it reads every template, finds the missing canonical tag or the duplicated H1, and fixes it in place. Setup 2 below covers this.

On model choice, think in tiers rather than version numbers. Use the largest model for drafting, briefs, and any judgment call about intent or quality. Use the smallest fast model for bulk classification, such as tagging 2,000 keywords by intent, where speed and cost matter more than prose. The middle tier handles most editing well. The current lineup by name, with quality, speed and cost per article compared, is in our guide to the best Claude model for writing. If your team drafts in ChatGPT instead, the same breakdown for that lineup is in which ChatGPT model to write with.

2. Where Claude fails at SEO, and what to do about it.

Four failure modes account for nearly every bad Claude SEO result. None of them is a bug the next release fixes. They follow from what a language model is: a system that predicts plausible text from what it was trained on and what you put in front of it. Plan around them and the rest of this guide works. Ignore them and you publish confident mistakes.

It has no real search volume data.

There is no keyword database inside the model. Ask how many people search a phrase each month and you get a number shaped like a measurement, produced the same way as every other sentence. Sometimes it is close. You cannot tell which times. Tell it never to estimate volume or difficulty, and bring measured numbers in from a keyword tool export, your own Search Console data, or a connector (setup 3).

It fabricates statistics and citations unless you make it verify.

Its own documentation is plain about this. Claude's guide to reducing hallucinations says even the most advanced models can generate text that is factually incorrect, and recommends having the model find a supporting quote for each claim and retract any claim it cannot back with one. In SEO drafts this shows up as a tidy percentage attributed to a real-sounding industry report, or a citation to a page that exists but says something else. The fix is a separate verification pass, which is what an automated fact-checking step runs on every draft rather than whenever somebody remembers.

It cannot see your live rankings.

Web search lets Claude read public pages. It does not see your Search Console positions, impressions, or click-through rate, and a search it runs is not a rank check for your market. Ask where you rank and the honest answers are “I do not know” or a guess. Put the numbers in front of it: paste the export, or connect Search Console over MCP.

It is inconsistent across 30 articles a month.

Each conversation only knows what it was given. Across a month of drafts the tone drifts, the structure changes, internal links go to whichever pages happened to be in context, and the banned phrases creep back. That matters beyond style. Google has said on its Search Central blog that its focus is on the quality of content rather than how it is produced, and Google's guidance on using generative AI content warns that generating many pages without adding value for users may violate its spam policy on scaled content abuse. The same guidance asks for accuracy, quality, and relevance in titles, meta descriptions, and structured data as well as the body. Consistency comes from fixed inputs (a Project, saved prompts, a checklist) and a gate that refuses the drafts that fail.

3. Setup 1: Claude chat plus a Project for briefs and drafts.

A Project is a workspace with its own chat history and knowledge base. As Claude's Help Center describes it, you upload documents to the project knowledge and set instructions that apply to every chat inside it. Project limits depend on your plan. For SEO this is the fix for the consistency problem: every brief and every draft starts from the same brand, the same rules, and the same facts.

What to put in it:

• A brand and voice doc. Who you sell to, how you sound, and three paragraphs you consider exactly on-voice.
• Style rules. Banned phrases, heading conventions, how numbers are written, and anything your editor keeps correcting.
• A product facts sheet. Features, pricing, integrations, and what you do not do, so drafts stop inventing features.
• A URL list with titles. The pool for internal links, refreshed monthly from your sitemap.
• An approved sources file. The studies, docs, and datasets you trust, with links, so claims have somewhere true to come from.
• Two or three of your best articles. As reference for depth and tone, not as templates to copy.

Then write project instructions that turn those files into rules. A version that works as a starting point:

Turn web search on for briefs. Per the Help Center, when Claude searches, its responses include citations you can open, which makes the brief checkable. Open them: a brief is only as good as the pages it read. The shape a brief should produce (answer first, H2s from the follow-up questions, one sourced number per claim) is the same shape the blog SEO process applies to every post.

4. Setup 2: Claude Code SEO skills for work inside your codebase.

Claude Code is, in its own documentation's words, “an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools”, available in a terminal, an IDE, a desktop app, and the browser. For SEO it can work on the thing that actually renders your pages: the Next.js layout that sets every title, the WordPress theme file that prints the H1, the component that should be emitting JSON-LD and is not.

Skills make that work repeatable. A skill is a folder with a SKILL.md file inside it: a short header with a name and a description, then plain instructions. Put it in .claude/skills/seo-audit/ inside the repository, or in the skills folder under ~/.claude/ to use it in every project, and according to the Claude Code skills documentation, Claude loads it when a task matches the description, or when you type the skill name as a slash command. The description line is what decides when it fires, so write it as the situation, not the feature.

The SEO skills worth writing are the checks you would otherwise repeat by hand on every template: a unique title and meta description per route, exactly one H1, a self-referencing canonical, no stray noindex, alt text on content images, valid JSON-LD, and every indexable route in the sitemap. Our technical SEO audit checklist is a good source list: turn each line into an instruction. The audit itself can be as simple as this:

Run the fixes on a branch, read the diff, and run the build before merging. A title change in a shared layout changes every page on the site. Source code is also not what Google sees: a route can look right in the template and still render the wrong title once its data loads, so check the rendered pages too, with a crawl or our free SEO audit of the live site.

5. Setup 3: MCP connectors that bring real data in.

The Model Context Protocol is the fix for the missing volume data and the missing rankings. The MCP specification defines hosts (the AI application), clients inside them, and servers that expose tools and data, so Claude can call a keyword source or your Search Console property directly instead of guessing. The architecture and the safe setup sequence are covered properly in our guide to how an SEO MCP server works.

There are two places to connect one. In the Claude apps, remote MCP servers are added as custom connectors, and connector limits depend on your plan. In Claude Code, a single claude mcp add command registers a server, and the project scope writes it to a shared config file your team checks in, so everyone working in the repository gets the same servers.

Connect Search Console first: your own queries, impressions, clicks, and average positions, which turns “where do we rank” into a lookup. Then add one source of measured keyword volume and difficulty. Most Search Console MCP servers are community-built wrappers around Google's public Search Console API, so read the code before you connect one and grant read-only access. The roundup of MCP servers for SEO, compared by job shows which ones are maintained by the company that owns the data.

Treat every server as software with access to your accounts. The Claude Code docs warn that servers which fetch external content expose you to prompt injection, and the connectors article notes that a malicious server can carry hidden instructions aimed at the model. Read-only scopes, one server per job, and no write access until a workflow genuinely needs it.

6. Running these setups every day.

Each setup above works for one article. Run them every day and the manual steps become the job. Here is what The SEO Agent does in place of each one.

• Setup 1, the Project. Instead of a URL list you refresh by hand, it checks the pages already on your site before it plans, so it never duplicates one, and it picks internal links only from URLs that exist.
• Setup 2, Claude Code skills. This one stays with you. Template fixes are codebase work, and Claude Code handles them well.
• Setup 3, MCP connectors. Keyword data is built in: live monthly volume and difficulty for your market, never an estimate.
• The prompts and the fact-check. Every draft gets a fact-check pass with cited sources, then a quality gate that refuses the drafts that fail.
• The paste into your CMS. Articles publish natively to WordPress, Webflow, Shopify, Wix, Ghost, or Framer on a daily cadence.

The automated SEO content pipeline walks through each stage in order.

7. Claude SEO prompts you can copy.

Seven prompts for the jobs above. Each one follows the same four rules, which matter more than the wording: supply the data rather than asking Claude to recall it, forbid invented numbers, fix the output format, and ask for a check before the answer. Replace everything in square brackets. The file names assume the Project from setup 1.

1. Cluster a keyword list by intent.

2. Build a content brief.

3. Draft from the brief.

4. Fact-check a draft.

5. Rewrite titles and meta descriptions.

6. Generate schema markup.

7. Map internal links for a new article.

Prompt 4 is the one never to skip. Run it in a separate chat from the draft, so the check starts from the page itself rather than from the reasoning that produced it.

A worked example: one keyword cluster, from raw list to fact-checked draft.

Take a fictional booking app for mobile dog groomers, Groomly, with a Project set up as in setup 1 and Search Console connected as in setup 3. The business is invented and the output below is trimmed to fit, but the prompts are the ones from the section above, and the shape of what comes back, including the two places it goes wrong, is what you should expect.

Two things in this run are worth noticing. First, the draft read cleanly with all eleven claims in it. Nothing about the prose gave away the three it made up, and only the separate check found them. Second, every correction depended on something from outside the model: the sources file, pages opened with web search, and Search Console. That is the whole pattern. Claude supplies speed and structure, and the facts have to come in from somewhere it can quote.

Common mistakes when using Claude for SEO.

1. Asking Claude for search volume. It answers, confidently, with a number it cannot have measured. Treat any volume it did not read from a data source as a guess, and tell it in the project instructions never to estimate.
2. Publishing the draft the day it was written. A clean draft and a correct draft look identical. The claim check is a separate step with its own prompt, and it is the step that finds the invented statistic.
3. Starting every article in a fresh, empty chat. No brand doc, no style rules, no URL list: every draft reinvents your voice and your product, and the internal links point wherever the model guesses. Put the fixed inputs in a Project once.
4. Letting Claude Code edit templates without a review. Work on a branch, check the rendered page rather than the source, and run the result against a step by step SEO checklist before you merge.
5. Connecting MCP servers you have not read. A connector with write access and an unfamiliar maintainer is a risk to your accounts, and servers that fetch web pages can carry injected instructions. Use read-only scopes and one server per job.
6. Scaling output before the gate exists. Thirty unchecked drafts a month is the pattern Google's scaled content abuse policy describes. Volume is only worth having once every piece clears the same bar.