The short answerClaude is a strong SEO research and drafting assistant. It clusters keywords by intent, summarises top-ranking pages, and drafts outlines and meta copy. Without a connected data source it has no volume or ranking data, and it can report wrong numbers or build pages that should not exist. Keep a human on every change to a live URL.
"Claude SEO" means two different things in search results right now. One is a set of open-source plugin bundles that run SEO audits inside Claude Code. The other is a practical question: which SEO jobs can you hand to Claude, and which ones will it quietly get wrong? For the second, the answer depends on the task.
Claude for SEO, task by task
| SEO task | Which part of Claude | What you give it | What comes back | What a human still owns |
|---|---|---|---|---|
| Keyword clustering and intent | Chat or a Project | A keyword export from your SEO tool, with volumes | Clusters grouped by search intent, with a suggested page per cluster | Which clusters deserve a page, and whether an existing page already covers one |
| SERP and competitor summary | Chat with web search, or Claude Code with a SERP data connection | The query and the URLs ranking for it | What the top pages share and what they leave out | Whether the gap is worth a page at all |
| Content brief and outline | A skill | The target query, your SERP notes, your own expertise | A first-pass brief and outline in your house structure | The angle and the first-party proof |
| Title tags, meta descriptions, alt text | A skill | The page copy and your rules for length and voice | Draft metadata for each page | A review before anything is published |
| Schema markup | Chat or Claude Code | The page and the schema type you want | Draft JSON-LD | Validation, and a check that the markup matches the visible page |
| Search Console, GA4 and Ads analysis | Claude Code with data fetchers or an MCP connection | API access or JSON exports | Answers to cross-source questions, such as where paid search is your only presence | A spot check of every number against the source data |
| Technical audit summary and hreflang | Claude Code | Crawl exports, your URL list, your locale map | A summary of issues and a draft hreflang XML sitemap | Every redirect, canonical tag and internal-link change |
| Content refresh | Claude Code with a Search Console connection | The live page plus its query data | What is stale on the page and a revised draft | What goes live, and when |
Read the last column first. Every row hands Claude the work that produces a draft or an answer, and every row keeps the decision that changes your site with a person. That split explains why some teams get real value from Claude while others end up with cloned pages nobody visits.
What "Claude for SEO" actually covers: four parts of one product
Anthropic sells chat apps, a coding agent and an API, and SEO work touches all of them. They behave differently, so pick by the job.
Chat (Claude.ai and the desktop and mobile apps). The window most people start in. With web search on, Claude searches when it helps and attaches citations to what it finds, so you can check the source yourself. It's fine for one-off jobs, like clustering a keyword list you paste in. The cost shows up on the tenth run, when you are re-explaining your brand and your formatting rules for the tenth time.
Projects. Anthropic describes Projects as "self-contained workspaces with their own chat histories and knowledge bases." You upload documents (a style guide, say, or a list of your product pages) and set instructions once. For an SEO lead working mostly in chat, a Project per site or per client is the cheapest upgrade available.
Skills. A skill is a folder holding a SKILL.md file with instructions, plus optional scripts and resources, that Claude loads only when the task calls for it. Anthropic's developer docs explain the loading in three levels: a short description is always visible, the body loads when the skill triggers, and bundled files load only when needed. Skills work across the Claude apps, Claude Code and the API, and Anthropic published the format as an open standard in December 2025. Ahrefs puts the practical value well: a skill "turns Claude from a chat window you keep re-explaining yourself to, into something you train."
Claude Code plus connectors. Claude Code is Anthropic's agentic coding tool. It reads and edits files and runs commands, in the terminal, an IDE, a desktop app or the web. It reads a CLAUDE.md file at the start of every session, which is where your standing rules live. It can also connect to outside services over MCP. The Model Context Protocol is the open standard Anthropic released in November 2024 for connecting AI applications to outside systems, and through it Claude can pull data from services like Ahrefs or DataForSEO instead of guessing. You don't need to be a developer to use it for SEO. You do need to be comfortable reading what it did before you accept it.
If you are deciding whether to buy an SEO agent product instead of building on Claude, that's a different question, and our guide to what an SEO agent can and can't do for B2B SaaS covers it. If you've already chosen Claude, the job is making it do good work.
Where Claude does good SEO work
The practitioners writing honestly about this agree on the shape. Will Scott, CEO of Search Influence, wrote in Search Engine Land that "Claude is one of the best research assistants an SEO can have." His list of good jobs: keyword clustering, search intent classification, content gap analysis against competitors, technical audit summaries, first-draft outlines, and copy.
Three patterns stand out from the first-hand reports.
Analysis across data sources you'd never join by hand. In an earlier Search Engine Land walkthrough, Scott had Claude Code write Python fetchers that pull Search Console, GA4 and Google Ads data into JSON, then asked it questions across all three. He says the setup takes about an hour, and that he "never had to read the API documentation." On a higher-education client, a paid-versus-organic gap analysis found 41 content gaps where paid search was the only presence. The analysis took about 90 seconds. Questions you'd never have budgeted an analyst afternoon for become cheap enough to ask every week.
Repetitive technical builds with a clear right answer. Elmer Boutin of Perry Johnson described five ways he uses Claude, including building an hreflang XML sitemap. His verdict on that one: "I double-checked the output and found it to be excellent on the first go." He also used it to translate and localise content, with quality "on par with what we were getting with Google Translate, as verified by our in-country experts." Note who verified it.
First drafts in your own structure. Briefs and outlines have a known shape. So do meta descriptions and schema. Give Claude the shape once (in a Project or a skill) and it fills it fast. Skills pay off here, because the structure is written down once instead of retyped.
What these share: real data or a clear spec goes in and a draft comes out. A person checks it before it matters.
Where Claude breaks
This is the part that costs money.
It acts with confidence on bad calls. Scott's later article is built around a failure. He asked Claude to recommend keywords from Search Console data and build pages for them. Claude "cloned the homepage into two new URLs, /seo-grader and /content-grader, changed the title tag and H1 to match the new keywords, and reused most of the homepage's body copy." Six months later, those pages had earned zero impressions and zero clicks. It happened again on a second, unrelated site. In his words: "Claude did it anyway, with total confidence, no caveat, and no flag that the 'new' page was actually a clone." He is clear that this is not a Claude-only bug: "This is a category problem with autonomous AI execution, not a single vendor's bug."
It can misreport your own numbers. Scott writes that he has seen Claude Code "confidently report a number that didn't match the JSON file," and adds that it is rare. Anthropic says the same about its own model: Claude "can sometimes generate text that is factually incorrect or inconsistent with the given context." Anthropic's hallucination guidance tells you to let Claude say it doesn't know, to make it pull direct quotes before it analyses a long document, and to make it retract any claim it can't back with a quote.
It has no SEO data of its own. Claude's web search returns pages with citations. None of Anthropic's product pages I read describe a keyword volume, ranking or backlink dataset inside Claude. Ask a plain chat window for search volume and you get a plausible-sounding guess. That data has to come from a vendor through a connection, which is exactly why Ahrefs wired its MCP into a skill pipeline, in its words, "so the skills could pull real data instead of hallucinating it."
It doesn't know when it doesn't know. "An AI model has no internal signal that tells it, 'I don't actually have a good answer to this,'" Scott writes. A prompt can't fix that. Only a check can.
So the line to draw is simple. Scott's rule: "Keep AI in the research and drafting seat, and keep a human responsible for what goes live." He lists the execution items that need a person: new page creation, title tag and H1 changes, redirects, canonical tags, internal linking changes, and anything that touches a live URL.
How to set Claude up for SEO work
Most of the gap between a toy and a tool is setup. Do these in order.
1. Connect real data before you ask a data question. Ahrefs hosts an MCP server for Lite plans and above, with setup guides for Claude on the web, desktop, mobile and Claude Code. Calls through it use API units the same way direct API calls do. DataForSEO publishes an MCP server too: in Claude Desktop you add a custom connector, paste its MCP URL and approve the authorisation, and Claude asks for your approval the first time it calls a tool. For Search Console and GA4, Scott's route (Claude Code writes the fetchers, the data lands as JSON) works without any vendor at all. Anthropic's own warning applies: "Only connect services you trust and need for your workflows."
2. Write your standing rules down once. In chat, that is a Project with instructions and a knowledge base. In Claude Code, it's CLAUDE.md. Put your brand voice and house structure in there, plus a list of the pages that already exist, so Claude stops proposing a page you already have.
3. Turn each repeat task into a skill. The frontmatter carries the skill's name and its description, and the description decides whether it fires. Ahrefs shows that a vague description ("Helps with writing") never fires, and one with real trigger phrases fires reliably. A minimal example:
---
name: meta-description-rewrite
description: Rewrite meta descriptions for existing blog posts. Use when the user asks to fix, shorten or rewrite meta descriptions, page descriptions or SERP snippets.
---
1. Read the page's current title, H1 and first section.
2. Write a meta description that states what the page answers, in the brand voice from CLAUDE.md.
3. Return the old and new versions side by side. Never publish.For any skill with consequences, such as publishing to a CMS or sending a message, Ahrefs recommends disable-model-invocation: true so Claude can't fire it on its own.
Ahrefs is also blunt about where the effort sits: "The markdown takes ten minutes." The expertise behind it takes the rest of your career, and wiring it to live SERP data, a CMS and a style guide is a separate project on top.
A worked example: one content refresh, from data to approval
Take one job: a refresh of a page that is slipping in Search Console. You can run it yourself with the setup above.
- Pull the evidence. Claude Code fetches the page's Search Console queries and positions, plus the live page. You ask: which queries is this page getting impressions for that the copy never answers?
- Check the SERP. Through a SERP data connection, Claude summarises what the current top results cover that your page doesn't. You confirm the gap is real by opening two of those pages yourself.
- Draft the change. A refresh skill rewrites the stale sections in your house structure and voice, and lists every factual claim it added with its source.
- Check the claims. Every number in the new copy gets matched to its source. Anything Claude can't source comes out. Treat any reviewer's "suggested improvement" the same way: when it asserts a new fact about someone, it's a claim to check before it goes in.
- Approve and ship. A person reads the final page and decides whether it goes live. Title tag and H1 changes get the same review as the body, because Google applies its review advice to metadata too.
The model does most of the typing in steps 1 to 3. The steps that protect you (4 and 5) are the ones a person or a deliberate check owns.
What Google says about content written with Claude
Google doesn't penalise content for being AI-assisted. Its 2023 guidance says: "Appropriate use of AI or automation is not against our guidelines." What it targets is the purpose. Its spam policies define scaled content abuse as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users," and the first example listed is using generative AI tools to generate many pages without adding value for users.
Google's guidance on generative AI content, updated in October 2026, is more direct than most vendors: "Keep in mind that generative models don't retrieve facts, but predict a likely sequence of words based on their training data." And: "It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing." That review covers title elements, meta descriptions, structured data and image alt text as well as the body. Google also says generative AI can be useful when researching a topic and for adding structure to original content.
Put Scott's cloned pages next to that policy. Two near-duplicate pages, built for a keyword, adding nothing for a user. Claude didn't break a rule anyone told it about. Nobody told it.
Getting your own site cited by Claude
Using Claude for SEO and showing up in Claude's answers are separate problems, and the second one starts with Anthropic's crawlers. Anthropic's crawler page lists three bots:
- ClaudeBot collects content that could contribute to model training. Blocking it signals that your future content should be excluded from training data.
- Claude-User fetches pages when a person asks Claude a question. Blocking it may reduce your visibility for user-directed web search.
- Claude-SearchBot crawls to improve search result quality. Blocking it may reduce your visibility and accuracy in user search results.
All three honour robots.txt, set per subdomain, and Anthropic says blocking its IP addresses may not reliably work as an opt-out. My reading of that page (Anthropic doesn't say this outright): a B2B SaaS site that wants to stay out of training can block ClaudeBot and still allow Claude-SearchBot and Claude-User, the two bots Anthropic ties to search visibility.
When Claude answers with web results, it cites them, and every citation carries the source URL. That makes the usual AI search work apply here too: clear answers near the top of the page, and facts that can be lifted in one sentence from pages the bots can reach. You can see which pages these bots actually request in your server logs, which beats guessing. On loudface.co, in the seven days to 2 October 2026, ClaudeBot read robots.txt 338 times and the sitemap 205 times, and Claude-User fetched robots.txt 132 times. ClaudeBot made no llms.txt request that week and Claude-User made 2, so read what that file does and who reads it before you build one. For how the other engines pick sources, see where ChatGPT gets its information and how Perplexity works.
The system around Claude matters more than the model
Every first-hand report above lands in the same place. Claude is a strong component. A chat window isn't a system.
Our position, from running AI-written content for our own site and for clients: AI-written content can beat human-written content, but not from a chat window with Claude or ChatGPT. It takes a system with rich data, thought leadership, a knowledge base and a well-built loop working together. And it still benefits from a person reviewing the high-level direction and the quality. Take that person out and performance drops meaningfully. Anyone telling you humans aren't needed is half right. Very few people know how to build and run a system that produces high-performing content.
The outside evidence points the same way. Ahrefs puts the work in the expertise and the data wiring. Scott and Google want a person checking before anything publishes, and Anthropic wants Claude to retract what it can't quote.
Judge any Claude setup by what it moves for the business (signups and booked demos) rather than by how many pages it produces. Output volume is the easiest number to grow and the least useful one to report. Scott's description of AI citation data fits every AI metric: "Wind sock, not GPS."
If you'd rather have a team run that system for you, LoudFace runs SEO and AI search programs for B2B SaaS, and engagements start from $5k a month. If you're building it yourself, start with one skill, one real data connection and one person who signs off on anything that touches a live URL. Then add the next one.
