Claude SEO: Concrete Use Cases, Copy-Paste Prompts and Limits

I use Claude every week in my SEO client work. Three things matter more than the choice of model: what you give it to read, how you phrase the request, and what you forbid it to do. Here are the use cases that hold up, each with the prompt I use, followed by the limits.

How to use Claude for SEO

The rule that decides everything fits in one sentence: give it the data, not the question.

Ask a model for the search volume of a keyword and it will hand you a plausible, wrong number. Give it the export from your keyword tool and ask it to spot the commercial-intent queries your site covers poorly, and it will do work that would have taken you two hours.

That distinction explains almost every failure I see. People use these models as search engines when they are reasoning engines. The value is not in what they know, it is in what they do with what you give them.

In practice, you can use Claude for SEO at three levels:

• In the chat (claude.ai or the app): you attach a CSV export, the text of a page, a list of URLs. This is where most of the use cases below start.
• Connected to your tools through the MCP protocol: Claude queries the crawler, Search Console or your keyword tool itself, with no copy and paste.
• In Claude Code: the agent works on the site's files and applies the fixes, from two hundred title tags to a redirects file.

8 Claude SEO use cases, with the prompt to copy

Every prompt follows the same structure: a role, numbered work, rules, then your data pasted underneath. Replace the brackets and attach the file when the prompt asks for it.

1. Auditing a site from a crawl

This is the use case I recommend first: tedious for a human, fully verifiable, no editorial risk. I run a crawl, export the result and ask for a three-level ranking. That hierarchy matters more than the list of issues: a tool report shows a hundred alerts with the same visual weight, when three of them explain most of the lost revenue.

Prompt: technical audit

You are a technical SEO consultant. I am giving you the export of a crawl of [site].
1. Sort the issues into three levels: blocking indexing, harmful, cosmetic.
2. For each blocking issue, give the number of URLs affected and the exact fix.
3. Finish with the three workstreams to tackle first, ranked by traffic impact, not by ease.
Rules: no action whose effect cannot be measured. If you are missing data to decide, say so instead of assuming.
[Attach the crawl export]

The final decision cannot be delegated. Whether a page deserves to be fixed, rebuilt or deleted depends on its commercial role, your resources and your timeline. Claude prepares the decision, it does not make it.

2. Working through a Search Console export

Pages that are crawled but get no impressions almost always tell a story: content that is too thin, misunderstood intent, or cannibalization with another page. Queries between positions 5 and 15, for their part, are the closest wins. It is cross-checking work that Claude does well and fast.

Prompt: Search Console analysis

You are an SEO consultant. I am giving you a Search Console export (pages and queries, last three months).
1. List the queries in positions 5 to 15 with many impressions and few clicks.
2. Spot the pages whose click-through rate is clearly below the average for their position, with a hypothesis on the cause.
3. Flag the queries for which two pages of the site share the impressions.
Rules: back every line with a figure from the export. Do not invent any data.
[Attach the export]

3. Clustering keywords by intent

Claude does not know search volumes: they come from your tool (Ahrefs, Semrush, SE Ranking, Google Ads Keyword Planner). On the other hand, it is very good at grouping a list of several hundred queries by intent and by target page, which takes hours by hand.

Prompt: keyword clustering

You are an SEO consultant. Here is a keyword export with volume and difficulty, and the list of my pages.
1. Group the keywords by intent: informational, comparison, transactional, local.
2. Within each group, gather the queries a single page can target.
3. For each group, name the existing page that should carry it, or the page to create.
Rules: copy the volumes from the export without changing them. Do not add any keyword or any figure.
[Attach the export and the list of pages]

4. Writing an article brief

When it comes to writing, Claude produces a solid outline and structure. That is exactly what content that fails to rank lacks most often. Give it the headings of the pages at the top of Google: it quickly sees what they all cover and what none of them covers.

Prompt: content brief

You are an SEO content lead. Target query: [keyword]. Here are the titles and subheadings of the top five Google results for this query.
1. Give the dominant search intent in one sentence.
2. Suggest an H2 and H3 outline where each heading contains the question it answers.
3. List the questions to cover in an FAQ, phrased the way a searcher types them.
4. Point out the angle the top pages are missing.
Rules: no filler sections.
[Paste the headings]

5. Writing title tags and meta descriptions

Prompt: tags

You are an SEO consultant. Here is the content of my page and its main query: [keyword].
Suggest three title and meta description pairs:
- a title of 60 characters maximum, starting with the query;
- a meta description of 120 to 155 characters, that gives a reason to click rather than a summary.
Rules: show the character count of each option. No superlatives, no exclamation marks.
[Paste the content]

Count the characters again yourself: language models are bad at counting characters.

6. Building internal links

Prompt: internal linking

You are an SEO consultant. I am giving you the list of my pages with their URL, title and topic.
1. For each page, suggest the three most relevant incoming internal links, with the exact anchor text.
2. List the orphan pages, which receive no links.
3. Flag the pairs of pages at risk of cannibalizing each other, and which one to keep.
Rules: varied, descriptive anchors, never "click here" or "learn more".
[Paste the list]

7. Generating structured data

Article, product, local business, breadcrumb: Claude writes clean JSON-LD in a few seconds. The risk is that it fills the gaps with plausible information, hence the final rule. Always validate the result with Google's Rich Results Test.

Prompt: structured data

You are a technical SEO consultant. Here is the content of a page from a [type of site].
1. Name the most suitable schema.org type and explain why.
2. Write the complete JSON-LD block, ready to paste.
3. List the properties you could not fill in for lack of information.
Rules: only use information visible on the page. No invented ratings, reviews or prices.
[Paste the content]

8. Reviewing before publication and making content citable

Last use case, and not the least: the review. Claude spots unsourced claims, unkept promises and paragraphs too long to be picked up by an answer engine. It is also the starting point for getting cited by ChatGPT or Google's AI Overviews, which I cover in detail in my article on AI for SEO.

Prompt: review

You are a proofreader. Check the text below before publication.
1. Unverifiable or unsourced claims.
2. Repetitions and sentences that say nothing.
3. Promises in the introduction that are not kept in the body.
4. The opening paragraph: does it answer the question within the first two sentences?
Rules: do not rewrite the text. Quote the passage and suggest a short fix.
[Paste the text]

To build a custom prompt from one of these tasks, paste your context into the generator: it assembles the role, the expected work, the rules and your data.

Connecting Claude to your SEO tools: MCP, Claude Code and skills

Copy and paste works, but it gets tiring fast. The real shift is connecting the tools to the model, which the MCP protocol makes possible. If the concept is new to you, I explain it in what is an MCP server.

• A crawler. Claude launches the crawl, reads the results and works on them without you exporting anything. The step-by-step setup is in my Screaming Frog MCP tutorial.
• Search Console. Impressions, clicks, positions and pages, queryable directly in the conversation, through MCP servers that are often built by the community.
• A keyword tool. Ahrefs, for example, offers an official MCP server, starting from its Lite plan: volumes and difficulty come from the real tool, not from the model's memory. In Claude Code, one command is enough: claude mcp add ahrefs --transport http https://api.ahrefs.com/mcp/mcp.
• Your files. Giving access to the site folder lets Claude read the actual code rather than your description of it.

With Claude Code, the next step becomes possible: the model no longer just analyzes, it makes changes. Fixing two hundred title tags in a repository, generating the structured data for a catalog, producing a redirects file from a list of old URLs. These are tasks where the human added value is zero and the risk of typing errors is high. To get started, follow my tutorial to install Claude Code.

Finally, the procedures you repeat (your audit checklist, your brief format) can become skills: instruction files that Claude loads when the task calls for them, instead of explaining everything again in every session. The how-to, with examples of SEO skills, is in Claude Skills.

How Claude improves your SEO

Three effects add up, and none of them is magic.

Speed of analysis. What took half a day now takes an hour. On an audit, the difference is not the quality of the diagnosis, it is the number of sites I can handle properly in a month.

Consistency. A monthly check of internal links or tags never happens when it costs three hours. It happens when it costs twenty minutes. This is the most underestimated gain: what pays off is not the tasks done faster, it is the ones that finally get done.

Quality of structure. A solid outline, headings that announce their content, answers placed at the start of sections: that is what pages that fail to rank are missing, and what makes a text quotable by an answer engine.

Best practices for reliable SEO prompts

Five rules, learned by getting things wrong.

Impose an output format. A prompt that does not say how to answer gets an essay back. Ask for a table, a ranked list, a precise number of options.

Explicitly forbid invention. The sentence "if you are missing data to decide, say so instead of assuming" removes most false claims. It works because it gives the model a way out.

Work on real material. The content of your actual pages, your actual figures. A generic example produces generic advice.

Give the business context. Who buys, at what price, with what decision cycle. The same keyword does not have the same value for a tradesperson and for a software publisher, and the model cannot guess that.

Review what will be published. In an audit, an error shows up when you implement it. In a text published under your name, it commits you. I never publish content I have not reviewed line by line.

What separates a prompt that works from one that disappoints is almost never its length. It is the presence of negative rules: saying what you do not want works better than piling up details about what you do want. The same prompts also work in ChatGPT, and I have gathered more of them in my article on ChatGPT prompts.

What Claude cannot do for SEO

The list of failures is more instructive than the list of successes, so here it is, unfiltered.

Market data. Search volumes, keyword difficulty, a competitor's ranking: all of that comes from a tool, never from a model. The answer will look right, it will be wrong, and that is the worst possible scenario.

A publishable article in one go. A text generated in a single block gives itself away with three signs: it announces what it is going to say instead of saying it, it balances its sections artificially, and it takes no risks. Fixing it takes longer than writing from a solid outline.

Prioritization judgment. Which page deserves three days of work? That depends on your margin, your seasonality and your production capacity. The model knows none of this and will produce a ranking that looks rational but is disconnected from your reality.

Keeping up with news. A model has a knowledge cutoff and SEO moves faster than that. For what changed last month, you have to give it the sources, or go and read them yourself.

How I organize my week with Claude

In practice, my use falls into three moments, and this routine is worth more than a long theory.

Monday, monitoring. Twenty minutes per site I follow: changes in impressions and clicks, new crawl errors, pages that are slipping. Claude reads the exports and flags what has moved. This is the check that did not exist before, for lack of time.

Midweek, production. Briefs, outlines, reviews. I never write the first draft from nothing, and I never publish an unreviewed draft. The model sits in the middle of that chain, not at either end.

End of the month, measurement. The panel of questions put to answer engines, to find out who gets cited on my clients' topics. It is manual, a bit slow, and it is the only reliable data I have in that area.

The benefits for an SEO consultant

I will be precise rather than enthusiastic, because the topic is saturated with promises.

The time saved is real and sits between three and eight hours a week in my case, mostly on analysis and formatting. It is not where people expect it: full article writing remains slow, because the time shifts from writing to reviewing and checking.

The quality gain lies elsewhere. The thankless tasks, checking internal links, keeping tags consistent, verifying structured data, are now done systematically instead of whenever there is time left. Over a year, that gap weighs more than any one-off speed gain.

And the risk, since it has to be mentioned: ease of production pushes you to publish more. That is the opposite of what you should do on a web saturated with average content. The tool is there to do better, not to do more. If you would rather delegate the full diagnosis, that is what I do in an SEO audit, with these methods.

Written by

Anthony Courtin

SEO, GEO & AI automation consultant

I help small businesses and independents with their visibility (Google, ChatGPT, Perplexity) and with the real-world use of AI at work. I work on-site with companies to train teams and rethink processes. Everything I write here, I use in client work.

Follow me on LinkedIn My YouTube channel (in French)