AI Search and GEO · How to
How to use Claude for SEO: 6 workflows with prompts and guardrails
Part of the guide Generative engine optimization (GEO): a practical guide for 2026
Use Claude for SEO work that needs careful reading and structured output: analysing a Search Console export, building a fan-out query list, turning competitor pages into a brief, rewriting sections answer-first, and drafting schema or robots.txt rules for you to check. Keep every number tied to your own data, verify claims before publishing, and set clear rules before letting Claude Code edit your site.
What Claude is good at in SEO, and what it is not
Claude is strong at reading long material you give it and returning structured, careful output. That suits SEO tasks built on your own data: exports, page copy, competitor pages you paste in. It is weak where SEO needs live market data it does not have, such as search volumes, current rankings or backlink counts. A model asked for a keyword's volume will produce a plausible number, and a plausible number is worse than a blank. Why these tasks matter for AI search is explained in the GEO guide.
One rule covers most of this page: numbers come from your data source, never from the model.
Workflow 1: Find quick wins in a Search Console export
Export the Queries and Pages tables from Search Console's Performance report for the last 3 months. Paste the rows (or attach the file if your Claude plan supports uploads) with this prompt:
Here is a Search Console export: query, page, clicks, impressions, CTR, position.
1. List queries with position between 8 and 20 and at least 50 impressions,
grouped by page. These are striking-distance queries.
2. List queries in the top 10 with CTR below 2%.
3. For each page with three or more striking-distance queries, suggest which
H2 sections to add or rewrite. Do not invent any numbers; use only the data given.
Return tables.
Check the output against the export before acting. Claude is reliable at filtering and grouping, but spot-check three rows: if one number is off, recheck the lot.
Workflow 2: Build a fan-out query list
AI engines split a prompt into shorter searches, so you want to rank for those. Ask Claude to generate candidates, then verify them yourself:
I sell [product] to [audience]. Here are 5 questions buyers ask me: [paste].
For each, list the short web searches an AI assistant would likely run
to answer it: "best [category] [year]", "[competitor] alternative",
"[A] vs [B]", "[brand] reviews", "[category] pricing".
Mark which ones need a comparison, alternative, integration or problem page.
Treat the result as a list of ideas. Search each query yourself to see what ranks. The background is in query fan-out explained.
Workflow 3: Turn the top results into a brief
Paste the text of the three pages that rank for your target query (copy the main content, not the navigation). Then:
Here are the three pages ranking for "[query]".
1. Table: each H2 on each page, and the question it answers.
2. Which questions all three answer, and which only one answers.
3. What none of them answer that a buyer would still ask.
4. A brief for a better page: H2s phrased as questions, the fact each
answer needs, and where we need our own data. Do not write the article.
Point 3 is where your page earns its place. Google's AI optimization guide asks for "non-commodity content" with "unique expert or experienced takes that go beyond common knowledge"; restating the top three pages adds nothing.
Workflow 4: Rewrite sections answer-first
Claude is good at restructuring without inventing, if you tell it not to:
Rewrite each section below so the first 2 to 3 sentences directly answer
the heading. Keep every fact, number and name exactly as given. Do not add
facts. If a section has no clear answer, say "NO ANSWER" instead of writing one.
Replace vague categories with names only where the names appear in my text.
The "NO ANSWER" instruction is the important part: it surfaces sections where you need to do research rather than letting the model fill the gap. The method is explained in how to write answer-first content.
Workflow 5: Draft schema and robots.txt rules for review
Claude can write JSON-LD and robots.txt from a description, and it saves time, but check both before they go live.
For robots.txt, give it the facts: Anthropic's own documentation lists three agents. ClaudeBot collects content that may be used for training, Claude-SearchBot indexes for search quality, and Claude-User fetches pages when a person asks Claude a question. Anthropic says its bots honor robots.txt, and it supports the non-standard Crawl-delay directive. If you want to stay visible in Claude's answers while opting out of training, the pattern is:
User-agent: ClaudeBot
Disallow: /
User-agent: Claude-SearchBot
Allow: /
User-agent: Claude-User
Allow: /
Anthropic notes that blocking Claude-User or Claude-SearchBot can reduce your visibility in Claude's web answers. For structured data, validate the JSON-LD with Google's Rich Results Test and confirm it matches the visible text.
Workflow 6: Let Claude Code make site changes, with rules first
Claude Code can edit a site's repository directly, which is useful for bulk fixes like adding "as of" dates to price tables or fixing internal links. It is also how pages quietly lose things. Write the rules before the first task, for example in the project's CLAUDE.md file:
| Tier | Examples |
|---|---|
| Allowed | Read files and data, draft pages and emails, build query lists, run lint checks |
| Needs my approval | Publish, change a title or URL, edit a comparison table, send anything under our name |
| Never on its own | Delete a page, rewrite rankings in bulk, touch robots.txt or canonicals |
Add one more rule: before any rewrite, list the elements the page exists for (a table, a calculator, an FAQ, specific internal links) and confirm each one is still there afterwards. Work one page per change so a broken page is easy to find.
Guardrails that apply to every workflow
- Verify every claim about a third party against their own page, and date it.
- Do not mass-publish unedited output. Google's spam policies name "using generative AI tools or other similar tools to generate many pages without adding value for users" as scaled content abuse.
- Keep your own voice and data in. Paste your notes, test results and real examples into the prompt; they are what make the page worth citing.
- Never paste secrets. API keys, client data under NDA and personal data stay out of prompts.
Using Claude inside Postbing
If you would rather not run these workflows by hand, Postbing plans, writes and publishes articles and can run on your own Anthropic key: with bring your own key you assign Claude to tasks such as writing, planning or editing. Postbing does not have an MCP server or a CLI, so it does not plug into Claude Code directly; it publishes through its CMS connectors and webhooks.
FAQ
Is Claude good for SEO writing?
It is good at structured drafting and rewriting from material you provide. It does not know your search volumes or rankings, so pair it with Search Console data and verify every fact before publishing.
Can Claude do keyword research?
It can generate candidate queries and group them, but it cannot give you real search volumes. Use its list as ideas and check volumes and rankings in your SEO tools or Search Console.
How do I make my site visible in Claude's answers?
Allow Claude-SearchBot and Claude-User in robots.txt and make sure your pages return the text in raw HTML. Anthropic says blocking those agents can reduce visibility in Claude's web answers.
Does blocking ClaudeBot remove me from Claude search?
Anthropic describes ClaudeBot as its training crawler and Claude-SearchBot and Claude-User as the agents used for search and user requests. Blocking ClaudeBot signals your content should be excluded from training.
Sources
- Anthropic: Does Anthropic crawl data from the web, and how can site owners block the crawler?: ClaudeBot, Claude-SearchBot, Claude-User, robots.txt, Crawl-delay, accessed 16 November 2026
- Google: Optimizing for generative AI search: non-commodity content, accessed 16 November 2026
- Google: Spam policies: scaled content abuse, accessed 16 November 2026
- Nicholas Dulait, "The KING of GEO" post on X (13 November 2026): agent permission tiers, one page per run, numbers from data not the model, accessed 16 November 2026
Related reading
Last updated: November 16, 2026. Spotted an error? Email support@postbing.com and we will fix it. Editorial policy.
Try Postbing free
Postbing reads your site, plans a month of topics from what your customers search for, writes in your voice and publishes to WordPress, Shopify, Webflow or Ghost. Your first article is free.
Try Postbing freeNo card needed. No link exchanges, ever.