25 prompts to make Claude your SEO expert https://x.com/i/article/2115333631883595777
How to run your SEO from Claude (25 prompts)
In this article I'll walk you through the full setup, with the 25 prompts I run on my own Search Console, numbered in the order you'll need them.
Search Console has all the data you need to grow your traffic. It just doesn't tell you what to do with it.
You open it, you see a chart, and the chart doesn't come with a list of what to fix.
This article is the setup I use to turn that data into decisions, with Claude doing the analysis:
- how to connect Claude to your Search Console, GA4 and keyword lists
- how to see where your site actually stands, beyond the headline numbers
- how to find the quick wins sitting on page 2
- how to find what's losing traffic, and why
- how to tell when two of your pages are fighting each other, and when they aren't
- how to decide what to write next, from queries you already get
- how to read what AI search is doing to your traffic
- how to run all of it as one weekly review
Every prompt is in here, ready to copy. There are 25 of them, numbered in the order you'll use them. The screenshots are real answers from my own site, with page names and keywords blurred.
Why Search Console is hard to use
Search Console is free, it's accurate, and it's the closest thing you have to Google telling you what it thinks of your site.
The hard part is getting anything out of it, and it breaks in the same four places.
1. It tells you what happened, not what to do.
You can see that clicks went up and impressions went down. You can't see why, or which of your pages caused it, or what you should change.
2. The report stops at 1,000 rows.
The performance report shows you the top 1,000 queries and pages. On any site with real traffic, the long tail is where most of the opportunities are, and it's the part you can't see.
3. Every real question needs a spreadsheet.
"Which pages lost rankings on queries they used to win" means exporting two date ranges, joining them, and pivoting. Add pages, queries and time together and it's an afternoon of moving rows around.
4. So it doesn't get done.
Because each answer takes hours, most people check the chart, feel vaguely good or bad about it, and close the tab. The analysis that would actually move traffic happens once a quarter, if that.
What changed
Three things are true now that weren't true a year ago.
Claude can read your SEO data directly.
CrawlRaven's connector puts your Search Console, GA4 and your Ahrefs or Semrush keyword lists inside Claude, and its checks already flag cannibalization, decaying pages and keywords close to page 1. You ask a question in plain English and the answer comes back from your real data, not from a guess.
It can cross thousands of rows in seconds.
Pages against queries against two date ranges, the job that took an afternoon in a spreadsheet, is one prompt.
It turns the finding into the fix.
It doesn't stop at "this page lost clicks". It tells you whether the page lost rankings, lost demand or lost clicks at the same position, and writes the new title, the brief or the redirect plan.
The way to think about it: Claude is the analyst. CrawlRaven is the data and the SEO engine underneath. The same prompts also work in ChatGPT.
Who this is for, and who it isn't
This works best when:
- your site already gets search impressions (a few hundred a day is plenty to start)
- you run a SaaS, a content site, a store, or SEO for clients
- you'd rather make one good change a week than read reports
It works badly when:
- your site is brand new and Search Console has almost nothing in it yet
- you're looking for AI to write a hundred pages for you
If you're in the second list, start with the basics and come back when Search Console has a few months of data.
The system
The whole thing is six jobs.
The analyst is the only part that touches everything else. It reads your data, runs the checks, and writes back the fix. If you swap one tool for another later, the prompts stay the same.
The rest of this article is the setup, in eight steps.
Step 1: connect Claude to your data
What it does: gives Claude your real numbers, so it stops answering from general SEO advice.
Why it matters: a Claude that can't see your data can only tell you what SEO blogs say. A Claude that can see it tells you what's happening on your site.
How to do it:
In Claude, add a custom connector, choose the remote option, and paste CrawlRaven's server address:
https://mcp.crawlraven.com/mcp
Claude opens CrawlRaven's sign-in in your browser. Approve the permissions you need (Search Console is the main one; analytics is a separate permission if you want the GA4 prompts) and you're connected. You don't paste any Google tokens or passwords anywhere.
Using ChatGPT, Codex, Cursor or another AI agent? The setup for each one is here: crawlraven.com/docs/mcp
you can still rinithout CrawlRavenExport your Search Console performance report and paste it in. Every prompt below still works on the rows you give it. You'll be limited to those rows, with no GA4, no history and none of the automatic checks, but it's a real way to start.
Here's what it looks like when you hit enter. Claude calls the tools on its own and pulls what it needs:
Then give Claude its standing brief. Put this in a Claude Project's instructions so every chat starts with it.
Prompt 1: the SEO analyst brief
You are my SEO analyst for [site]. My goal is [signups / sales / leads], and the pages that matter most are [list them].
Every number you give me must come from my connected data. Quote the actual rows. If you can't get a number, say so. Never estimate one.
Compare the last 28 days with the 28 days before unless I ask for a different window.
For every finding, give me one fix, the page it applies to, and how I'll know in 28 days whether it worked.
Keep answers short: the finding first, then a table, then the fixes.
The rule that matters most is "never estimate a number". Without it, AI fills gaps with plausible numbers. With it, every figure traces back to a row in your data.
Step 2: know where you stand
What it does: gives you the real picture of the last month, not just the headline.
Why it matters: the headline numbers hide most of what's going on. A month can look fine overall while the last two weeks are falling.
How to do it:
Prompt 2: the 28-day health check
Give me a health check for [site]. Compare the last 28 days with the 28 days before: clicks, impressions, CTR and average position. For each one that moved, tell me which pages and queries drove the change, quote the numbers, and end with the 3 things I should look at first.
Here's what it gave me on my own site:
What you get: the four numbers that matter, what moved them, and a short list of where to look.
On my site the headline said clicks were up 10%, 449 against 408. Claude split the window in half and found the real story: 297 clicks in the first two weeks and 152 in the last two, with average position sliding from around 17 to 36. The monthly total was hiding a drop. I'd have missed it looking at the chart.
Prompt 3: which pages actually bring customers
Join my Search Console and GA4 data for the last 90 days. For each landing page that gets organic traffic, show clicks, sessions, key events and the conversion rate. Which pages bring traffic but no signups, and which convert well but get very little traffic?
What you get: the pages worth protecting, and the pages worth pushing. A page with a strong conversion rate and low traffic is usually your best SEO bet. This one needs GA4 connected.
Prompt 4: your real keyword map
For my top 20 pages by impressions, show which queries each one ranks for. Flag any query where the page that ranks isn't the page I'd want ranking, and any page that ranks for queries unrelated to what it's about.
What you get: what each page actually ranks for, which is often not what it was written for.
Step 3: quick wins
What it does: finds the changes most likely to move clicks in the next month.
Why it matters: a keyword sitting at position 11 is invisible. The same keyword at position 6 gets clicks. Pages already close to page 1 are where small changes pay off fastest.
How to do it:
Prompt 5: keywords one push from page 1
Find the keywords on [site] sitting between position 8 and 20 over the last 28 days with at least 200 impressions. Skip queries that are another company's brand name. Show a table (query, page, position, impressions, clicks) sorted by impressions, and for the top 5 tell me the one change most likely to move each onto page 1.
What you get: a short list of queries with real demand that are almost there, and one specific change for each.
On my site, Claude found six of these. Five of the six had zero clicks, and the fixes were small: a heading that didn't use the words people search, an FAQ that didn't answer its own question, a page offering one example when the search asked for several. None of it needed a new page.
The "skip other brands" line matters. Without it, your list fills up with competitor names you'll never win.
Prompt 6: titles nobody clicks
Find pages with at least 500 impressions in the last 28 days whose CTR is low for their position. For each one, read the current title, show the top 3 queries it appears for, and write 3 new titles that match what those people are searching for.
What you get: pages that already rank but lose the click, with replacement titles written from the actual searches.
Prompt 7: things you rank for but never targeted
Show me queries where I get impressions but none of my pages is really about that topic. Group them by theme, and tell me which ones deserve a new page and which just need a new section on an existing page.
What you get: demand Google already connects to your site, that you've never written for.
Prompt 8: internal links to your near-winners
For my pages sitting between position 8 and 20, which of my stronger pages should link to them? For each, suggest the page the link goes on, where on that page, and the anchor text.
What you get: the cheapest ranking boost there is, using pages you already have.
Step 4: what's losing traffic
What it does: finds the pages that are dropping and tells you why each one dropped.
Why it matters: "this page lost clicks" can mean three different things, and each one needs a different fix. Rewriting a page that lost demand is wasted work. Leaving a page that lost rankings alone loses more.
How to do it:
Prompt 9: the drop diagnosis
My clicks dropped. Find where: split the change by page and by query, and find the day it started. Tell me whether it's one page, one group of pages, or the whole site, and what each of those usually means.
What you get: whether you have a page problem or a site problem. One page dropping is usually competition or freshness. A whole group dropping together usually points to a change on your side, or a ranking update.
Prompt 10: what's losing clicks, and why
Which pages on [site] are losing clicks? Compare the last 28 days with the 28 before. For each page that dropped, tell me whether it lost rankings, lost impressions, or lost clicks at the same position, and what that points to.
What you get: every page that dropped, sorted into the three kinds of drop:
- lost rankings: someone overtook you. Refresh the page, or strengthen it with links.
- lost impressions at the same position: fewer people are searching for it. There's nothing to fix on the page.
- lost clicks at the same position: something on the results page is taking the click, often an AI answer or a better title. Rewrite the title and the snippet.
On my site, one page went from 76 clicks to 24 because it fell from around position 10 to 23. That's a rankings problem, so it's the one worth working on. Two others lost clicks at the same position, which is a title problem, not a content problem.
Prompt 11: lost queries
List the queries I had clicks for in the 90 days before last, and have none for now. Group them by page and tell me which pages lost the most.
What you get: the specific searches you've stopped showing up for, which a page-level view hides.
Prompt 12: stale pages losing ground
Which of my pages are declining and haven't been updated in six months? Sort them by the clicks they've lost.
What you get: your refresh list. A declining page nobody has touched in six months is usually the easiest win on this list.
Step 5: pages fighting each other
What it does: finds queries where several of your pages show up, and tells you whether that's actually a problem.
Why it matters: this is where most people make things worse. They see two pages ranking for the same query, merge them, and lose traffic both pages were getting. Checking the intent first is the whole step.
How to do it:
Prompt 13: find every split query
Find every query where two or more of my pages show up and the top page gets less than 70% of the impressions. Sort them by impressions.
Prompt 14: does it need fixing, and who owns it
Take the biggest one. Which pages are splitting it, what share does each one get, and does it actually need fixing? If it does, which page should own it?
Here's what happened when I ran this on my own site:
CrawlRaven flagged a query where 23 of my pages showed up, with 3,856 impressions over 90 days and zero clicks. On paper, it's textbook cannibalization.
Claude looked at the intent and said it wasn't. The query was someone searching for another company's own tool. Nobody typing it wanted my site, which is why there were zero clicks at every position. The homepage already showed up for 97% of those searches.
So the right plan was mostly to leave it alone: merge nothing, redirect nothing, and make one page clearly answer the one version of that search I could honestly serve. If I'd merged pages to "fix" it, I'd have put real traffic at risk to chase clicks that were never coming.
Prompt 15: the merge, redirect or rewrite plan
For that query, give me the plan: what to merge, what to redirect, what to rewrite, and which internal links to change. Tell me how much traffic each change puts at risk.
What you get: a plan that protects the traffic you already have. The "traffic at risk" line is what stops you from merging away a page that's quietly doing well.
Prompt 16: target keywords on the right page
Here are my target keywords and the page I want ranking for each: [list]. Check which ones land on the right page, which land on a different page, and which don't rank at all.
What you get: the gap between your plan and what Google actually did with it.
Step 6: what to write next
What it does: turns the queries you already get into a list of pages worth writing or updating.
Why it matters: the best content ideas aren't in a keyword tool. They're in your own Search Console, as searches Google already shows you for.
How to do it:
Prompt 17: every question people search to find you
List every query that's phrased as a question (who, what, why, how, can, does, is) or is 6 words or longer. Group them by topic, and for each group tell me whether an existing page answers it or whether it needs its own page.
What you get: the exact questions your audience asks, grouped into page ideas. These long, specific queries are also the kind AI answers pull from.
Prompt 18: content gaps
Which queries get impressions where my best position is worse than 30? Group them, and for each group tell me whether a new page would have a real chance or whether the topic is out of reach for my site right now.
What you get: topics where Google already sees you as relevant, with an honest read on which ones are winnable.
Prompt 19: the refresh brief
Write a refresh brief for [page]: the queries it ranks for now, the ones it used to rank for and lost, what's missing compared to what people search, and the sections to add or rewrite. Don't change what's already working.
What you get: an update plan that adds what's missing without breaking the parts that rank.
Prompt 20: a brief from your real ranking data
I want to write a new page about [topic]. Using the queries I already get impressions for around it, write a brief: the main query, the questions to answer, the sections, and which of my existing pages should link to it.
What you get: a new page planned from real demand, with its internal links decided before you write a word.
Step 7: AI search
What it does: shows how AI search is affecting your traffic, and helps you get quoted in AI answers.
Why it matters: more searches now end in an AI answer. You want to know whether your pages show up there, how much traffic AI tools send you, and how to write so they quote you.
How to do it:
Prompt 21: what the AI report says
Read my imported Search Console AI report. Which of my pages show up in AI Overviews and AI Mode, how has that changed, and how does it compare with their regular search impressions?
Google started rolling out these AI reports in Search Console in June 2026, to some sites first. They show impressions only, with no clicks or queries yet. If you have the report, CrawlRaven can import it, and this prompt reads it.
Prompt 22: how much traffic AI sends you
How much traffic did AI assistants send [site] in the last 90 days? Look at GA4 sessions from chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai. Show sessions and key events for each source, the landing pages that traffic arrived on, and how it compares with the 90 days before.
What you get: real numbers on what AI tools send you, and which pages they send people to. This one needs GA4 connected.
Prompt 23: make your page quotable
Take my page [page] and the queries it ranks for. Rewrite its opening section so the first two sentences answer the main question directly, with the specific facts an AI answer could quote. Keep my voice, and don't add any claim I can't back up.
What you get: a section that answers first and explains second, which is the shape AI answers tend to quote.
Step 8: the weekly loop
What it does: turns all of this into a short weekly routine, and remembers what you changed.
Why it matters: SEO changes take weeks to show up. If you don't write down what you changed and when, you'll never know which fix worked.
How to do it:
Prompt 24: log the change, judge it in 28 days
Log this change: on [date] I [what changed] on [page]. In 28 days, compare that page's clicks, impressions and position with the 28 days before the change, and tell me whether it worked.
CrawlRaven stores these as notes on your timeline, so the next review knows what changed and when.
Prompt 25: the weekly review
It's my weekly SEO review. Compare the last 7 days with the 7 before, and the last 28 with the 28 before. Tell me what moved, which logged changes are old enough to judge, the one thing most worth fixing this week, and what I should leave alone.
What you get: one prompt that checks everything, once a week. CrawlRaven syncs your Search Console data daily, so it always runs on fresh numbers.
What to watch out for
Made-up numbers. If Claude can't reach a number, it may fill the gap. That's why "never estimate a number" is rule one of the brief. Ask it to quote the rows when something looks off.
Judging too early. A change you made last week hasn't shown up in Search Console yet. Wait 28 days before you call it a win or a loss.
Chasing impressions. Impressions are the easiest number to grow and the least useful. Judge your changes on clicks, and on signups if GA4 is connected.
Changing everything at once. Fix one thing per page and log it. If you change ten things together, you won't know which one worked.
Merging too fast. Two pages on the same query isn't automatically a problem. Run prompt 14 before you merge anything.
The full setup
1. connect Claude to your data and give it the analyst brief
1. run the health check to see where you actually stand
1. work the quick wins: page-2 keywords, weak titles, internal links
1. find what's losing traffic and sort each drop by its cause
1. check split queries for intent before you merge anything
1. plan new pages from the queries you already get
1. measure what AI search is doing to your traffic
1. log every change and run the weekly review
If this was useful:
- repost it for anyone still exporting Search Console into spreadsheets
- follow me for more on running SEO with AI
Tools used in this setup: Claude and CrawlRaven.com