Deeper Analysis with Claude:
From Data to Decisions
Tutorials 1 and 2 focused on Claude for individual tasks — reviewing title tags, summarising audits, writing client copy. This tutorial goes a level deeper: using Claude for multi-step reasoning and strategic analysis. You'll learn how to upload and interrogate Search Console data, identify keyword cannibalisation, compare page structures, and generate full technical SEO briefs. The goal is to use Claude not just as a writing assistant, but as an analyst.
- Structure complex, multi-part analysis prompts effectively
- Use Claude to identify cannibalisation patterns in Search Console data
- Compare competing pages and produce a content consolidation recommendation
- Generate a complete technical SEO brief from a client intake form
- Understand when to break analysis across multiple prompts vs. one long prompt
1The Shift from Tasks to Analysis
There's an important distinction between using Claude for tasks and using it for analysis. Tasks are discrete and well-defined — "rewrite this title tag", "fix this schema". Analysis is open-ended — "what's going wrong with our rankings and why?"
Discrete, well-defined. Claude executes a specific instruction. Output is directly usable. One prompt, one output. Examples: rewrite, summarise, translate, format.
Open-ended, reasoning-heavy. Claude interprets data, identifies patterns, and makes recommendations. Often involves multiple follow-up prompts. Output informs decisions rather than being the final deliverable.
Analysis-level use requires a different prompting approach: you need to give Claude more context, guide it through a reasoning chain, and ask follow-up questions to drill into findings. The rest of this tutorial teaches exactly that.
2Preparing Your Data for Claude
Before any analysis prompt, the quality of data you provide determines the quality of insight you get back. Here's how to prepare the three main data sources used in this tutorial:
Go to Performance → Search results. Set your date range (last 3 months is ideal for spotting trends). Click Export → Download CSV. In the export, you want the Queries tab. Optionally, also export the Pages tab separately.
For cannibalisation analysis, use the Search type: Web filter and enable both Clicks, Impressions, CTR, and Position columns. The more date range you include, the more Claude can spot trends vs. one-off spikes.
Claude can handle up to roughly 200–300 rows pasted directly. For larger exports, filter to your most important queries first (e.g. top 200 by impressions), or run multiple focused analysis prompts.
For page structure comparisons, you don't need to paste entire page HTML. Instead, paste: the page title, meta description, all heading tags (H1–H3) in order, and the first paragraph of each main content section. Claude can infer content structure from headings far more efficiently than from raw HTML.
A quick way to grab headings: open the page, open browser DevTools, and run Array.from(document.querySelectorAll('h1,h2,h3')).map(h => h.tagName + ': ' + h.innerText).join('\n') in the Console tab. Copy the output.
For the brief generation workflow, you'll need: the client's website URL, industry and target audience, their top 3–5 business goals, any known technical issues, their current CMS and hosting setup, and any previous SEO work done. A one-page intake form filled out before an onboarding call gives Claude everything it needs.
Keyword cannibalisation happens when two or more pages on the same site compete for the same search query — splitting clicks and confusing Google about which page should rank. It's one of the most common issues on established sites and one of the hardest to spot manually at scale.
This workflow uses a two-prompt chain: the first prompt identifies cannibalisation signals, the second produces specific recommendations for each case found.
How to structure the analysis chain
When two pages are cannibalising, or when a page is underperforming against a clearly stronger competitor, you need to understand why at a content level. This prompt takes the heading structure and key content sections of two pages and produces a direct, actionable comparison.
What to paste: Use the DevTools console trick from Section 2 to quickly extract headings from both pages. For each page, also paste the meta title, meta description, and URL. You don't need to paste full body copy — headings alone give Claude a strong structural signal.
One of the highest-value uses of Claude in agency work is turning a client intake conversation into a structured, comprehensive SEO brief — the kind that previously took a senior strategist a half day to write. This prompt generates the full brief in one pass, ready for review and light editing.
The intake form template
Before running this prompt, fill out this intake form for the client. The more detail you provide, the more actionable the brief will be:
BUSINESS - Company name: [NAME] - Website: [URL] - Industry / niche: [INDUSTRY] - Main products or services: [LIST] - Target audience: [WHO ARE THEIR CUSTOMERS?] - Key geographic markets: [e.g. UK, US, Global] SEO GOALS - Primary goal: [e.g. Increase organic leads by 30% in 6 months] - Secondary goals: [e.g. Improve rankings for [keyword cluster], fix site speed] - Any upcoming events: [e.g. site migration, product launch, rebrand] TECHNICAL CONTEXT - CMS: [e.g. WordPress, Shopify, custom] - Hosting / CDN: [if known] - Site size (approx. pages): [NUMBER] - Known technical issues: [e.g. slow page speed, duplicate content, no schema] - Previous SEO work: [e.g. Audit done 12 months ago, on-page work only] COMPETITIVE CONTEXT - Main organic competitors: [LIST 2–4 DOMAINS] - Where do they currently rank vs competitors? [e.g. Page 2 for most head terms] CONSTRAINTS - Development resource available: [e.g. One developer, 2 days/month] - Content resource available: [e.g. In-house writer, 2 articles/month] - Budget level: [e.g. Limited / Medium / Significant]
Editing the brief: Claude's first draft will be ~80% ready. The most common edits needed: (1) making quick wins more specific to the actual client's stack, (2) adjusting timelines to match real developer availability, (3) adding any client-specific context Claude couldn't infer. A 15-minute edit is all that's usually needed.
6Getting Better Reasoning from Claude
For complex analytical tasks, there are several techniques that produce significantly deeper and more reliable reasoning:
Ask Claude to reason before concluding
Adding "think step by step" or "reason through this before giving your answer" prompts Claude to work through the problem explicitly rather than jumping to a conclusion. This is especially useful when the answer isn't obvious.
"Which of these two pages should we consolidate into?"
"Which of these two pages should we consolidate into? Before giving your recommendation, reason through: backlink profile, content quality signals, URL structure, and historical ranking performance."
Ask for confidence levels
Claude can't be certain about things it can only infer. Adding "flag any assumptions you're making and your confidence level in each recommendation" produces more honest, trustworthy output — important when the output informs real client decisions.
Ask Claude to steelman the opposite view
For strategic decisions, ask: "Now argue the case for the opposite recommendation — what would make consolidation a bad idea here?" This surfaces risks and edge cases Claude might not have volunteered.
Break very large analyses into focused sub-questions
| Instead of one giant prompt… | Use a sequence of focused prompts |
|---|---|
| "Analyse everything wrong with this site's SEO" | Prompt 1: "Analyse crawlability issues only." → Prompt 2: "Now analyse the on-page issues." → Prompt 3: "Summarise the top 5 priorities across both analyses." |
| "Tell me why our rankings dropped" | Prompt 1: "Here's our GSC data — what queries declined most?" → Prompt 2: "For those queries, what are the most likely causes?" → Prompt 3: "Which cause is most likely given that we did [X] on [date]?" |
7Practice Exercises
Run the cannibalisation analysis on a real client's Search Console data:
- Export the last 3 months of query data from GSC for one client
- Filter to the top 150–200 queries by impressions
- Run Prompt A1 and review Claude's cannibalisation findings
- Pick the most severe case and run Prompt A2 as a follow-up
- Compare Claude's recommendation to your own instinct — do you agree? If not, push back and ask Claude to reconsider given your additional context
Generate a full technical SEO brief for a new or existing client:
- Fill out the intake form template from Section 5 for a real client
- Run Prompt C1 and read the output brief carefully
- Identify 3 things the brief got right without you having to spell them out
- Identify 2 things that need editing or are inaccurate — make the edits
- Estimate: how long would this brief have taken to write manually? What's the time saving?
Experiment with the reasoning techniques from Section 6:
- Take any analysis prompt from this tutorial and run it without any reasoning instructions
- Run the same prompt again, adding "reason step by step before giving your final answer"
- Compare the two outputs — is the reasoning more transparent? Is the recommendation different?
- Try adding "flag any assumptions and your confidence level in each point" to the second run
- Share both outputs with a colleague and discuss which they find more useful
8Summary
Key takeaway: The difference between Claude as a task tool and Claude as an analyst is in how you prompt it. Provide rich context, ask for structured reasoning, chain your prompts from broad to specific, and always apply your own SEO expertise when reviewing the output. Claude is the analyst — you're the senior partner who reviews the work.