How to use Claude to find keyword cannibalization — and actually fix it.

The most direct way to learn how to use Claude to find keyword cannibalization is also the simplest: export your Google Search Console Performance report grouped by query and page, paste it into Claude, and ask it to flag every query where more than one URL earns impressions or clicks. Claude groups the rows by query, calculates the click split, identifies which URL performs better on CTR and position, and recommends whether to consolidate via 301 redirect, add a canonical tag to the weaker page, or differentiate the two pages’ keyword targets.

To get the export right, open GSC and navigate to Search results in the Performance menu. Enable clicks, impressions, CTR, and position. Click the Pages tab, then export — this gives you a query-by-page grouped CSV where each row is a unique query-page combination. Paste the CSV rows into Claude with the prompt: “Find keyword cannibalization in this GSC data. Group by query, flag every query where two or more pages rank, identify the keeper URL, and give me one fix per conflict: consolidate, canonical, or differentiate.” Claude’s 200K-token context means even large exports clear in a single pass — no pagination, no sampling.

The Cannibalization Detector skill (available from Hawk Academy and Ryze AI’s open-source skill library) turns Claude Desktop into a dedicated cannibalization analysis tool. Install it with a single curl command into your ~/.claude/skills/ folder on Mac or %USERPROFILE%\.claude\skills\ on Windows, restart Claude Desktop, open a new conversation, and type: “Find keyword cannibalization in my Search Console data.” Claude then prompts you to paste your GSC export and returns a structured report: true cannibalization vs. harmless overlap, the keeper URL per conflict, and the single recommended fix.

The advantage over a raw paste is consistency. The skill primes Claude’s behavior so every audit uses the same analysis logic, the same output format, and the same severity classification. If you run cannibalization checks for multiple clients or sites, this removes the need to re-engineer the prompt each time. You can also skip the install entirely by copying the full skill prompt into the System Prompt of a Claude Project — which takes the workflow to approach #9 in this list and makes it shareable across a team.

If your blog lives in a file-based system — Next.js MDX files, Hugo markdown, Jekyll posts, or similar — Claude Code’s /blog cannibalization ./content/blog command (from the Claude Blog skill) scans every post in your content directory. It extracts primary and secondary keyword targets from each file’s title, H1, and meta description, runs semantic similarity scoring across the full corpus, classifies each post’s search intent (informational, commercial, transactional), and outputs a conflict matrix showing which posts compete — plus a merge, redirect, or differentiate recommendation for each cluster.

This approach shines for teams that treat SEO as part of their engineering workflow. You can run it pre-deploy to catch cannibalization before a new post ships, or pipe it into a CI/CD step that fails if two posts target the same primary keyword. The limitation is coverage: it reads what is in your files, not what Google actually ranks — so it can miss cannibalization where an old page ranks for a keyword it was never explicitly optimized for. Pairing it with a GSC-based approach (method #2 or #3) closes that gap. For internal links between those posts, see our guide on Claude keyword research automation.

Screaming Frog crawls your site and exports every URL with its title tag, H1, H2s, and meta description in a single CSV. Feeding that export to Claude — with the prompt “Identify URLs with overlapping primary keyword targets in their title tags and H1s. Group by likely search intent, flag cannibalizing pairs, and recommend whether to consolidate or differentiate” — gives you a purely on-page view of cannibalization that GSC data misses entirely.

This matters because not every cannibalizing page has ranking history in Search Console yet. A new product category page and an existing blog post may both be optimized for “best running shoes for flat feet” in their title tags even if neither has sent measurable GSC signals yet. Catching it before Google has to arbitrate is far cheaper than fixing it after rankings split. The Screaming Frog approach works on any site regardless of whether you have GSC access — useful for agency audits of client sites you’ve just onboarded. See also our post on connecting Claude to your marketing data via MCP for a more automated version of this pipeline.

Standard GSC-based cannibalization analysis tells you which URL has better click-through rate and average position. Ahrefs adds the third dimension: which URL has more referring domains, higher UR (URL Rating), and more estimated organic traffic from other keywords. When you paste an Ahrefs Top Pages or Organic Keywords export alongside your GSC data and ask Claude to cross-reference them, it can recommend keeping the URL that is both the better performer and the stronger link equity holder — which sometimes are not the same page.

This matters most on sites where old blog posts have accumulated backlinks over years but newer, better-optimized pages have outranked them on position. Redirecting the old page to the new one loses those backlinks in the short term; consolidating content onto the old URL preserves them. Claude can model both scenarios and recommend which path loses less authority. This level of analysis is difficult to do manually at scale and is one of the clearest demonstrations of how to use Claude to find keyword cannibalization in a way that goes beyond what standard SEO tooling surfaces. Pair this with our Claude keyword research automation guide to keep new content from recreating the same conflicts.

Semrush Position Tracking exports show weekly rank history by keyword and URL. When you feed a multi-month export to Claude and ask it to identify keywords where the ranking URL changed between sessions — especially cases where a new page displaced an older page for the same term — you get a time-series view of cannibalization as it developed. This is particularly valuable for content teams publishing frequently: you can pinpoint which new article started competing with an existing post and exactly when the rank split occurred.

The prompt to use: “Analyze this Semrush position tracking export. Find keywords where the ranking URL changed across dates. Flag cases where two different URLs from my domain ranked for the same keyword in the same period. Identify the displacement direction and recommend whether the original or newer URL should be the canonical target.” Claude handles the temporal clustering that would take a human analyst hours in a spreadsheet. The limitation is coverage: Semrush Position Tracking only includes keywords you added to your project, so it is best used alongside a comprehensive GSC export to avoid blind spots.

The manual version: search Google for site:yourdomain.com “target keyword” for each suspected cannibalizing query, copy the results, and paste them into Claude with the instruction to identify which pages are competing and recommend a resolution. It is slow and incomplete by nature — Google’s site: operator has been notoriously unreliable since 2023, often returning only a fraction of indexed pages — but it requires no tool access whatsoever, which makes it the right starting point for competitor research or for auditing a new client before you have their GSC credentials.

To scale this approach slightly, use the Semrush or Ahrefs free tiers to pull a sample of your top 50 keywords, search each via site: operator, compile the results, and paste the combined list to Claude. Even with sampling, this surfaces the most egregious cannibalization conflicts — often the ones costing the most clicks. For a more thorough and automated version of this workflow, the GSC paste method (#2) or the Cannibalization Detector Skill (#3) will outperform it significantly.

Claude Projects lets you save a system prompt that loads automatically for every conversation in that project. For cannibalization detection, this means you define the analysis rules once — how to classify true cannibalization vs. overlap, what severity thresholds to use, what output format to produce — and every team member who opens the project gets the same expert-level behavior from Claude without needing to re-engineer the prompt each session.

Set up a project called “Cannibalization Audits,” paste the full Cannibalization Detector skill prompt as the system prompt, and add a brief like: “You are an SEO analyst specializing in keyword cannibalization. When given GSC data, produce a structured report with: (1) executive summary of total conflicts found, (2) high-severity conflicts table, (3) medium-severity conflicts table, (4) recommended fix per row, (5) a prioritized action plan.” Every team member then simply opens the project and pastes the new client’s GSC export. The output is consistent enough to drop directly into a client deliverable. For connecting these workflows to live data pipelines, see our guide on Claude MCP integrations.

For teams running cannibalization audits across 20, 50, or 200 client sites, building a Python script that pulls GSC data via the Google Search Console API, chunks it into Claude API calls, and writes results to a Google Sheet or Airtable base is the only practical solution. The architecture is straightforward: authenticate to GSC API, pull the query-page performance report for each site property, send chunks of 500 rows at a time to Claude with the cannibalization detection prompt, parse the structured JSON output Claude returns, and aggregate into a master conflict table.

Claude Haiku handles this task at roughly 60% of the accuracy of Claude Sonnet 4 at 20% of the cost — for bulk pre-screening, Haiku finds the high-severity conflicts well. Route flagged cases to Sonnet 4 for the final recommendation. With a 15-minute scheduled Lambda function, you can have a weekly cannibalization report for your entire client portfolio ready in your inbox every Monday with no manual work. This is the developer’s equivalent of what Ryze AI delivers out of the box for non-technical users, and understanding Claude MCP is the fastest path to this level of automation without writing a custom script from scratch.