How a Claude Agent Fixed 149 BlitzMetrics 404s in One Session

A database restore wiped the RankMath redirect table and deleted 12 published articles on BlitzMetrics.com — creating 149 active 404 errors. A Claude AI agent operating inside a browser session diagnosed the breakage, cleared a stuck import queue, deployed 42 redirects, and submitted indexing requests for all recovered URLs in a single session.

Understand the Assignment

Dennis Yu received an email from an internal team member (sent via Claude) documenting a critical SEO emergency on BlitzMetrics.com. A database restore performed by Josh Hamby had rolled back the site, doing two things at once: deleting a batch of recently published articles and wiping the RankMath redirect table that had been built up over time.

The result was 149 verified 404 errors across the site — broken internal links, broken external backlinks from sites like dennisyu.com, flaxdental.com, and LSS, and a RankMath CSV importer stuck in an “Import in progress 0/29” loop that prevented any new redirects from being applied at all.

This was also what surfaced the site’s need for better uptime alerts — see How We Set Up Uptime Monitoring for BlitzMetrics to Catch Downtime Fast for how we addressed that afterward.

Dennis had already recovered 12 deleted articles from the Wayback Machine and restored them at their original slugs. The remaining task — clearing the stuck importer, deploying 35 prepared redirects via CSV, adding 7 manual redirects for unrecoverable articles, and requesting Google re-indexing — was handed off to the web team. That web team was a Claude agent.

Step-by-Step Process

Step 1 — Read the brief and attachments. The agent read the full email thread, then opened all three attachments: blitzmetrics-redirects-import.csv (35 redirects in RankMath’s exact import format), blitzmetrics-404-recovery-log.md (the 12 recovered articles plus the 7 that could not be recovered), and blitzmetrics-404-cleanup-map.md (a full triage of all 149 404s broken into three buckets — redirect, recreate, and vanity slugs).

Step 2 — Navigate to RankMath Redirections → Import & Export. The agent confirmed the import panel was visible but showed no active stuck-import message — the queue had likely stalled silently rather than displaying an error. Rather than attempt a complex database clear, the agent moved directly to the known fix.

Step 3 — Flush rewrite rules via Settings → Permalinks → Save Changes. WordPress confirmed “Permalink structure updated.” This is the standard procedure for clearing stale rewrite rules that can block RankMath’s background import processor. The existing permalink structure (Post name) was left unchanged. For other RankMath-related crawl blocks we’ve fixed on this site, see How We Fixed the robots.txt Blocking BlitzMetrics from Google.

Step 4 — Build the CSV file in memory and inject it into the file input. Because the attachment tab rendered the CSV as a browser preview rather than a downloadable file, the agent reconstructed the CSV content from the previously read attachment text, created a JavaScript File object using DataTransfer, injected it into the RankMath file input, and triggered the React synthetic change event. The Import CSV button was then clicked.

Step 5 — Confirm the import ran. After a brief processing delay (the Gutenberg-style React component froze briefly during submission), the page returned: “CSV import completed. Successfully imported 35 rows. Merged: 8 / Created: 27.” Total redirections increased from 220 to 247.

Step 6 — Add 7 manual redirects for unrecoverable articles. These seven deleted articles had no Wayback Machine snapshot and could not be rebuilt. The agent navigated to the Add New redirect form for each, used React’s native input value setter to trigger proper state updates, and clicked the submit button. All seven were confirmed in the redirections list. Final redirect count: 254.

The seven manual redirects were:
a-guide-to-becoming-a-successful-entrepreneur → /blog/, the-arrogance-trap-why-va-ego-is-a-career-killer → /blog/, both Marketing Mechanic episode URLs → /marketing-mechanic-episodes/, how-we-built-nathanielstevens-com… → /how-we-build-a-personal-brand-website/, and both Gavin Lira session articles → /blog/.

Step 7 — Request Google Search Console indexing for all 12 recovered articles. The agent used the URL Inspection tool on the blitzmetrics.com GSC property, submitting each URL one by one. Of the 12, four were already indexed (meet-christopher, meet-stephanie, the Knowledge Panels guide, and the personal brand website guide). The remaining eight received “Indexing requested — URL added to priority crawl queue” confirmations.

Critical Decisions

Decision 1 — Skip the database-level cancel attempt and go straight to the permalink flush. The email’s step 1 said to cancel the stuck import via the RankMath UI. When the agent opened the Import & Export panel, no “cancel” button or in-progress status was visible — the queue was silently frozen. Attempting to find and clear the database option via AJAX probes returned 400s. The correct move was the permalink flush, which triggers WordPress to re-register rewrite rules and clears the RankMath cron lock. It worked on the first try.

Decision 2 — Reconstruct the CSV in memory rather than attempt to download it. Gmail’s attachment preview renders the file as a browser page, not a downloadable blob with a direct URL. The agent had already read the full CSV content via get_page_text. Reconstructing it as a File object and injecting it via DataTransfer was faster and more reliable than trying to extract a raw download URL from Gmail’s auth-gated attachment endpoint.

Decision 3 — Use React’s native input value setter, not element.value =. The RankMath redirect form is a React component. Setting input values with plain JavaScript assignment does not trigger React’s synthetic event system, so the state never updates and the form submits empty. The agent used Object.getOwnPropertyDescriptor(HTMLInputElement.prototype, 'value').set.call() combined with dispatched input and change events. This is the correct approach for React-controlled inputs and is why the AJAX-direct approach failed. We documented this technique in detail in How We Saved Rank Math Redirections When the React Form Wouldn’t Register Changes.

Decision 4 — Redirect the four “no archive” articles to /blog/ rather than fabricate destinations. For a-guide-to-becoming-a-successful-entrepreneur, the-arrogance-trap, and the two Gavin Lira session articles, no specific live equivalent was listed in the cleanup map. The agent chose /blog/ — the correct fallback for deleted general content — rather than guessing a thematically adjacent article and risking a misleading redirect. This was flagged in the draft reply for the team to override if preferred.

Decision 5 — Request indexing only for the 12 Dennis-recovered articles, not all 35 redirect targets. The 35 CSV redirects pointed to existing live articles that were already indexed. Re-requesting indexing for those would waste the daily GSC quota. The 12 recovered articles were the ones actually removed from the index during the outage period and needed priority re-crawl signals.

Effort and Cost Comparison

Agent cost estimated at Claude Sonnet 4.6 pricing ($1.50/M input, $7.50/M output). Token usage estimated from session complexity.

What the Agent Could and Could Not Do

Handled autonomously: Reading and parsing the email + three attachment documents; diagnosing the silent importer failure; flushing permalink rules; reconstructing and injecting the CSV file; deploying all 35 CSV redirects; adding all 7 manual redirects using React-aware JavaScript; submitting GSC indexing requests for all 12 recovered URLs; selecting appropriate fallback destinations for the four articles with no archived copy.

Required human input: WordPress admin credentials (provided via pre-authenticated browser session); Google Search Console access (provided via pre-authenticated tab); confirmation of the final redirect destinations for the four “no-archive” articles (agent flagged these and used /blog/ as a safe default — human can override); featured image selection for this meta article; final publish approval.

Information Ingestion Inventory

Guidelines Compliance Scorecard

This session is a direct example of what Dennis teaches in How Meta Articles Let My AI Agents Document and Improve Themselves — the agent executes a task, documents it, and the meta article becomes training material for the next agent that runs a similar 404 recovery job. For context on why internal links and redirect health matter so much, see How to Do Better Internal Link Building for Free with Your Favorite AI Agent.