RankOps is an autonomous SEO agent in production that connects to Google Search Console, identifies underperforming pages, generates optimized metadata using Claude and LangGraph, deploys changes to Webflow CMS, then A/B tests the results and auto-reverts any change that doesn’t improve click-through rate. It’s an agent that governs its own output.
The Problem
SEO metadata optimization is high-volume, low-creativity work. A typical site has hundreds of pages with suboptimal titles and descriptions. Manually rewriting them takes an SEO team weeks. And after rewriting, there’s no systematic way to know if the changes actually helped. Most teams just deploy and hope.
The Approach
ETL pipeline pulls search performance data from Google Search Console: pages, queries, impressions, clicks, CTR, position. Identifies pages that get impressions but low clicks (the biggest opportunity).
LangGraph agent takes an underperforming page, retrieves the current content and metadata, generates optimized title and description using Claude, and produces a structured output.
Webflow CMS integration deploys the change automatically. No manual copy-paste, no developer involvement.
A/B testing runs for a configurable period (default: 2 weeks). Compares CTR of the new metadata against the baseline. If CTR doesn’t improve meaningfully, the change is auto-reverted.
OpenTelemetry + Sentry instrumentation tracks every step: which pages were selected, what was generated, what was deployed, what the test results were. Full observability into the agent’s decision-making.
Key Decisions
Auto-revert on failure. The agent doesn’t just deploy and forget. Every change is measured. If the new metadata doesn’t improve CTR, it’s rolled back automatically. This makes the system safe to run unsupervised. Bad changes are temporary.
LangGraph for orchestration. The workflow has multiple steps with conditional branching (skip pages that are already optimized, retry on API failures, route to different prompts based on page type). LangGraph handles this state machine cleanly.
Trigger.dev for scheduling. The ETL pipeline, agent runs, and A/B test evaluations are all scheduled jobs. Trigger.dev handles retries, dead-letter queues, and monitoring.
Results
- Autonomous operation runs without human intervention
- Self-governing reverts changes that don’t work
- Full observability every decision logged and traceable via OpenTelemetry
- Measurable impact every change has a before/after CTR comparison
Retrospective
The auto-revert mechanism is what makes this production-safe. Without it, an agent that makes bad changes is worse than no agent at all. The measurement-first approach (test every change, revert failures) is a pattern I’d apply to any autonomous system. If rebuilding, I’d add more sophisticated segmentation: different strategies for different page types, and I’d track downstream conversion impact, not just CTR.