
Weekend Showcase Challenge: Search Console Agent
An AI-powered SEO agent that authorizes with Google Search Console once, then delivers weekly, actionable reports on what to fix on your site — pulling live data from Search Console, sitemaps, and rendered HTML.
Vision: what it does
Most site owners know SEO matters, but few have the time to read Search Console dashboards every week and translate raw numbers into action. Search Console Agent closes that gap. You connect your Google Search Console account once through OAuth, and from then on the agent works quietly in the background: once a week, it pulls your latest performance data, reads your sitemap, fetches and inspects your actual page HTML, checks how your pages render, and turns all of that into a clear, prioritized report — delivered straight to your mail inbox for your all sites — telling you exactly what needs to change on your site for better search visibility.
The idea was to remove the two biggest barriers to consistent SEO work: remembering to check the data, and knowing what the data means. One authorization, and the insight comes to you.
How I built it
The flow is intentionally simple on the surface, even though a fair amount happens underneath it:
- One-time authorization — a user signs in with Google OAuth (
webmasters.readonlyscope), and the app persists the linked account and refresh token so no repeated logins are needed.
- Site discovery — the backend lists every verified Search Console property tied to that account.
- Weekly analysis pipeline — on a schedule, the agent fetches 30 days of Search Console performance metrics, reads the site's sitemap, retrieves and parses the actual HTML of key pages, and checks rendering behavior to catch issues that raw metrics alone would miss (broken tags, missing metadata, indexability problems, and similar on-page issues).
- AI-generated recommendations — that combined data is handed to an LLM through a LangChain pipeline, which reasons over it and writes a plain-language report of what to fix and why.
- Delivery — the report goes out automatically by email, and progress is also streamed live to the frontend via Server-Sent Events, so a user watching the dashboard can see each stage happen in real time: fetching credentials, pulling Search Console data, scraping the site, thinking, and finishing.
The trickiest part wasn't any single step — it was making the pieces trust each other. Search Console data alone can tell you a page is underperforming; it can't tell you why. Combining that with a live fetch-and-render check of the actual page gave the agent enough context to make specific, useful recommendations instead of generic SEO advice. Keeping OAuth credentials, the app's own session system, and the scheduled Lambda job cleanly separated also took some iteration — they're deliberately independent so a revoked Google token doesn't silently break the app's own login state.
AWS services used
The system is fully cloud-deployed:
- AWS Amplify hosts the Next.js frontend with CI/CD.
- AWS Elastic Beanstalk runs the FastAPI backend.
- AWS RDS (PostgreSQL) stores users, linked accounts, credentials, and sessions.
- AWS Lambda, triggered by AWS EventBridge on a daily/weekly schedule, runs the report-generation pipeline.
- Amazon Bedrock powers the AI analysis step that turns raw SEO data into recommendations.
- AWS CloudWatch handles logging and monitoring across the pipeline.
Architecture, in short: the browser talks to Amplify, which talks to the Elastic Beanstalk API, backed by RDS. On its own schedule, EventBridge wakes a Lambda function that pulls from the Search Console API, runs the analysis through Bedrock, and emails the result via SMTP — independent of whether anyone is looking at the dashboard at that moment.
What I learned this summer
Across the Summer Build Series, the shift has been from "can I ship something this weekend" to "can I ship something that keeps running after the weekend." The Wishlist build was about an idea; the creative app was about shipping something people could try; this agent was about building something that runs on its own schedule, touches a real external API under real auth constraints, and produces something genuinely useful without a human in the loop each time. Wiring OAuth, scheduled Lambda execution, and an AI reasoning step into one dependable weekly loop was the biggest step up in complexity — and the most satisfying to see finish a cycle unattended.
Try it / see the code
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article