Six months ago I was spending 15 hours a week on keyword research, gap analysis, and content planning. Now the agent does it, and I spend 45 minutes reviewing what it found.
The agent is named Kai. It's a Claude AI running on a cron job that fires every Sunday at 6 AM. Here's what it actually does:
1. Hits the Google Search Console API with my domain and pulls impression/click/position data for the past month 2. Cross-references with my content calendar (a spreadsheet I maintain) 3. Runs DataForSEO queries on keywords where I'm ranking 4-8 (low-hanging fruit territory) 4. Audits my existing content for keyword gaps and opportunities 5. Spits out a report with ranked recommendations ordered by potential impact
The whole thing runs in about 8 minutes. Costs me $0.80 per run in API fees. Would cost me a full workday if I did it manually.
Why I built it
The original friction point: I'd sit down Friday afternoon with the best intentions to do keyword research, get three tabs deep into SEMrush or Semrush, and lose 4 hours to the rabbit hole. I'd find great opportunities but never act on them. The data would sit. It was friction at the moment of action.
An agent removes that friction by doing the part that bores me (API calls, data compilation, cross-referencing) and leaving only the part that requires judgment (should I actually write about this, and how would I angle it).
The stack
- Claude Opus via API (initially tried Sonnet, but it kept missing patterns in the data)
- Google Search Console API for impressions/clicks/positions
- DataForSEO API for SERP data and keyword metrics
- A shell script that orchestrates everything and handles retries
- S3 bucket where the reports land, and I get a link via email
- I have API budgets (GSC is free, DataForSEO costs $20-30/month at my volume)
- I understand the APIs enough to debug when something breaks
- I'm comfortable letting an AI make recommendations I can easily override
- I've invested time in building the infrastructure
- Agent that generates content briefs (outputs outlines for new posts)
- Agent that monitors SERP changes for client keywords
- Agent that builds link prospect lists from competitor analysis
No fancy agent framework. No LangChain or similar. Just a shell script that calls the Claude API with instructions, waits for the response, formats it, and ships the output.
What changed
First week: The agent recommended I write about "RV trailer weight distribution." Niche, 50 searches/month, low competition. I wrote it. It's now my highest-traffic post in that category, ranking position 2.
The agent caught a pattern I'd missed: my site ranks well for problem-solution content but not for list posts. So it now flags every opportunity where the top-10 results are lists, and I've started writing more list content. Traffic from those posts is up 40% in 90 days.
Best part: the agent learns my preferences. I marked "avoid these topics" in one report comment, and now it deprioritizes them automatically.
Why you probably shouldn't copy this exactly
This works for me because:
If you're not comfortable with APIs or cron jobs, there are off-the-shelf tools that do similar work. SEMrush has a content assistant. HubSpot has gap analysis. They're slower and more expensive per-run, but they're less of a setup burden.
That said, if you have even basic technical comfort, building this yourself takes a weekend. The payoff is real.
The actual thing I learned
The agent works because it removed my personal friction, not because it's doing something impossible. A human SEO could do this work faster than the agent if they cared enough. The agent wins because humans don't. I don't. So I outsourced the part that was going to not happen anyway.
That's the real usefulness of agentic AI right now. Not replacing expertise. Removing the friction between knowing what to do and actually doing it.
What's changed for my clients
When I pitch the agent to new clients, I don't say "I have an AI." I say "You get weekly keyword opportunity analysis without you doing anything."
That matters more to them than knowing the mechanism. They don't care if it's an agent, a human, or magic. They care that they get intelligence delivered to them on a schedule.
The agent runs even when I'm on vacation. It doesn't take days off. It doesn't get tired. It processes data consistently the same way every time.
That consistency is more valuable than perfection. A perfect analysis that never happens is worth zero. A good-enough analysis that happens every single week is worth everything.
The roadmap
I'm planning to add three more agents in the next quarter:
Each one saves 2-3 hours per week per client. At this rate, I'm automating myself out of busywork and back into strategy.
That's the goal anyway. The agents are just tools to get there.
The most important thing: I don't stay behind the same desk waiting for work to accumulate. I'm building systems that work when I'm not paying attention. That's how you scale a one-person agency to a real business.