Multi-agent SEO: how we drove S$350,000 incremental revenue with no ad spend.

Most “AI for SEO” advice is content-team advice with a chatbot bolted on. Write more posts, faster, with ChatGPT. The result is what every SEO team is now drowning in: more posts, lower rankings.

What worked at Hashmeta was the opposite move — fewer posts, with a multi-agent system that compounds editorial discipline instead of replacing it. Twelve months later: S$350,000 incremental revenue, no ad spend, 5M organic clicks/year, #1 rankings on 12+ competitive Singapore terms within 45 days, first paid client in 60.

This is the architecture.

Five agents, two stop-points, one expert in the loop.

• Strategist agent. Keyword research, search-intent classification, brief generation. Output is a brief, not a draft.
• Writer agents. Long-form drafts in the Hashmeta voice, with citation discipline. Output is a draft, not a publish.
• Editor agent. Fact-checking, internal linking, schema markup. Output is a polished draft.
• Publisher agent. Final formatting, image selection, publish scheduling.
• Me. Stop-point 1 at brief approval (90 seconds); stop-point 2 at final-draft approval (5 minutes).

Three things made this work that don’t show up in most AI-SEO advice:

1. The editor agent was the bottleneck, not the writers. We over-invested in writer-agent quality at the start. The actual leverage was downstream — fact-checking and internal-linking discipline. When the editor was strong, the senior pass took 5 minutes per article. When it drifted, the senior pass took 25 minutes and rankings followed it down.

2. Brief quality dictated everything. Tight briefs → fast senior pass → consistent output → ranking. Drifty briefs → slow senior pass → inconsistent output → traffic plateau. We curated 50 briefs by hand before we automated brief generation. That’s the order; reverse it and you fail.

3. Day-one metrics loop, not week-three. The strategist agent only got smart once the GSC + Ahrefs feedback loop was instrumented. We built the loop in week three; we should have built it in day one. Three weeks of unguided publishing is a lot of pages to clean up.

The system isn’t a prompt template. It’s a stack of OpenClaw skills, a Python orchestration layer, and a senior reviewer who is paid to disagree with the agents. The combination is what compounds.

If you want the full method, the playbook is documented as The AI SEO & GEO Blueprint — the lead magnet on this site. If you want to build it with hands on the keyboard, SEO-1 — AI SEO & GEO Blueprint is the four-hour workshop. If you want it shipped for your business, send a brief.