I Let an AI Agent Run 5 Dead Websites for 30 Days — Siraj Raval on Wins, Failures, and Next Research

Siraj Raval's video locks the framing up front. He gave an AI agent five websites and zero instructions for 30 days. Three lost money, one barely broke even, one was a real surprise. This is explicitly not a passive-income hype piece. What he wants to show is where this kind of agent makes smart architectural decisions, where it breaks, and what that tells us about where the technology actually is in 2026.

SEO is historically among the most labour-heavy parts of running an internet business — keyword research, content briefs, technical fixes, internal linking, ranking tracking, months of compounding work. The question Siraj wanted to answer was whether an autonomous agent can do this work without him in the loop, and what the failure modes look like when it does. He bought 5 sites from Flippa that had been built but never properly operated (deliberately in different niches, $2,500 total), typed one command per site on day 1, and closed the laptop. The agent was Auto from SearchAtlas, who sponsored the video — disclosed up front.

The standout win is site 3, a local services site untouched in years. Siraj's day-one command was clean: dominate local SEO for city + service, fix Google Business Profile alignment, build location keyword landing pages, rank in the local pack. On day 9, the agent surfaced a structural insight on its own — local-pack rankings attach to the Google Business Profile, not the website. So rather than pouring effort into website pages first, it prioritised GBP alignment, migrated reviews, and built location pages that backlinked the GBP rather than competing with it. Result: 47 leads through the contact form, sold to a partner trade business at $25/lead, $1,200 total — above acquisition cost. The agent figured out the niche-correct strategy unaided.

The failure on site 4 is more instructive. Affiliate comparison site, day-one brief in the same shape as the others — identify the highest-opportunity review keywords in the niche, write authoritative comparisons, build internal linking that concentrates authority on the money pages. Auto executed flawlessly: 22 comparison pages, internal linking optimised, schema markup added. Technically all correct. Result: $230 in affiliate revenue against a $700 target. The agent did what was asked. The brief itself was structurally wrong. The agent had no way to notice.

30-day totals: $2,500 in, $4,120 out, $1,620 net — a 1.64× return excluding Auto subscription costs. Not the viral 'make money while you sleep' shape. But for sites that were generating $0 going in, the numbers are real.

Three takeaways segmented by audience. (1) For businesses: agents like Auto are the right-shape tool for the marketing labour problem — test it on one underperforming page with the free 7-day trial. (2) For marketers: the leverage moves to writing the right brief. The agent will execute. Your value shifts to strategy, niche selection, and prompt quality. That's not a worse job — that's a more interesting one. (3) For agent builders: the failure mode on site 4 is the architectural problem worth solving. An agent that can recognise when its own brief is structurally bad — not just execute inside it — is the next interesting research direction. Siraj calls out Mistral, Anthropic, and OpenAI by name: that primitive doesn't exist yet.